StarNet v0.13.5

Download

Platform	Installer
Windows (x64)	StarNet_0.13.5_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.13.5_aarch64.dmg
macOS (Intel)	StarNet_0.13.5_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.13.5

Before you update: if you might ever go back to 0.13.3 or older, save a station backup first (SETTINGS › APP & BACKUP › STATION BACKUP). Older versions don't know SEND TO or SPLIT THE JOB and delete those settings the next time they save your station. If agents are working when you install, choose WAIT FOR AGENTS.

The big one. 0.13.3 was security only; this is the feature update that was promised with it, polished end to end. Your agents can bring crew into a chat and hand work to each other, conveyor lines can hand finished jobs to other lines, building is drag-and-drop, the station draws smoothly even when StarNet isn't focused, and Claude Code runs are faster and cheaper. The whole app also got an organization pass: one place for each job, one name for each thing, and plain, true wording wherever people got stuck. Plus a long list of fixes from walking the whole app as one product.

Updating from 0.13.3

• Update from inside StarNet: UPDATE CENTER › INSTALL UPDATE. If agents are working, WAIT FOR AGENTS only closes the warning; press INSTALL UPDATE again once NEEDS YOU and COMMS are idle.
• Using the downloaded installer instead? It closes StarNet right away, so make sure no agent is working first. If you quit from the tray to install, press RESUME afterwards to restart routines and away work.
• Your station, conversations, keys and settings carry over as they are. The save format is unchanged; 0.13.5 only adds a few new session settings with neutral defaults.
• Speaker on in COMMS? Replies now use the Voice chat model with reasoning off by default, typed messages included. To keep your usual effort, set SETTINGS › AI & MODELS › Voice chat model › REASONING to SAME AS TYPED.
• Telegram or Discord bots: messages sent during the minute StarNet restarts to update aren't processed. Send them again afterwards.
• Going back to 0.13.3 later? Save a station backup and copy your whole workspaces folder first: 0.13.3 doesn't know some 0.13.5 settings (SEND TO, SPLIT THE JOB, new skill states) and drops them when it saves.

Agents & chat

• Bring crew into a conversation by asking. Say "add FORGE to this chat" and your lead adds them. They join the moment its reply ends, or the chat says why they couldn't. ADD AGENTS shows each agent's portrait.
• Delegated work has its own home. Work your lead hands off gets its own threads, folded under the conversation that asked, and one live card in that conversation: who is doing what, Stop while it runs, Approve once or Deny when a worker asks, and See work to unfold each worker's answer and what it did.
• Workers build on each other. When workers run one after another, the next one starts from what the earlier ones returned and the files they saved, and can open those files.
• NEEDS YOU is one list of everything waiting on you. It sits at the top of SESSIONS, and NOTIFICATIONS › NEEDS YOU shows the same list. It holds approvals and questions from your conversations, group chats that need you, delegated work and routine runs that failed, an agent waiting at STEP-IN, extensions to approve, builds to keep or discard, quests an agent says are done and goal-loop results to review. Each one is a row in one quiet NEEDS YOU panel, and an agent's builds share one row with their count: click a row to go where it waits (a build opens its own session, where you keep or discard it), and the keys to answer, retry or dismiss sit right on the row.
• Asking your lead to hand work to a named crew member now does it.
• Naming a file doesn't cost your lead its crew. A file name like notes.md or main.py is no longer mistaken for a website, so your lead can still hand the job off and search the web.
• Group chats: the lead can build and change lines from the group, each agent keeps its own room's gear (a researcher can search the web there), replies show their tool calls like one-on-one chats, and the bottom strips share one tray.
• Question and approval cards read at a glance. The question comes first, long ones fold behind "Show the whole question", answers are full-width keys with the agent's ★ suggestion lit, and the reasoning folds last. A one-time grant reads "Approve once" on every approval card in the app. When an agent asks for the same tool again in a chat, the card says what Always does.
• An answer that can't land says so. If a worker's approval on the DELEGATED card already expired or was answered somewhere else, its row says the approval is no longer waiting and your answer wasn't applied, instead of doing nothing.
• SEND NOW. A message you type while an agent works still waits its turn. Each waiting message now has a send now key that hands it to the running task, so the agent reads it before its next step. If the agent hadn't read it yet when you press STOP, reload or restart StarNet, it comes back to the message box.
• STOP gives your words back. Messages waiting behind a stopped run go back into the message box instead of disappearing, and so does one you sent now that the agent hadn't read yet. Open approval cards say "the run ended before an answer — nothing was approved".
• Typing an answer answers. A reply typed in COMMS while the lead is asking you something answers that question instead of starting the job over.
• The first message after a restart goes through. It waits for the conversation to load and sends once, instead of being refused. If the station answers with an error or never answers, you're told, and your message stays in the conversation with a Try again key.
• A goal breakdown never covers your work. When an agent offers to break your goal into steps after a run, the panel sits above the message box and steps aside when an approval or a question arrives. Esc means not now, and one Esc closes only the thing it was pressed on.
• COMMS says which conversation you're in. The title reads "COMMS ·" and the session's name, and two untitled sessions read "New session · NOVA" and "New session · MIRA" instead of looking the same.
• Fewer empty sessions. An agent your lead or a workflow hires doesn't get an empty session until you pick it ON THE LINE in COMMS, and a starter idea fills the blank chat you're in. When an agent is busy somewhere else, the chat says where ("NOVA IS BUSY IN …").
• Keys say what they do. In a group chat the stop key says "Stop every agent in this chat", and an @ from General says it opens a group chat beside General.
• /goal is a session goal, and says so. It works while StarNet is open and stays through a reload, a GOAL LOOPS key sits right there for goals that should last, and each step shows as a station message, not as words you typed. /loop, /routine, /cron and /away answer with a key into the right AUTOMATE tab.
• RECRUIT speaks one language. RECRUIT NOVA adds a new agent; RE-SPEC MIRA INSTEAD changes the one you have.
• The AGENT DOSSIER says where each setting lives, and switches that change every agent say so ("for every agent", "all agents"). An agent's PURPOSE and house rules say only that agent follows them; your standing orders say every agent does.
• Autonomy says what will happen, not what unlocks. You read "While you're away I'll do up to 3 small jobs a day", the dial says SMALL JOBS, and REACH says LOCAL ONLY for work that never leaves this machine.
• Moving PACE or REACH no longer lifts an E-STOP, and any settings click that does lift one says what runs again.
• A worker running in the background keeps its approval open after the lead's turn ends, instead of being denied a second later where you couldn't see it.
• Telegram, Discord and other chat apps: in a direct chat, a plain follow-up joins the running task instead of cancelling it. "stop" or /stop still stops it.
• CONNECT and RESUME work with long agent instructions. Telegram, Discord and the other chat apps no longer fail with "Failed to fetch" when an agent's instructions are long, and CHANNELS says in words when a request got no answer.
• Long chats keep your earlier instructions. Messages that no longer fit the model's window are carried forward in your own words, instructions first, instead of silently disappearing.
• Memory tells the truth. Remembered notes carry the date they were saved and the newer one wins when two disagree. An agent that says it saved something without saving it is told to, and the "remembered" receipt lists only this run's saves. Forgotten memories stay forgotten when you restore a backup. "Never mind" and other passing remarks no longer become standing orders; "remember that…" still does.
• Approvals carry into the next design. Approving work, including ★ KEEP THIS STYLE, records what was actually made (files, colours, fonts, image prompt) and the whole crew draws on it. Image generation can use an approved image as a reference on OpenRouter, the ChatGPT plan and the OpenAI Images API, and says so where a connection can't.
• Skill changes wait for you. A skill an agent creates or changes becomes a proposal in CONNECT › ABILITIES › AGENT SKILLS. It needs a passing evaluation and your approval before it's used; the current version keeps working meanwhile.
• **Spoken turns...

StarNet v0.13.3

Download

Platform	Installer
Windows (x64)	StarNet_0.13.3_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.13.3_aarch64.dmg
macOS (Intel)	StarNet_0.13.3_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.13.3

A security update. Everything else stays exactly as it is in 0.13.2. The big feature update (SEND TO between lines, drag-and-drop building, memory and long-chat fixes, clearer stops, local model fixes) is coming next as its own release.

Security

This release tightens StarNet's local security:

• Local access checks are stricter. The station only hands its access key to the app running under your own account, and checks every request.
• Outside content stays contained. A run that has read outside content, such as a web page, a webhook or a forwarded message, can't create or change routines or skills, and that content no longer feeds what agents learn on their own. A forwarded Telegram message now counts as outside content.
• Credential folders are off-limits to file tools. Agents' file tools can't read well-known credential folders (SSH, GPG, cloud CLIs, saved logins, signing keys) unless you give them Full Power.
• Imported station configs carry only the permissions you curated. Standing approvals to run commands, send or spend are skipped and listed in the import notes.
• Project files are checked before anything is opened, so a network path in a project can't make the station reach out to another machine first.
• Code tools are more careful with untrusted projects. The Rust language server only starts for a project you've chosen to watch.
• The browser page only runs its own scripts, and the desktop app holds fewer raw permissions.

Update to get these fixes.

StarNet v0.13.2

Download

Platform	Installer
Windows (x64)	StarNet_0.13.2_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.13.2_aarch64.dmg
macOS (Intel)	StarNet_0.13.2_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.13.2

Before you update: if the in-app update from 0.13.1 stops with "Update paused - no verified recovery point was created", download the full installer from the releases page and run it over the top (your station is kept). If agents are working, choose WAIT FOR AGENTS, not INSTALL ANYWAY. macOS 10.15 Catalina is no longer supported: StarNet needs macOS 11+ with Safari 16.4+.

A large repair update: credits and spending tell the truth, in-app updates are safer, routines get their keys and tools back, the station runs lighter, and your agents can now drive StarNet for you.

Behavior changes

• Autonomy runs on your local day, not UTC, and a focus you declare for it is enforced: off-focus ideas are dropped.
• One "While you were away" session. Routines, away builds, loops, autonomy drafts and failures fold into one pinned report instead of a session each.
• Autopilot cabinet writes now really run when a cabinet is placed and file writing is granted, with consent, a checkpoint and UNDO. Before, they always fell back to a desk draft.
• A resumed run whose earlier history can't be verified starts locked and says why.
• From 0.13.2 on, in-app installs see all running work (routines, workflows, channels, /v1 calls, background workers): WAIT FOR AGENTS holds the install, and INSTALL ANYWAY stops that work and lets it book its spend before writes freeze.
• A greeting that names an agent ("hey nova", "thanks claude") may start a task run with tools, and a retry or redirect like "ok, try it now" is now treated as a task.

Before you downgrade: settle any interrupted run in SETTINGS › SPENDING LIMITS first, or 0.13.1 refuses every run that has a spending limit set.

Fixed

Updates & recovery

• From 0.13.2 on, in-app updates no longer stop on one unreadable file. The recovery point skips locked or unreadable entries and lists them, and captures a WORKSPACES folder that is a link (#91).
• No stranded spend across an update. From 0.13.2 on, INSTALL ANYWAY lets stopped runs book their spend first, and the "saved by a newer StarNet" screen offers DOWNLOAD LATEST when the in-app update can't finish (#91).
• SAVE-NET recovery details check whether the local engine actually answers instead of always saying REACHABLE.
• macOS: an outdated web engine is named as such, never "a broken file".

Credits, StarNet link & spending

• A funded account is never told it is out of credits. A failed balance check says so, a refused link says relink, a balance held by your own running runs is called held, and a PER RUN cap above your balance is clamped to it.
• Slow link checks no longer hide your account. WAKE, the STORE, the connect screen and SETTINGS wait out a slow link repair, and a link that met a down account service repairs itself; RETRY AUTO-RESUME re-reads it (#76).
• The desktop app keeps its StarNet link across restarts (#93).
• StarNet Managed models: a saved model like gpt-5.5 is matched to the managed catalog instead of being wiped to "not available", and CHOOSE MODEL no longer opens and instantly closes.
• Interrupted runs can be settled. SPENDING LIMITS lists each one: enter its charge and SETTLE, or COUNT AS its per-run limit. It no longer refuses runs that have a limit set with "Spend history is unavailable"; a graceful Quit books stopped runs' spend.
• Spend stops point at SETTINGS › SPENDING LIMITS, and SAVE LIMITS saves only the limits you changed.
• LINK DOWN says why the station service is gone and, on desktop, offers a restart. Pairing failures name which side failed.

Providers & models

• Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 conversations no longer fail with "Invalid signature in thinking block" on the Anthropic provider (#73, #94), and Sonnet 5.5 with thinking OFF works. Haiku 5.5 history edits and 5.5 price estimates are corrected.
• Claude Code agents can see images, and reuse their cached prompt across runs (#68).
• Gemini no longer refuses every request with a station-planning tool attached.
• Mistral, OpenRouter and other OpenAI-compatible providers no longer reject runs over extra message fields.
• Groq retries a badly formed tool call instead of ending the run.
• ChatGPT: an expired sign-in shows ⏼ RECONNECT on its card (#92).
• Key checks tell the truth: a refused Anthropic key says HTTP 401, a network failure names its cause (#90); OpenRouter keys verify through tunnels (#62); Kimi uses your account's region (#70).
• Anthropic quick picks offer current models (Opus 4.1 and Haiku 3.5 were retired).

Routines & workflows

• In browser mode, routines, Run Now and autonomy use the key you saved in the browser (#89): StarNet keeps a copy in WORKSPACES/.secrets/provider-keys.json, and REMOVE deletes it. A line's schedule runs on its agent's provider.
• Routines saved before 0.13.2 get their web tools back (#58): their saved tool list is repaired and saved once on first start. Unknown toolset names are refused by name, and Run Now gets the same tools as the scheduled run.
• Workflows with no WORKING FOLDER tell the agent its files stay private and how to save into a trusted project (#60, #81).
• Delegated images and voice files land in the trusted project, so the lead and the next Conveyor stage can read them, and a worker in ASK mode is told why it has no shell (#77, #88).

Station browser & web

• The station browser heals itself. An orphaned browser is cleaned up, a held profile falls back to a fresh temporary one, and RESET STATION BROWSER is named when needed (#61).
• Clearer failures: failed loads, tunnel refusals, TLS errors and unreachable search engines name the real cause and host.
• Tabs behave: new-tab links open in the same tab, a closed first tab no longer strands the browser, and clicks and typing never land on a different tab.
• One-site requests keep the browser, and a page quoting an error is no longer read as a dead host.

Station & UI

• Faster, lighter station. The world draws at 30 fps focused and 5 fps in the background, and conveyors, lighting and the session rail cost far less (#69).
• QUEST LOG no longer freezes the app.
• Window text is selectable, and OPEN THE FULL CONVERSATION opens a workflow's session (#87).
• Hover hints work on touchscreen laptops driven by a mouse.
• Copy diagnostics works on macOS. NEEDS YOU counts match the rows shown.
• Connectors: editing keeps saved secrets, errors say why, OAuth sign-in survives a port change, and Intercom signs in with OAuth.
• API keys are used with Windows syntax in shell commands.
• E-STOP says when a run stopped itself and offers no TRY AGAIN.

New

• Your agents can drive StarNet. Ask in plain words and the agent opens any window for you, changes settings, and handles E-STOP, the away queue, cleanup, groups and notifications. Agents now understand every part of StarNet, and a lead can read which sessions are busy (#55).
• Ctrl+K (Cmd+K on Mac) FIND searches every window, agent and conversation.
• Approvals show the whole request. A long command or path is no longer cut at 80 characters: the card shows one short line, and "Inspect complete request" holds every character (known secret patterns hidden), with COPY COMMAND for shell commands. Group chats carry it too; Telegram and Discord approvals show up to 3000 characters and say when they cut.
• /v1 and ACP editor runs can now use the agent's tools under your PERMISSIONS and approvals (#96); a model without tools still answers in text.
• HINTS switch in SETTINGS › LOOK & SOUND turns hover bubbles off, also from chat (#79).
• ERASE EVERYTHING in SETTINGS › APP & BACKUP (desktop app) deletes all StarNet data on this computer and restarts fresh (#65).
• One-click desk assignment: a desk's card shows WHO SITS HERE chips.

Changed

• Autonomy remembers what it already built, focuses on work you repeat, and keeps one version of a build waiting.
• Permissions ask one question, with the same three answers in SETTINGS and each agent's Dossier; AUTONOMY leads with ON ITS OWN.
• Custom classes are one page; tools and skills are built in.
• DELIVERABLES uses the new glass look.
• macOS 11 Big Sur is now the minimum (with Safari 16.4+, which macOS 13.3+ has built in).

Thanks to @bonjurroughs for prompt caching and hover hint fixes, and to @xLagerFeuer for the station status idea.

StarNet v0.13.1

Download

Platform	Installer
Windows (x64)	StarNet_0.13.1_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.13.1_aarch64.dmg
macOS (Intel)	StarNet_0.13.1_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.13.1

A repair update: web requests work again, group chats save and show who is in them, COMMS reads like a conversation, crew no longer freeze in the halls, each floor OUTBOX shows only its own line's work, schedules let you pick who they run as, and spending limits hold during retries.

Fixed

Web requests

• web_request and web_fetch work again. Since 0.12.5, every web request an agent made could fail with "fetch failed", even for sites like google.com. Requests now connect normally, and on networks without IPv6 they fall back to IPv4.
• Clearer network errors. When a request fails to connect, the error now includes the reason code (such as ECONNREFUSED) and, when known, the host or address, instead of a bare "fetch failed".

Group chats

• Adding an agent saves at once. + ADD and REMOVE now save the moment you press them. Before, an add was only staged until a save key that could sit below the window's edge, so minimizing StarNet lost it. IN THIS CHAT lists only members the station has confirmed.
• @ works in every chat. Type @ in any chat to pick an agent from a menu over the message box (arrow keys, Enter or Tab to pick, Esc to close, @ALL in a group). Picking an agent who is not in the chat adds them first: a direct chat becomes a group, and from General a new group opens beside it.
• Names with spaces. "@researcher 2" now reaches RESEARCHER 2, not RESEARCHER. A crew agent outside the chat is named as "not in this chat yet", and an unknown @name tells you how to send it as plain text.
• You can tell it is a group. Group rows carry a [GROUP] tag, the COMMS header leads with [ GROUP ] and lists the members in their colour (past two lines, the rest fold into +N), and the transcript speaks the same conversation style as COMMS.
• Groups you are not viewing still reach you. Replies, questions and approvals in another group now show on its row (Working, Reply needed, Approval needed) and mark it unread. Before, an approval there could go unseen until it was automatically denied after 5 minutes.
• A steadier transcript. Text you select stays selected while agents reply, every message has a copy key, and back-to-back replies from different agents each show their name.
• Pauses are explicit. A paused group shows one CONTINUE, a reply the pause stopped is named under its message with a RETRY, and sending to a paused group no longer starts work queued before the pause ahead of your message.
• Approvals fit. Approval keys sit on the first line of their row and rows that need you sort first, so ALLOW ONCE and DENY are no longer clipped in a short window.
• A group survives crew changes. Deleting an agent that sat in a group no longer breaks the group: it leaves the group, @ALL and plain messages go to the members still on the crew, and if it led the group the lead passes on.
• Turning a chat into a group no longer fails on one oversized or missing attachment; the readable files are shared and the message names the one that was not.

COMMS readability

• Reads like a conversation. Messages show a 12-hour time, and a [ TODAY · 3:42 PM ] divider marks the first message, a new day, or a gap of 30 minutes or more.
• Your turn stands out. Your message is a gold > prompt line on its own rail, and the agent's reply mirrors it: name and time above, the text on a phosphor rail over a light dither.
• No more boxes. Station lines are ruled terminal tags, code is lit without outlines, choices are [ BRACKETED ] keys, and a reply split across rows reads as one block under one name.
• The run summary stays under its reply. On fast runs it used to land between your question and the answer.
• Smaller rating. The rating prompt for a finished task is one slim row with two thumbs under its run, not a large centered card.
• Steadier hover. Hovering a message no longer makes it glow; its time brightens and the copy key appears.

Crew movement

• Crew no longer freeze in the halls. Two or three crew whose paths crossed could each wait for the other forever and stand still for minutes. They now never wait on someone who is waiting on them.
• A backstop for stuck crew. A crew member trying to walk somewhere who has not moved for 12 seconds drops that path: crew at work re-route to their desk and idle crew pick something new. Seated crew, conversations, gatherings and crew waiting on your approval are left alone.

Floor OUTBOX & deliverables

• Each OUTBOX is its own line's. Clicking a floor OUTBOX opens the workflow that ships into it, with its newest result and last jobs, instead of every deliverable on the station. An OUTBOX on no line says so.
• Per-line counts. Each OUTBOX's SHIPPED pallet, crate stack and hover count only its own line's work. Routines, chats and away runs wait in DELIVERABLES › TO REVIEW, never on a chute.
• Workflow jobs are filed in DELIVERABLES too. Every finished workflow job also gets its own row in DELIVERABLES (under the workflow filter) with what you asked, what the line delivered and the files its steps wrote, with OPEN IN WORKFLOWS. It stays after you collect the crate. A job where no step ran is not listed.

Schedules & recipes

• RUN AS. Choose which crew member a schedule runs as when you make a routine from a recipe, and change it later from the routine's edit form. Before, it always used the active workstream's agent (usually the Overseer), and the only way to move it was asking in COMMS.
• EDIT and DELETE are easy to find. On your own recipe, ✐ EDIT and ⌫ DELETE sit in the recipe's header. An edit saves in place, and from its next save on, the recipe shows when it was last saved.
• Same-name guard. Saving a recipe with a name that already exists warns you first; save again to keep both. An imported recipe whose name is taken lands as "Name (2)" and says so.
• An honest send warning. Scheduling a recipe whose directive looks like it may send or write now says plainly that unattended sends are refused by default, so it can only draft, but that it really sends if the agent has Full Access or you grant the routine connected tools. FULL BYPASS also lets it send.

Spending, providers & delegation

• Spending limits hold during retries. A run now checks its per-run limit and the station's spending caps before every retry and recovery attempt, not only between turns, so a run stops retrying once a limit is reached.
• Quest refresh respects spending caps. Quest refresh, automatic or by hand, no longer calls a provider once a spending cap is reached; the Quests panel marks the cycle skipped.
• Provider status tells the truth about keys. In Settings, OpenRouter now tests your saved key against OpenRouter itself and can show VERIFIED. A custom endpoint whose model list is public no longer shows VERIFIED for a key it never tested.
• Delegated work can be verified. A lead asked to confirm work it dispatched can now look up the finished run, instead of concluding the worker never ran. The worker's Dossier RECORD tab shows that run up front as "delegated by" its lead.

Known issues

• Images on Claude Code. Agents running on Claude Code cannot see images; image analysis says so and suggests another way.

StarNet v0.13.0

Download

Platform	Installer
Windows (x64)	StarNet_0.13.0_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.13.0_aarch64.dmg
macOS (Intel)	StarNet_0.13.0_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.13.0

A simpler dock, a new glass look, apps your crew builds, StarNet Remote for your phone, Claude Code as a provider, and a lead that builds rooms and changes settings from chat. Heads-up: stations that never set autonomy now start at SUGGEST, routines stop after 90 minutes, a StarNet-credits run stops at $2 unless you set a per-run limit, old notifications are cleared, and going back to 0.12.5 drops made props and plugin terminals. See Heads-up below.

Highlights

• One dock, fewer doors. The dock keeps its four menus (CREW, WORK, BUILD and SYSTEM), and MY WORK, AUTOMATE and CONNECT each open one window with tabs in place of six separate dock buttons. A new station's dock adds buttons the first time you need them.
• A new glass look. Frosted glass panels, CREW as agent cards, COMMS replies laid out like reports, and session rows that lead with the agent's name in its own colour.
• APPS. Describe an app (a dashboard, a tracker, a tool, a game) and your crew builds it in its own window. It can update itself on a schedule, and you change it by describing the change.
• Your lead runs the station from chat. Ask for a setting change, a new room, a whole station layout or a workflow line, and the lead does it through the same paths your buttons use, asking your approval first unless the agent has Full Access.
• Workflows you can just use. WORK › AUTOMATE › WORKFLOWS lists every line with one key to send it a job. SET IT UP FOR ME builds a line from a description, and NEEDS CHANGES turns what you did not like into suggested fixes.
• Build Mode upgrades. Box-select many things, move or duplicate them as one edit with one undo, drag to lay a row of furniture, resize or furnish a room by clicking it, and make your own props on StarNet credits.
• StarNet Remote. Pair a phone by QR code to chat with your crew, answer approvals and get notified when an agent needs you or finishes a task you sent.
• Claude Code as a provider. Agents can run on your own Claude subscription through your installed Claude Code. ChatGPT, Grok, Kimi and Claude Code can each hold several sign-ins and move to the next one when an account hits its usage limit.
• One shared browser. You and your COMMS agents use the same station browser, by default a real Chrome window with its own profile. STEP-IN hands you an agent's browser when it needs you to sign in.
• Feedback that sticks. Every thumbs up or down and every correction is saved, and your most recent ones shape every agent's later runs.

New

Dock, windows & look

• ONE MENU dock. CREW holds AGENTS, RECRUIT and YOU. WORK holds MY WORK, AUTOMATE and QUESTS. BUILD holds BUILD MODE, CONNECT and NEW APP. SYSTEM holds FIELD MANUAL, SETTINGS, UPDATES and NOTIFICATIONS. An APPS menu appears once you have an app.
• Tabbed windows. MY WORK is TASKS · DELIVERABLES · RECIPES. AUTOMATE is WORKFLOWS · SCHEDULES · GOAL LOOPS · AWAY WORK. CONNECT is ABILITIES · CHANNELS. Tabs swap in place, and a tab click never throws away text you have not saved. The OUTBOX is now DELIVERABLES › TO REVIEW, shown only when something waits for your verdict.
• A dock that grows with you. A brand-new station starts with the buttons a newcomer needs; the others come online the first time they matter, with a NEW SYSTEM ONLINE notice. Nothing is locked: SHOW EVERYTHING (in SETTINGS › LOOK & SOUND and QUESTS › Progress) puts every system in the dock, and a station that was already set up opens with everything shown.
• Glass everywhere. CREW, COMMS and every window share one frosted-glass style with matching keys, lit rows and one drawn icon set. CREW shows each agent as a card with its status lamp, and the station view carries only the CINEMA key, with the link state in the top bar's UPLINK.
• Station art. The station's hull and wall materials are redrawn, with three new hull finishes, and 44 crew skins get steadier walk cycles in which heads, arms and held items hold still from frame to frame.

COMMS & sessions

• Per-agent threads. Clicking a CREW row narrows the session list to that agent's sessions. A group chat lists under every member as soon as they join.
• Inbox rows. Each session row leads with the agent's name in its own colour beside the time, with the session title under it and the square session lamp (read, unread, working, reply, approval, failure).
• Reports-style replies. Headings read as section labels, bold stands out, code blocks wrap in their own well, and tables are easier to read. Rating a reply is two thumbs on one row, and REMEMBERED starts collapsed.
• 12-hour clock. Message times read on a 12-hour clock, with the date on older messages.
• Quieter header. The COMMS header no longer repeats the transcript's working spinner and timer; it shows only faults, such as a station it cannot reach. Two icons sit there instead: + adds agents and the globe opens the BROWSER (plus the HUD key in the desktop app).
• A "yes" runs the work. Answering "yes" or "do it" to an agent's offer runs it as a task with tools, in COMMS and on Telegram and Discord. Every run also knows today's date.

Apps

• NEW APP. Give an app a name and say what it should do; the lead builds the page, fills it with real content and opens it in its own glass window. Asking for an app in plain COMMS works too.
• AUTO-UPDATE. Each app can refresh itself every 15 minutes, hourly, every 6 hours, daily or weekly (or never), following what you wrote, and an update may redesign the page itself. A refresh keeps running if you close the window.
• Change it by describing it. The bar under every app takes a described change, refreshes now, and says when the app last updated and when it runs next.
• Apps are only for things inside StarNet. A site to host, a React project, a script or a browser extension is built as real files instead.

Your lead, from chat

• Change any setting from chat. The lead can change about forty things that used to need a click: an agent's model, personality, name, skin, approval or reach; renaming, pinning, archiving or deleting sessions; the look; memory; backup models; spending caps; routines on or off; connectors, abilities and skills; trusted folders; and more. Each change uses its button's own path and is reported done only once the station has actually changed.
• Bigger changes ask separately. Anything that widens what agents may do or spend (Full Power, Full Access for an agent, spending caps, the autonomy dial, standing approvals, trusted folders, plugin approval) gets its own approval card every time unless the agent already has Full Access, is never covered by an earlier "always", and never runs unattended. A run that has read untrusted content asks again even under Full Access. The lead never enters a credential, answers an approval or lifts E-STOP.
• Station builder. The lead can see the floor and build where you ask: furnished preset rooms, rooms described part by part ("the left side cozy, the right side a line that builds and tests code"), restyling, refurnishing, clearing and removing rooms, and hallways. It can make a missing prop on StarNet credits, with the object and price on the approval card. The card draws the plan before anything is built, and each build is one undo.
• Whole-station layouts. The lead can lay out the whole station as a diamond (every room on an even grid round the bridge, keeping its shape as rooms are added) or a concourse, every room furnished in a style. "No, undo that" takes back its last build.
• Lines from chat. The lead can build a line you describe, set up its steps, budget and routing, set what starts it (a schedule, a folder or a webhook), send it one test job with your approval, read the result and fix the line.

Workflows & lines

• WORKFLOWS window. Every workflow is listed with who works it, an honest status and one key: SEND A JOB or FINISH SETUP. NEW WORKFLOW places a starter line or one SET IT UP FOR ME builds from your description, and one agent can do every step.
• Results you can steer. A job shows the step working on it, then the whole result. NEEDS CHANGES suggests exact step changes to use or put back, KEEP THIS STYLE makes the line match a result you liked, and SEND IT AGAIN runs the job beside the last result. Jobs are kept across restarts.
• Workflow panel. The panel shows the line as a diagram of its real machines. The + adds a step, a branch or a sorter; a step's card can move it, add a reviewer, rename it or remove it, each as one undo. The INBOX card opens on SEND IT A JOB, shows live progress ("Now: step 1 of 2") and returns the whole result plus each step's own reply, and a test job has its own STOP key.
• Lines that fit. A ready-made line that does not fit where you click is laid out nearby, routed exactly as drawn, and work presets come ready to run with a setup guide. Joining two machines with the BELT tool makes the connection, a moved machine re-lays its belts, and a loose belt routes nothing and says so.

Routines & automation

• Plain-English schedules. "every day at 7am", "weekdays at 8:30am", "the 15th of every month at noon" and "every 2 hours between 9am and 5pm on wee...

StarNet v0.12.5

Download

Platform	Installer
Windows (x64)	StarNet_0.12.5_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.12.5_aarch64.dmg
macOS (Intel)	StarNet_0.12.5_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.12.5

Workflow lines you can step-test and start from a folder or webhook, Google Workspace early access, steadier long runs, and a large security update. Heads-up: metered runs now stop once spend in the last 24 hours reaches $25, until you press RESUME in Budget (editable; 0 turns it off). Some plugins, chat-channel bots and code.run ask for approval or pairing again. The heads-up items at the end list what to check.

Highlights

• Workflow studio. Each line gets one docked Workflow panel. You can step-test a line one stage at a time (pause, continue, edit the hand-off, re-run or rewind), one agent can crew several bays, and a line's INBOX can work in a trusted project folder (#25).
• Lines that start themselves. A file dropped in a watched folder, or a call to the line's webhook, starts one run of that line. Bay lamps, crate cards, INBOX plates and a TODAY row show what each line did.
• Google Workspace early access. Gmail (full or send-only), Calendar, Docs, Sheets and Drive can be connected before Google finishes verifying StarNet.
• Reasoning levels that match the model. Grok (xAI), DeepSeek and known OpenAI API models offer the levels each model accepts, and StarNet Managed models that support reasoning offer levels too. DeepSeek thinking mode works with tools, and setup errors say what is actually missing.
• Steadier long runs. Conversations are saved turn by turn, and interrupted runs show up in history and can be continued. Recovery from outages, stalled models, context overflow and cut-off tool calls is better, and COMMS shows when a model is slow or retrying.
• Spend safety. Metered runs stop once spend in the last 24 hours reaches $25. Loops that change nothing pause, and runaway unattended runs stop and say why. Conversation summaries no longer send screenshots as text, which could cost hundreds of thousands of tokens per summary.
• BYOK stations are no longer refused with "Out of managed credit" (#24). A specialist set to Follow station default runs on the Overseer's model and provider in COMMS, workflow-line hand-offs, RUN A SAMPLE JOB and chat channels.
• Windows commands report real results. Commands report their real exit code (0.12.4 reported exit code 0 for most commands that failed), and a cd carries over to the next command.
• A large security update. It covers network pinning, removing the API token from URLs, a locked-down desktop window, owner pairing for every chat channel, untrusted-content tracking through helpers, resumed conversations and hand-offs, value-based secret redaction and keychain storage for Slack and Matrix tokens.
• Desktop and station. The Windows installer is about 40% smaller (382 MB, down from 640 MB). The app no longer draws the station while minimized or in the tray on Windows, recovers from WebView2 crashes, and has a new icon. There are fourteen new agent skins and smoother agent motion.

New

Workflow lines & routines

• Workflow panel. Clicking a line's INBOX or a BAY opens one panel for the whole line, docked beside the live station, in place of the separate INBOX and BAY cards. It shows an editable line name, a plain "how it runs" sentence, a readiness pill and the cost of tested steps.
• The panel's flow strip follows the line's real run order, with loop back-arcs, and + inserts a real BAY between two stages as one undo step. Bays the INBOX never reaches are set apart as not connected (grouped under NOT CONNECTED when there are several). Each BAY reads as GETS → DOES → HANDS OFF → TO, with role starters, an agent picker, RECRUIT and ⊕ ADD A WORKSTATION HERE, and what you write under HANDS OFF is added to the brief that bay's agent receives.
• Agents appear as their own skin portraits throughout the panel. Clicking part of a line on the floor selects it in the panel, and selecting it in the panel highlights it and pans the floor to it. Opening the panel minimizes the Build Library, which comes back when the panel closes.
• Step test. ▶ START STEP TEST walks a line one stage at a time. With PAUSE AT EVERY HANDOFF you can CONTINUE, or CONTINUE WITH MY EDIT to change the exact text the next bay receives (UNDO MY EDIT takes it back). You can also ↻ RE-RUN STEP, ↺ RE-RUN FROM an earlier agent, or ▶▶ RUN TO END. Try this step (▶ RUN THIS STEP) runs a single bay with a test input, and E-STOP stops test runs.
• A paused step test follows floor edits, so a BAY added mid-test is used when you continue, once it has an agent. It previews the exact text the next bay will get, and says where and why a test stopped short of the OUTBOX. Test runs count as TESTS rather than jobs, and their spend counts toward the line's LINE BUDGET. If StarNet restarts mid-step, a cost the step already recorded still counts; otherwise the test says its spend is unknown.
• One agent, several bays. The same agent can now crew more than one bay; the old DUP AGENT fault is retired. Each bay gets its own brief, and work addressed to one bay rides past the agent's other bays. An agent without a desk uses the tools of the room its current bay is in; an agent with a desk keeps its desk room's tools at every bay.
• A multi-bay agent stands at its home bay, the bay with its current work glows, and its other bays read "NAME ↔ home". The agent picker marks an agent that already crews another bay and names that bay. If a desk-less agent's bays sit in different rooms, the floor suggests PLACE A DESK — TOOLS FOLLOW THE DOCK.
• Working folder (#25). A line's INBOX can pick a trusted project folder, and every stage of the line works in it, including chat-channel runs, RUN A SAMPLE JOB, scheduled routines and Run Now. Each stage re-checks that the folder is still trusted before calling a model. Changing the folder is one undo step, and the choice survives restarts.
• Line triggers. A line's INBOX offers ⊕ WATCH A FOLDER and ⊕ ADD A WEBHOOK. A file dropped in the folder, or a POST to the line's webhook, starts one run of the line. The file's text or the request body, up to 32 KB, becomes the job. For a binary file (such as a PDF, Office document, image or archive) only its path and size are sent, and webhook bodies over 256 KB are refused.
• Each trigger allows 20 fires an hour by default (up to 120), and queues up to five items behind the one running. Before each fire, and again when a queued item starts, it checks the line's daily spend cap, the station E-STOP and that the line is still on the floor.
• A watched folder must be in your home folder or a trusted project. It can never be a system folder, a drive root, StarNet's own data or a folder inside a line's working folder. A file fires once its writes settle, and temporary files and dot-files are ignored. A file that has fired does not fire again while its name, size and modified time are unchanged, even if it is moved away and back; an edited or renamed file fires again. Unreadable files are retried with growing waits, and a vanished folder shows as blocked until it can be read again.
• A webhook's key is shown once (COPY KEY, then I SAVED IT), stored only as a hash, and replaced with NEW KEY; COPY URL copies the webhook's address. The webhook listens on 127.0.0.1, so calling it from outside your computer needs a tunnel you set up yourself; email triggers are not offered yet. Each trigger row shows its fire count, last fire, last outcome and last error, with PAUSE, EDIT and DELETE.
• Line watch. Every crewed bay wears a status lamp with a hover plate: IDLE, WORKING, WAITING, FAILED or PAUSED (step test). FAILED stays lit until the next success or until you acknowledge it, even after a reload.
• Clicking a crate riding the belt opens a card with its job, line and route. It shows the run working on it (with live cost) or its outcome and failure reason, with a link to the agent's logbook (and, for a step-test crate, to the Workflow panel).
• Each line's INBOX carries a one-line plate of today's runs; hover for spend against the daily cap and the median run time. The Workflow panel header shows a TODAY row: runs, shipped (reached the OUTBOX with proven work), failed, tests and median run time since local midnight, and $ today (a UTC day).
• Routines. ✎ EDIT TASK in AUTOMATION → ACTIVE ROUTINES renames a routine or changes what it should do (✓ SAVE CHANGES). Only the fields you changed are sent, so an agent's newer edit to the other field is not undone. Routines on a clock-time schedule show their time zone, and a routine with results waiting to be delivered shows how many and when the next retry is.

Models & providers

• Reasoning levels for Grok (xAI), OpenAI API and DeepSeek. The model dock offers the levels each model accepts, instead of showing OFF while the model ran at its own default. Grok and DeepSeek levels come from xAI's and DeepSeek's model lists, and OpenAI API levels from a table built into StarNet (models not in the table show AUTO). A model with no reasoning setting is labelled AUTO, or OFF if its provider lists it as not reasoning.
• StarNet Managed models that support reasoning offer reasoning levels too.
• When OpenAI refuses a reasoning level StarNet sent, StarNet reads the accepted levels from OpenAI's error message, rese...

StarNet v0.12.4

StarNet v0.12.4

Full update notes: https://github.com/androoAGI/starnet/releases/tag/v0.12.4

Projects now bring their orchestrator conversation, crew, and delegated work together in COMMS. This release also adds selected Google file access and optional Windows accessibility computer control, alongside extensive conversation, station-saving, and recovery improvements.

Projects and crew coordination

• Open a project into one persistent orchestrator conversation with the project's working context.
• Choose and save a preferred crew for each project.
• Follow delegated work inside COMMS, with inline status updates and expandable activity details showing the agent, task, result, tools used, outputs, and directions.
• Send a direction to a working agent or request that its work stop from the activity view. Queued directions remain visible and are included in the orchestrator's review.
• Keep the station visible while moving between project conversation, crew, and activity views.
• Improve project control spacing, activity readability, and consistency with COMMS' glass appearance. Status announcements occur when an update actually changes.
• Fix saved projects getting stuck during loading because of an undefined session status.
• Preserve crew controls when reopening projects, keep activity order stable, and prevent duplicate worker controls during refreshes.
• Keep transcript navigation out of project detail views. Failed or revoked project opens no longer relabel the previously open conversation, and late responses cannot pull focus back after navigating away.
• Show unavailable project context and activity-refresh failures explicitly, retaining the last confirmed activity while reconnecting.

Overseer and delegated conversations

• Coordinate background workers from existing COMMS orchestrator conversations, with durable worker conversations and results returned to the originating conversation for review.
• Restore worker conversation history and project context when continuing or resuming delegated work.
• Reconcile delegated conversations and completed results without switching the conversation you are reading.
• Deliver review summaries without exposing internal task labels in the conversation.
• Refresh queued conversation context before execution and retain compatibility with legacy session dispatch and the first-save bridge.
• Keep automatic coordination scoped to COMMS orchestrator runs. Stop, halt, emergency recovery, and update handling respect coordinator boundaries.
• Refuse new overseer work if its durable halt state cannot be saved, and preserve worker/review state through restart and failure paths.

Conversations, attachments, and COMMS

• Restore conversation history before sending a new message, including when history arrives slowly or a conversation is reopened while a read is pending.
• Keep unsent messages when history cannot be restored and report the failure instead of sending without the expected context.
• Discard stale history responses after switching, clearing, or removing a conversation; preserve drafts and conversation identity through reconnects.
• Require confirmed completion before marking a streamed reply complete. Interrupted replies retain their partial content instead of appearing successfully finished.
• Drop files anywhere across the chat panel, with a visible attachment target, multiple-file handling, previews, and the existing attachment limits.
• Handle cancelled drags, unavailable files, disabled chat, and folder drops. Accidental drops outside an attachment target no longer navigate away from the station; normal text and link editing is preserved.
• Adjust COMMS layout and expand/restore behavior for narrow and short windows while retaining drafts and panel sizing.
• Retain bounded conversation-continuity and page-error diagnostics to help investigate interrupted sessions without recording message contents in the continuity trace.

Station building and movement

• Autosave station design changes while editing, before leaving build mode, including bursts of changes without losing conversation state.
• Reuse the existing environment render during object placement to avoid unnecessary floor rebakes in populated stations.
• Let agents yield, step aside, or retreat through narrow hallways so opposing traffic can reach its destination; improve following distance and handling around seated and working crew.
• Keep the website demo aligned with the station, placement, COMMS, onboarding, project, and reliability improvements.

Google files and connector storage

• Add Selected Google files in Abilities: use Google's file picker to grant access to specific Docs and Sheets through the limited drive.file permission.
• Provide 11 tools for accessible-file metadata/export, reading/creating/editing Docs, and reading/creating/editing Sheets. This is not whole-account Drive search; broad Drive, Docs, Sheets, Gmail, and Calendar connections remain deferred.
• Use CHOOSE GOOGLE FILES on a connected service to select more files. Cancelled, empty, or broader-than-supported grants preserve the previous connection.
• Require a durable refresh grant and recheck the limited permission after restart and refresh; refreshing cannot silently widen access.
• Encrypt desktop connector credentials using an encryption key backed by the operating system credential store, with migration from existing plaintext connector state.
• Preserve original credentials and recovery copies when migration, writing, key access, or JSON validation fails. A locked or malformed credential store reports an error rather than replacing the original key or state.
• Export encrypted connector settings through the running app with its credential store unlocked, while removing credentials and machine-specific permissions from recovery exports.
• Clarify privacy disclosures: encrypted credentials do not encrypt retained file contents. Disconnecting removes local credentials, not Google's grant or previously retained work.

Windows computer control

• Add optional CUA · Accessibility under Abilities → Computer Control for Windows x64, alongside Windows · Classic and Off. Installing CUA does not enable it; the selected driver persists across restarts and applies to new runs.
• Give authorized agents app/window discovery, screenshots with accessibility elements, semantic and pixel input, menus, window positioning, clipboard actions, and bounded state verification.
• Verify the downloaded driver's hash, version, and helper before activation. Check Driver and Reinstall CUA support diagnosis and repair; a failed replacement download preserves the current installation.
• Isolate native control sessions per run, serialize their actions, and close them on completion, cancellation, transport failure, or driver changes. Applications opened for you remain running after the agent's session closes.
• Recover expired sessions on a subsequent call, reject stale element references, and report refused or unverified actions without automatically repeating an ambiguous mutation.
• Keep driver telemetry disabled and provider credentials out of its child environment; report cleanup failures without logging desktop content.
• Native CUA control is currently Windows x64 only. It does not add always-listening audio or change voice permissions, and restricted runs cannot start the native driver.

Saving, permissions, routines, and recovery

• Preserve the last known saved station when reads fail or save acknowledgements are missing; distinguish local restore failures from a disconnected backend.
• Strengthen durable-store recovery so read/write failures cannot masquerade as empty state or confirmed changes.
• Serialize permission changes and reject stale responses, preserving the latest confirmed grant, revoke, or bypass decision.
• Validate routine, messaging, and other asynchronous control acknowledgements before showing success.
• Keep marketplace skills, project, connector, routine, and run-history reads from treating service failures as verified empty results. Disabled or unauthenticated connectors no longer count as ready capabilities.
• Preserve spending authority through failed writes, incomplete or truncated ledger reads, and hard restarts. Uncertain accounting no longer appears as a trustworthy zero balance or silently restores permission to spend.
• Allocate limited routine execution slots to the oldest due work to prevent starvation.
• Elect a single writer during concurrent stale-workspace recovery using immutable ownership generations; protect whole-root recovery and workspace activation from takeover races and crash windows.
• Refuse recovery when process ownership cannot be established, including ambiguous routine-lock ownership, rather than assuming another process is gone.
• Improve queue and browser-probe cleanup so failures do not strand pending work or timers. Isolate seeded lifecycle and recovery-test workspaces to avoid cross-session interference.

Desktop compatibility, packaging, and maintenance

• Bundle the shared specialty catalog with the desktop frontend and validate staged boot assets.
• Replace regular-expression lookbehind and newer timeout-API assumptions that prevented older macOS WebKit from parsing or running affected chat, widget, and history paths; preserve escaped-pipe table rendering.
• Improve onboarding sign-in-code contrast on dark glass in the desktop app and website demo.
• Stage production dependencies in desktop bundles and update Sharp and archive dependencies to patched versions.
• Expand regression coverage for persistence faults, ownership races, spending uncertainty, conversation continuity, project recovery, file drops, responsive layouts, native control, and credential migration.
• Improve QA isolation, screenshot readiness, shared-host gate deadlines, installed Mac version receipts, and upgrade comparisons that account only for the new...

StarNet v0.12.3

Download

Platform	Installer
Windows (x64)	StarNet_0.12.3_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.12.3_aarch64.dmg
macOS (Intel)	StarNet_0.12.3_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.12.3

This update remasters the station itself: industrial materials, a redesigned command bridge, a furnished default station with every essential ability already installed, a browsable Build library, remastered agent skins with grounded walk cycles, and an onboarding that goes from what you want to do straight to a first useful task. It includes fixes and hardening for customer-reported update, boot, rating, delegation and conversation-history problems reported since 0.11.2.

These notes cover the changes shipped since 0.11.2. The 0.12.0, 0.12.1 and 0.12.2 candidates were held before publication; this final build includes the complete overhaul and the final sprite-animation merge.

• Close-to-tray: finishing startup or reloading the document no longer reopens a window you closed to the tray. Initial display waits for the completed document; explicit tray Open and a second launch still reveal the app.

• Saved model choices: a temporary catalog error or offline fallback list no longer clears your selected model, including during startup and refresh.

• Rebuilt walk animations: 33 redesigned skins now use eight-frame alternating walk cycles in all eight directions, with corrected manifest pointers. Capybara retains its earlier walk frames; the retained original skins are unchanged.

A remastered station

• Industrial station materials: walls, floors, ceiling coping, corridor doorways and shells are rebuilt in a darker industrial style with orbital shells, clean station floors and consistent structural frames. Wall heights, doorway reveals and corner joins are aligned across every room shape.
• Redesigned command bridge and props: the workstation, tactical table, capability equipment (terminal, files cabinet, comms dish, memory core, media studio), lounge furniture, storage, lab and workshop machinery are re-authored at station scale with readable forms, fitted screen effects and truthful activity (screens, mechanisms and indicators follow real harness state).
• Pre-equipped default station: a new station starts as one furnished room with a workstation, all five essential ability props and plants, so a fresh Commander can hand over a first task without placing equipment. Seven station presets (DEFAULT, QUIET RETREAT, COZY WORKSHOP with a connected conveyor line, CREATIVE STUDIO with a prepared draft-and-review line, RESEARCH, ENGINEERING and OPERATIONS stations) can be applied from Build and the previous station restored.
• Build library: Build Mode is a spacious, searchable prop library sectioned into FURNITURE, EQUIPMENT and ABILITIES with category menus, previews, clear placement/cancellation feedback, inline move cancellation and keyboard Escape. Build selections now match the visible artwork.
• Remastered agent skins: 37 finalized skins with grounded strides, corrected cardinal facings, natural turns, talking and sitting motion, workstation seating with proper foot occlusion, and layered couch seating. Legacy skin selections keep working.
• Movement and lighting: agents route clear of raised doorway walls, keep a natural travel pace, reconcile their walk after collisions and plant correctly when stopped. Station lighting pools, fixture illumination, back-wall light and exterior navigation lights are rebalanced; the shipped CRT look is preserved.
• Renderer recovery and performance: silently lost station canvases are recovered and remaster materials restored before a re-bake; the canvas watchdog no longer performs synchronous readbacks; lighting caches are reused. On the 96-prop stress layout the integration measurements recorded frame intervals of 8 to 15 ms at p95 with zero render faults.

From an idea to a result

• Goal-first onboarding: setup follows what you want to do, with a cinematic agent arrival and a first-task handoff. The optional quick tour explains the equipment already in the station and never asks you to place props; a typed first task survives the tour and returns as an editable draft.
• Personal journeys: WORK → QUESTS connects ambitions, plans, next actions and outcome reviews (Now / Goals / Progress / History). Planning suggestions can be reviewed, adopted, edited and undone; drafts survive refreshes and delayed saves; finished steps link to the exact run output.
• Creative Studio guidance: a guided setup and verified sample workflow for the Studio, with a prepared draft and review conveyor in the furnished preset.

Everyday improvements

• Compact speech bubbles and fewer redundant delivery acknowledgments.
• Conversation history stays in order: refreshing or retrying a conversation no longer reorders or rewrites the local transcript (the "my turns vanish / an old message moved to the front" report).
• Composing stays in your chosen session: background workshop events and delayed connector checks respect the focused composer, unsent drafts and attachments. Explicit session navigation continues to work.
• Corrections replace old memories: explicit memory corrections preserve an audit history while keeping one current belief in recall. Approved requirements can stay pinned within their project or conversation, and stale reviews cannot overwrite newer edits.
• Incomplete work is explained: repeated promises without tool actions end with an explicit incomplete-work error. Unsupported claims that a preference was saved are checked against actual notebook-write receipts.
• Saved files stay clickable: the "▤ saved " and media rows an agent produced are replayed when a conversation is reopened, switched to, reloaded or retried, instead of surviving only in the model's prose.
• XP and ratings: earned XP refreshes in open views, skipped rating credit can be recovered, and interactive replies in a conversation that also contains scheduled routine activity can be rated on their own run record. Ineligible replies now say exactly why.
• Secondary panels render correctly at enlarged UI scale and narrow widths; provider text and dropdown affordances are readable; CRT curvature is accounted for in hover and click targets.
• Static CRT settings (0–200%) persist across reload independently of theme and glass.

Reliability and support repairs

• Boot guard self-heals a slow engine: if the desktop window loads before the station engine answers, the shared catalog is retried with backoff and the page reloads once on success instead of painting "STATION FAILED TO BOOT" (the shared/specialties.js reports on Mac and Windows). A genuinely broken file still shows the banner, now with the retry ledger for support.
• Account unlink recovery: unlink and status-read failures stay visible with a retry and a working LINK STATION path instead of silently hiding the controls.
• ChatGPT sign-in at setup: a live ChatGPT sign-in takes precedence over an old saved OpenAI API key when the key field is blank. A key you explicitly type still selects the API route, and connection errors identify the route used.
• Browser element discovery: the internal browser now finds JavaScript-driven cards, ARIA targets and open-shadow controls (the campaign-listing report).
• Telegram and channel leads receive the crew briefing matching the delegation tools they were actually granted; workers cannot delegate recursively.
• Local models: small-talk turns omit task-only prompt material and per-run ids ride last in the prompt so caches survive; Ollama output has a configurable ceiling and capped casual replies stop without automatic continuation.
• Response safety: auxiliary budgets and responsive fallback cancellation are preserved; durable unload saves are recognized when their acknowledgement was lost; the tutorial activity guard stays scoped to orientation.
• Delegated specialists keep access to connected MCP tools; cross-provider fallback keeps the selected route; task cancellation, review retries and routine/session binding repairs from the 0.11.2 follow-through are included.

Upgrade notes

• Existing stations keep their layout, props, crew, conversations, routines, goals and settings. The furnished presets never replace a saved station; they are applied only from Build and can be restored.
• Update from SYSTEM → SETTINGS → UPDATE CENTER, or download the installer for Windows or macOS from the release page. The desktop shell clears the stale WebView cache automatically on a version change.
• The installer is substantially larger than 0.11.2 because it includes the new station and character artwork. Review-only art is excluded; the textures needed to enable the remaster are included.
• The customary 48-hour attended soak was waived for this release. Candidate-specific installed smoke, scripted reliability and responsiveness receipts are recorded with the release preparation evidence.

StarNet v0.11.2

StarNet v0.11.2

Released September 11, 2026. This update expands agent personality and setup, refreshes the station's worlds and walking animations, improves report readability, and repairs save, provider, delegation, graphics and desktop-update behavior.

These notes cover the changes shipped since 0.11.1, including the final CRT static controls and Windows pointer repair.

Station visuals and performance

• Six rebuilt backdrops: THE NURSERY gains layered gas clouds, filaments, dust and twinkling cores; NIGHT CITY gains a lit street grid, building windows, rivers, bridges and beacons; OCEAN gains stepped waves, crest lines, reflections and pixel clouds; THE BELT gains detailed small asteroids and dust; THE MOON gains richer regolith, crater rims, ejecta and rocks; FOREST gains a dark aerial canopy, conifers, undergrowth, forest-floor detail and a river.
• Backdrop lag repair: heavy sky artwork is generated in a background worker, keeping the interface responsive while the scene is prepared. Ground scenery is cached rather than repainting every overlapping tree on every frame.
• Smoother camera movement: the ground cache moves with the camera and refreshes when zoom, window size or station footprint changes. Superseded artwork requests are discarded, with a compatibility path for environments without workers.
• Adjustable CRT static: Appearance settings now let you change and save static strength. The default remains 100%, preserving the existing look for stations that have not changed it.
• Smoother walking animations: expanded walk cycles add intermediate frames across agent skins, with turning cadence adjusted to preserve natural foot motion. Turtle rear-facing frames and wizard staff continuity receive specific corrections.
• More consistent glass styling: task-understanding, noticed/recommendation, away-work, decision and other transient cards now match the station interface. Station announcements are more readable, and compact recommendation labels remain compact.
• Panel sizing stays saved: docked panel height survives close/reopen and reload, with appropriate handling for maximized panels and smaller windows.

Overseer creation and agent personalities

• Rebuilt Overseer setup: a full-screen, two-step identity-and-connection flow brings the character gallery, name, personality, working style, provider, model and reasoning controls together.
• Clearer connection choices: StarNet account setup has a prominent branded action; alternative providers use recognizable local logos. Model selection supports keyboard navigation, custom endpoints and saved reasoning choices.
• Six distinct personality presets: Composed, Warm, Blunt, Dry, Unhinged and Upbeat now define more consistent behavior in conversation, disagreement, uncertainty, failures and successful work.
• Detailed personality tuning: adjust warmth, humor, formality, answer length and energy; choose no, occasional or frequent profanity; control emoji and directness; or write a custom communication style. Personality tuning is separate from the audible voice.
• Visible saved overrides: customized personalities are labeled, tuning survives preset changes, and Reset to Preset clears overrides. Older personality names map to the new presets. Unhinged's language choice retains an explicit confirmation.
• More reliable setup navigation: going back preserves identity, appearance and connection entries. Replaying the tutorial preserves the current agent name. Narrow-window layout fixes prevent character and identity sections from overlapping.
• Clearer setup recovery: stale provider-specific warnings and buttons retire when the connection, agent or missing setup step changes. A connected station without a model shows CHOOSE MODEL and opens the picker instead of asking for an unrelated API key. Recovery buttons work with both mouse and keyboard.

Providers, connections and discovery

• GitHub device sign-in: connect GitHub through its device authorization flow, with a visible code, expiry/cancellation handling and useful sign-in errors. Small windows keep the code accessible. If saving a new token fails, the existing connection is preserved.
• Discover improvements: popular platforms are easier to find, with clearer guidance into each platform's connection setup.
• Field Manual connection guide: work-app connections and messaging channels have dedicated setup guidance and navigation to the relevant panel.
• Provider presentation: refreshed provider logos, aligned connection statuses and clearer StarNet account actions make connection choices easier to scan.
• Model-selection race fixed: an older model-catalog response can no longer overwrite a newer provider or model choice, including overlapping picker requests and agent changes.
• Codex-to-OpenRouter fallback fixed: saved fallback choices now switch the provider route and credential source correctly when Codex reaches its quota. Paid fallback work retains its configured spending limits and cost record. Addresses issue #12.
• Claude continuation fixed: verification reminders and later host notes retain their proper place in the conversation on OpenRouter and compatible managed connections, avoiding the unsupported assistant-prefill request that could interrupt a task with HTTP 400.
• Delegated connector access fixed: specialists retain access to connected MCP tools through the lead agent's live approval flow, including dispatched, spawned and resumed work. Denials, connector removal and renewed confirmation after external content still apply. Addresses issue #13.

Saves, conversations and recovery

• Concurrent-save protection: stale clients cannot silently overwrite newer station edits. Conflicting local changes are retained for recovery, including edits made while conflict recovery is already open.
• Startup recovery: an unsynced local cache no longer prevents startup. Save-recovery actions use the station's own controls and retain the local recovery path.
• Queued-save restart repair: completed queued writes now retain the server's acknowledged revision, preventing false conflicts when the station restarts or updates.
• Conversation conversion preserves attachments: converting a direct conversation to a group retains historical attachment references and snapshots the required files before committing. If an attachment cannot be read, the original conversation remains intact.
• Conversion retries are safe: retrying the same conversion returns the existing result; a different request cannot accidentally reuse the earlier conversion's identity.
• Duplicate restored replies fixed: reopening a completed conversation reconciles locally combined streaming text with its durable turns instead of displaying both. Uncommitted text, attachments and stopped/error markers remain available.
• Loop review isolation: reviewing or undoing a loop iteration is protected against simultaneous execution, another review, or changes to the same loop, including while Git undo is still running.

COMMS, reports and task feedback

• Structured reports in COMMS: tables, ordered and nested lists, quotations and named links now render alongside headings, emphasis and fenced code.
• Readable narrow reports: wide tables stay within their own scrolling area, and table text remains readable at narrower panel widths.
• Exact report copying: copying a message preserves its original report text and Markdown; code-block copying retains its separate behavior.
• List-format fixes: plus-sign bullets and tab-separated list markers now render consistently during normal report display and restored history.
• Clear failed-session marker: failed sessions use a flashing X without repeating the same failure label beside it.
• Truthful partial results: failed or interrupted tasks retain the text, usage and evidence already produced instead of being presented as successfully completed work.

Crew movement, furniture and station behavior

• Doorway routing repaired: agent footsteps follow traversable doorway paths instead of cutting through adjacent walls or floor boundaries.
• Continuous movement and facing: waypoint transitions remain continuous, and look-back facing is held correctly instead of snapping prematurely.
• Unreachable desk recovery: agents wait in place while real work remains active, retry routes and recover when a refit restores access. Moving a desk no longer teleports the Overseer to it.
• Correct desk-facing poses: seated crew face their desk, using an appropriate directional pose when a skin lacks matching typing art. Working agents without a seat remain standing.
• Furniture cleanup: moving, rotating or deleting furniture releases stale seating claims, offsets and leisure references. Unchanged couches keep their occupants when the station origin shifts.
• Retired crew releases seats: removing plan-derived crew frees its furniture reservation so another agent can use the seat.
• Room deletion stays coherent: intersecting furniture and belt tiles are removed in the same undo snapshot, while neighboring furniture and logical agent links are retained. Undo, redo and reload preserve the intended result.
• Approval waits are visible: crew awaiting permission stand with the waiting-for-approval state and suppress work effects. Matching responses, concurrent prompts and terminal events clear the correct wait.

Images, usage and speech

• Image costs reach run receipts: actual image charges are included in local usage and completed-run costs without charging a second time.
• Image cancellation stops publication: cancellation propagates through generation, download and processing so a later result cannot publish a file after cancellation. Up...

StarNet v0.11.1

Download

Platform	Installer
Windows (x64)	StarNet_0.11.1_x64-setup.exe
macOS (Apple Silicon)	StarNet_0.11.1_aarch64.dmg
macOS (Intel)	StarNet_0.11.1_x64.dmg

These downloads are mirrored byte-for-byte from the validated StarNet distribution release. Already running StarNet? Update from inside the app; your crew, sessions, keys, and station remain in place.

---

StarNet v0.11.1

This update improves linked-account reliability, voice playback, and everyday station controls.

• Fixed managed chat using stale per-request server settings or credentials instead of the linked StarNet account. Custom providers keep their own connection settings.
• Fixed image generation rejecting a linked StarNet credits account. Ordinary wording is less likely to trigger an unnecessary image-generation requirement.
• Saved room equipment now remains available when a chat request omits placement data, and tool availability descriptions better match the actual grants.
• Spoken replies buffer complete sentences and recover audio in order, with clearer interruption diagnostics.
• Background replies preserve the conversation you are currently viewing. Inbox rows have more readable message previews.
• Fixed work ratings on resumed legacy stations, retaining the station identity through restart.
• The glass interface is now standard across COMMS, model selection and station windows. Window interactions, text readability and Appearance previews have been improved.
• Recipe Bay now offers 45 curated workflows with a simpler launch flow. Recruitment, abilities, settings, profiles, quests, channels, automation and task outputs have clearer menus and controls.
• File-change approvals and work feedback provide clearer action and run references.
• Station lighting, wall finishes and floor materials have been refreshed. New starter stations use a white shell, and prop details leave more room for browsing the build catalog.
• Global stop controls have been removed, with recovery for their legacy pause state. Individual work controls remain available.

Update through Settings > Updates > Update Center. After restarting, confirm that the installed version is 0.11.1.

Google Workspace connections remain deferred while Google verification is completed. Saved connections are retained. Google login for StarNet billing is separate.

Release verification and known limitations

Windows upgrades from 0.10.13 and 0.11.0, clean installation, and close/reopen behavior passed. Mac signing, notarization, and Intel installed-app acceptance passed.

This urgent release proceeded with the release owner's extended-soak and aggregate-readiness exceptions. Existing customer reports remain under investigation; the included fixes do not establish recovery for every affected account or device.

Complete source changes · Official signed build