Robin v0.3.1

Quitting Robin stops its server. After you quit the app, Robin's built-in server could keep running in the background, with the agents it had started. Quitting now stops it and everything under it, and if the app ever closes without stopping it, a crash or a forced quit, the server notices within seconds and stops on its own.

Windows 10 and 11, macOS 12 or later, and Linux x64. Robin updates itself from 0.3.0 and 0.2.x. The installers aren't signed yet: on Windows choose More info, then Run anyway; on a Mac open System Settings, Privacy & Security, and choose Open Anyway.

Robin v0.3.0

Your coding agents, working as one team. Say what to build. An agent plans it, and nothing is built until you approve. Then Claude Code, Codex, Gemini CLI and the others you use each take a task in its own branch, agree on how their pieces fit, review each other's changes across vendors and bring you only the decisions that are yours.

Goals, from start to merge. Type a goal, or paste a GitHub or Linear issue. Watch the planner read the project, then read the plan: each task with its agent, what it touches, what proves it done and what it is likely to cost. Edit any of it before you start, or the tasks no agent has picked up while the work goes on. When a goal closes, its recap says what was asked, what each agent did, what was agreed and what was learned.

A team, not a crowd. Agents on a goal share a thread. They ask each other questions, agree on contracts that are pinned for everyone, and a question nobody answers goes to the lead, then to you. Their team tools never ask your permission. Write to one agent or to all of them; they carry on when you leave, and "since you left" says what happened.

Reviewed by another vendor. Every change runs the project's own checks, then an agent from a different vendor reviews it. Its issues go back to the builder, and a second look confirms the fixes before the change reaches you. Review it beside the goal: the summary, the checks, the diff, the cost. Merge it, open a pull request, or send it back with a reason.

Questions with choices, from any agent. Codex's own questions, an MCP server's forms, and a new ask_person tool for every other agent arrive in Needs you, on your desk and on your phone, with their options. Your answer goes straight back to the agent.

Issues in, pull requests out. A GitHub issue labelled robin, or a Linear issue, becomes a goal; Robin keeps one comment on it up to date and closes it when the work is merged. Every pull request says who asked, which agent wrote each commit, who approved what, what the review found and what it cost.

It learns once. What the agents learn, a command that failed and the one that worked, your corrections and their agreements, goes into the project's AGENTS.md in the same change. Send a change back for how it looks, and your reason becomes a rule every agent on the project follows from then on.

Your settings and your config stay yours. A project's settings live in its repository (.robin/settings.json): protected files, the order of agents, a default budget and its own standing answers. Robin writes nothing into your checkout and nothing into your agents' own configuration; each session it starts gets its own. Standing answers now also apply to agents you start in your own terminal.

Budgets that hold. A goal's budget counts work on a subscription plan at what it would cost on the API, and says so; at 80 % you are told, at 100 % its tasks stop before their next step until you raise it.

The phone runs goals. Start a goal, read and start its plan, follow the thread, answer questions with choices, review a ready task and merge it or send it back, and raise a budget, all from your phone.

Run it where you like. robin serve runs Robin without the desktop app, on your machine or your own server, paired to your phone. The command line speaks goals: robin goal new, start, status.

Also in this release

• A message or a "Tell every agent" sent while an agent works reaches it at its next step, without interrupting it.
• Installers for macOS (Apple silicon and Intel) and Linux (AppImage and .deb), and every installer is far smaller.
• Escape no longer leaves a goal while a dialog or field has it; the review button goes away once a change is merged.
• Light or dark follows your system until you pick one in the top bar.

Windows 10 and 11, macOS 12 or later, and Linux x64. Existing projects, paired phones and agents carry over, and Robin updates itself from 0.2.x. The installers aren't signed yet: on Windows choose More info, then Run anyway; on a Mac open System Settings, Privacy & Security, and choose Open Anyway.

Cockpit v0.2.1

Agents pick up where they left off. After a restart, an update or an agent that exits, each embedded agent reopens its own conversation with its harness's resume command (claude --resume, codex resume, gemini --resume and the others that support it), in the same permission mode, without sending its first message again. An agent that stops shows what happened, with Resume conversation next to Restart. A pane you closed stays closed.

Cockpit learns from what went wrong. When an agent's command fails, Cockpit records it and finds the command that fixed it. The next agent that hits the same failure is told what worked, at the moment it fails. A failure that recurs across sessions becomes a project gotcha every agent reads in the brief. Supported through Claude Code's and Qwen Code's failure hooks and Gemini CLI's tool results. Everything stays on your machine; secrets and your home folder are stripped before anything reaches the project memory.

Memory you can see. The Brief has a Memory tab: the decisions, gotchas, preferences and facts agents saved, and what Cockpit learned, with who added each one and the evidence behind a lesson. Forget one and agents stop seeing it.

Keyboard and screen. One focus style across the app: a thin outline that follows each control's shape, drawn inside list rows, and the field border itself for text fields. The session tabs and phase strip move with the arrow keys. Plain words everywhere: no session ids or internal names on screen, no clipped text, and the session page reads in sentences.

Also in this release

• Your Anthropic API key for the Explain lens is stored encrypted with Windows' own protection instead of in plain text.
• Recordings older than your "Keep recordings for" setting are now removed, as Settings says.
• A hardened desktop shell: one check for every external link, no script evaluation in the production build, and a crashed window reloads itself.
• Menus and lists that are taller than the window scroll instead of running off screen.

Windows 10 and 11, x64. Existing projects, paired phones and connected agents carry over, and Cockpit updates itself from 0.2.0. On first run Windows may show a SmartScreen prompt: choose More info, then Run anyway.

Cockpit v0.2.0

Cockpit now works with 12 agent harnesses: Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot CLI, OpenCode, Qwen Code, Factory Droid, Amp, Cline, Goose and Aider.

One plan for every agent. Agents claim a priority, phase or sub-step under a lease, hand off with notes, and share project memory, through a 37-tool MCP server and the cockpit CLI.

One inbox. Questions, plan approvals and permission prompts from every session land in one place, each permission prompt with the exact command or file change it asks about.

Launch from the drawer. Start any installed agent primed on a sub-step; the first message is typed for you. "Ask before tools" now makes the agent ask even when its own settings default to skipping approvals.

Environment check per harness. Onboarding and Settings show which agents are installed and connected, checked in your login shell, with each agent's own install steps when one is missing.

Redesigned phone app. Inbox; a permission screen with Deny, Reply instead and Slide to approve; Agents with live terminals and replays; start an agent from the phone; Plan with editing; Settings; pairing and lock.

Also in this release

• Photos on new priorities, shown on the desktop and the phone.
• Undo on the phone restores what was really there, even if the plan changed in between.
• A phone answer the desktop confirms late shows as unconfirmed, never as failed.
• A revoked phone stops showing cached data and asks to pair again.
• Reconnecting an agent keeps your own settings on its Cockpit server entry.
• Stability and diagnostics fixes, including everything from 0.1.20 to 0.1.23.

Windows 10 and 11, x64. Existing projects, paired phones and connected agents carry over. On first run Windows may show a SmartScreen prompt: choose More info, then Run anyway.

Cockpit v0.1.23

First-run polish: launching a session on a machine WITHOUT Claude Code now shows a clear install guide right in the terminal (native installers, no Node.js required) instead of a raw shell error; and Pair-your-phone explains how to enable MagicDNS when your tailnet hasn't assigned a DNS name, instead of silently doing nothing. Includes the v0.1.20-0.1.22 stability + photos work.

Cockpit v0.1.22

Polish: the mobile photo viewer's close button now pins to the top-right on every device (safe-area handling inside portals). Includes everything from v0.1.20 (self-healing integration, resilient mobile transport, clean uninstall) and v0.1.21 (photos on New Priorities, desktop + mobile).

Cockpit v0.1.21

Cockpit v0.1.21 — Photos on New Priorities

A mock, a whiteboard shot, a bug on screen — sometimes the photo IS the brief.

Desktop

• Attach photos to a New Priority: drag & drop onto the dialog (with a live drop glow), paste a screenshot straight from the clipboard, or pick from disk. A clean thumbnail rail with hover-remove keeps the one-field flow intact.
• Priority briefs now render images — attached photos appear as a tidy gallery in the brief; click to expand in place.

Mobile

• Create priorities from your phone — new bottom sheet with dictation, and photo tiles that open your camera or library. This is the first time priorities can be created from mobile at all.
• The Steer + now offers two verbs: New priority and New directive.
• Priority cards show a photo count; expand for a thumbnail strip and a full-screen viewer.

Under the hood

• Photos are validated (image-only allowlist, 8 MB/photo, 12/priority), stored next to the project's cockpit data with server-minted names, and served through an authenticated, traversal-proof route with immutable caching.
• Everything ships with the v0.1.20 stability layer: self-healing Claude Code integration, resilient mobile transport, clean uninstall.

Beta note: installers are not yet code-signed — Windows SmartScreen will warn on first launch (More info → Run anyway).

Cockpit v0.1.20

Cockpit v0.1.20 — Self-healing integration, resilient mobile, honest first run

Terminal / Claude Code integration

• Fixed the hook-registry corruption that flooded every terminal with Cockpit.exe: No such file or directory errors. Cockpit's Claude Code hooks are now recognized by content (not a marker Claude Code strips), installs replace old registrations instead of stacking them, and dead runner paths are pruned automatically.
• Self-healing on every boot: moved installs (per-user → Program Files), duplicate registrations, and stale runtimes converge to a single clean registration. A corrupt settings.json/.claude.json is never overwritten — Cockpit refuses instead of destroying your Claude config.
• Clean uninstall: removing Cockpit now detaches all Claude Code hooks, statusline, and MCP entries — no more error spam after uninstalling.

Mobile (Cockpit on your phone)

• Disconnects now heal themselves. The live streams re-assert their credentials before every reconnect (fixes the silent forever-401 after iOS evicts cookies), reconnect immediately on network handoffs (Wi-Fi ↔ cellular), and the terminal's reconnect backoff no longer gets permanently stuck at its ceiling.
• If your pairing genuinely expired, the app now says so plainly with re-pair instructions instead of "reconnecting…" forever.
• Steer recovers on its own even with a dead stream (background poll fallback).
• Fixed a server-side timer leak on every phone disconnect that slowly degraded the desktop app over a long day.

First run & reliability

• Onboarding no longer hands you an npm command your machine can't run — it now uses Anthropic's native Claude Code installer (no Node.js required).
• First-boot server failures are captured in the diagnostic log (previously invisible), the boot window tolerates slow disks/antivirus scans (60s), and a corrupted install fails with a clear "reinstall" message.
• The installer pre-authorizes Cockpit in Windows Firewall; the uninstaller removes the rule.
• Security hardening across the local API and mobile surface (auth guards, size caps, CSP, navigation pinning).

Note: installers are not yet code-signed — Windows SmartScreen will warn on first launch (More info → Run anyway).

Cockpit 0.1.19 — self-healing mobile updates

Same ALTIMETER mobile redesign as 0.1.18, plus a fix so the installed mobile PWA refreshes itself on every future update (service-worker cache is now versioned per release — no more stuck-on-old after an update). Includes the stable Cockpit-Setup.exe download alias.

Update, then open /m on your phone over Tailscale (add to Home Screen).

Cockpit 0.1.18 — Mobile redesign (ALTIMETER)

The entire mobile app (/m) was rebuilt into a new design system — ALTIMETER ("an altimeter, lit on OLED black"): hero-numeral mastheads, a slide-to-commit beacon for every consequential action, milled MILLED/FORGED/QUENCHED materials + grain on true black, editorial hairline rows, and one spring motion engine. Every screen reworked: Pulse, Fleet, Steer, More, DVR, Lock, Install, Offline, Session. Also folds in the latest electron-stability fixes.

After updating, open Cockpit on your phone over Tailscale at /m (add to Home Screen for full-screen).