Repository navigation
Releases
openajar 0.5.3
openajar 0.5.3 gives the project a home and makes it easy to add to any MCP app. The website is https://openajar.com, and installing is now one short line: curl -fsSL https://openajar.com/install.sh | sh. A new command, openajar mcp, serves openajar to any MCP app that starts a command (Cursor, Claude Desktop, the MCP marketplaces): it starts the local service when it is stopped, and on a computer without openajar it says what to install instead of hanging. openajar is on npm (npx -y openajar) and listed in the official MCP Registry as com.openajar/openajar. Nothing changed for an existing install: Claude Code and Codex keep working as before.
[0.5.3] - 2026-11-21
Added
-
openajar.com: the project's website (site/, a Cloudflare Worker with static assets), with the short install line
curl -fsSL https://openajar.com/install.sh | sh(a redirect to scripts/install.sh on main). -
openajar mcp: openajar as a stdio MCP server, for any MCP client that starts a command (Cursor, Claude
Desktop, the MCP marketplaces). It bridges to openajar on this computer, asopenajar mcp-bridge local
does for Codex, and starts the local service first when it is stopped. On a computer where openajar was
never set up, it still answers: its instructions and its one tool tell the agent what to run (the install
line,openajar setup, and that VMs need a Hetzner Cloud token or Docker), and it says the same on stderr. -
openajar is ready for three registries: PyPI (
uvx openajar mcp; the sdist now carries only the package),
npm (npx -y openajar, a dependency-free launcher in integrations/npm that runs the installed CLI, or
explains what to install when there is none) and the MCP Registry ascom.openajar/openajar(server.json).
glama.json names the maintainer for glama.ai. The publish workflow puts each release on all three by
trusted publishing, after scripts/check_release_versions.py checks that every copy has the tag's version.
Changed
- The unit tests no longer pass or fail on the runner's speed: they wait for what they expect for up to
10 seconds instead of a fixed sleep, a speed check scales its bound on the macOS runners
(OPENAJAR_TEST_SLOW in CI), and the store lock check fails only on a loop really stuck on the lock.
docs/DEVELOPING.md says how a new test waits and bounds a speed check.
openajar 0.5.2
openajar 0.5.2 is a small release on top of 0.5.1. Each agent now has one name everywhere, so the web app says "Claude Code" where it used to say "claude-code in claude-code 2.1.328". Running the install line again now restarts a running local service, so it runs the new version on the same home and port (before, the old service kept running the old code until it was stopped and started by hand). A stopped machine's page drops the Network level that did nothing there, and the browser fingerprint probe waits longer for the sound server on computers that have none. No MCP tool or setting changed; the one new command is openajar local restart.
[0.5.2] - 2026-11-21
Fixed
- The web app names an agent the way the desktop does: Claude Code, not claude-code 2.1.328. One rule on the
node and in the web app (openajar.core.agent_names) reads the grant and the MCP client as the product's name,
with the computer as words and the raw values under Technical details. - Updating with the install line restarts a running local service, so it runs the new version on the same
home and port (openajar local restart, launchd or systemd). Before, the old service kept running the old
code until it was stopped and started by hand. It says so in one line, and when environments are open, what
the restart interrupted: their machines and jobs keep running, but an open browser there opens again on the
next browser call, and a recording, a tool call in flight or an unused upload or download link ends. - A stopped machine's page no longer shows the Network level (Full, Limited, None), which did nothing there:
one line says what it was. - The browser fingerprint probe gives
navigator.mediaDevices.enumerateDevices()20 seconds instead of 8. On a
computer with no sound server (a CI runner image) its first look at the audio devices waited past 8 seconds
and the fingerprint test failed; the values it reads are the same.
openajar 0.5.1
openajar 0.5.1 is a small release on top of 0.5.0. When an agent's machine dies (its container is killed
or its server is powered off outside openajar), the node now notices it from the provider, every tool says so
with one next step, and opening the same title again starts a new machine instead of handing back the dead
one. An app opened on the desktop keeps the keyboard. The web app gets a simpler navigation (ten places, with
Undo for settings that save on their own) and the rest of its screens in the Dawn design, and it shows an
environment whose machine stopped. No tool or setting is removed or renamed; screen_act gains window with
type and key, and keys_to in its result.
[0.5.1] - 2026-11-21
An agent's environment whose machine dies is now noticed and replaced, typing on its desktop lands in the
app the agent opened, and the web app has a simpler navigation and the rest of its screens in the Dawn design.
Added
- Four more tasks in the eval suite, ported from evals/ (t12 to t15: the machine dies in the middle of a
build, a file from the person's folder up to the machine and back, a report that takes seconds and a confirm
dialog, a price past a catalogue's first page), a harness step that runs beside the agent (t12's
docker kill), the person's own folder for a task, and a paid run on the person's ownclaudesign-in. docs/EVALS.md has the baselines and
what 0.5.1 changed, andbench/evals.jsonlkeeps the runs. - The web app shows an environment whose machine stopped. It leads the Open list and Home's Working now
with a red Machine stopped chip on the dimmed desktop; its page says when it stopped and why, and that the
agent starts a new machine when you say "continue <title> in openajar", and Close is what is left (Extend
goes). The list's environments carrymachine_gone(when, why) for it.
Changed
- The environment skill, the server's instructions and the open's
nextsay the defaults in so many
words: when the task is done and nobody asked to keep the machine open,environment_close()is the
last call, whatever thekeep; files are read and changed with the file tools, nevercatorsed
in a command; andmachine_gonemeans opening the same title again. screen_actwithtypeorkeysays where the keys went (keys_to, the window that had the keyboard),
and takeswindowto bring that window to the front first.- A simpler navigation in the web app: 13 places become 10. Home, Needs you, Computers and What happened
come first; then Your business (Accounts, Jobs that repeat, Brand), For developers (Runs and checks, Site APIs,
each hidden until first used) and Settings. Computers holds Environments and Machines as two tabs, Jobs that
repeat holds Routines and Email, Runs and checks holds Runs and Checks, and History is now called What
happened. Every old address still opens the same page, and the phone's tab bar is Home, Needs you, Computers,
What happened and More. - A setting that saves on its own now says what changed in a toast with Undo ("Quiet hours turned on"),
instead of Saved in the card's footer. Undo puts the old value back the same way it was saved; it never loosens
what an agent may do without asking, and what cannot be undone keeps its confirm dialog. - Accounts and one account in the Dawn design. Each workspace is a heading with one banner for what needs a
look, and its accounts are cards with the site's tile, the handle, a state chip, what it did last and Ask me
first. An account's page opens on the night band with the site's tile, its state and the live screen as a
stage with the control pill, and glass panels for who is using it, what happened and whether it is safe. The
workspace page, the login queue, the vault, Workers, Add an AI service (one page) and Proof follow the same
design, and an address with no such account says so in plain words. - The detail pages in the Dawn design: one run, one check, one machine, one routine, a new routine, Email,
one Site API, memory and standing approvals each open with their kind's tile, the title, the state as a chip
and one line of facts, and keep their skeletons in the page's shape while they load. - The phone's More opens a sheet from the bottom with the places as tiles in the sidebar's groups and the
machines' line and the organisation at its end, above the tab bar, which stays in reach. - What an agent did on an environment's desktop reads as what its hands did (typed, pressed keys, brought a
window to the front), and ascreen_acttype or key withwindownames that window in the call's line.
Fixed
-
An app opened on the desktop keeps the keyboard. The environment's browser starts in the background
when it opens, and its first window could come up after an app the agent had just opened with
screen_open(LXTerminal, say) and take the focus, so the agent'sscreen_acttyping went into the
browser. A browser window that comes up now takes no focus by itself: the app that had it stays in front
until the agent uses the browser, a person clicks it, or the dock's Browser is pressed, and a browser on a
desktop where nothing has focus still gets it.screen_opengives the app's window the keyboard before it
returns. -
A machine that stops is noticed, and said once, everywhere. When an environment's container or server
stops for good (killed, powered off, deleted outside openajar), the node finds out from the provider
(docker inspect, Hetzner's API, asked twice; never from one failed ssh call, which a network blip also
gives) and then: every call that needs the machine fails with the new errormachine_gone, whose
message says when it stopped, why, and whether a save exists to resume from;environment()shows
state: "gone"andmachine_gone; a job on it endslostwith "the job ended with its machine", not
"could not reach the machine over ssh to start it";terminal_execno longer answers ok with ssh's exit
255;environment_closeends it without asking fordiscard=true; andenvironment_openwith the same
title no longer hands back the dead machine asattached,ready: it ends it and starts a new one on that
title (from the newest save when there is one), withreplacedsaying what it replaced. The checks never
run on the machine, so they never count as the agent's use of it.
openajar 0.5.0
openajar 0.5.0, the first release under the new name (agentos until 0.4.0). It keeps everything 0.4.0 brought (prepared machines, checkpoints, replay, video proofs, the network policy, logins, Morph) and adds a desktop on every machine, screen tools, the Dawn web app and the rename. Before 1.0 a minor version may change tools or settings: this one renames the package, the command, the MCP server and the skills, and the old names keep working (docs/LEGACY-NAMES.md). The tool names 0.4.0 renamed are removed, as 0.4.0 announced: a call by one says the new name. To upgrade from 0.3.0 or 0.4.0, run the installer again.
The agent's computer has a desktop you can watch, a pointer that shows what the agent does, and a button
that turns it into a clip you can share. The web app has a new design, agents can work desktop apps that are
not web pages, and the product has a new name: openajar. Everything 0.4.0 brought stays, under the same
tools and settings.
Changed
- agentos is now openajar (a door left ajar: the agent works in its own room, and the door is open for
you to look in and step in). The package, the command, the MCP servers (openajar-local), the skills
(openajar-environment,openajar-lanes,openajar-offload,openajar-studio), the variables
(OPENAJAR_HOME, ...) and the docs carry the new name. What you set up keeps working: theagentos
command runs openajar and says so on stderr, everyAGENTOS_*variable is read when itsOPENAJAR_*one
is not set, and an existing~/.agentosand~/.config/agentosare used where they are, with their
environments, saves, keys and grants.openajar local startreplaces the olddev.agentos.localservice
on the same home and port, and registering replaces theagentos-localentries in Claude Code and Codex
and the old skills. Environments, saves, the machine image (agentos-env, used without a rebuild),
backups, links, cookies and tokens made before still work: docs/LEGACY-NAMES.md lists every name that
stays, and why. To upgrade, run the installer again; it replaces theagentostool withopenajarand
leaves your files where they are. From 0.4.0 that is all: its grant, web app (updated to 0.5.0's),
checkpoints, replays, prepared machines and network settings carry over, and a repository's
.agentos/environment.tomlis still read (.openajar/environment.tomlfirst). From 0.3.0 the agent
grant is replaced by one withsecretonenv/*lanes, sologinworks with noagentos connect.
- The web app's new design, Dawn. The night sidebar with the sunrise at its foot and the organisation
switch on glass, one violet for the agent, what is selected and the primary button, every button a pill,
larger type, and the Dawn picture behind Home's window, sign-in and first run. Home opens on a window with
one sentence and what needs you on glass cards; Environments shows each open one as a picture of its
screen; one environment opens on a night band with its live screen and one pill for who has control.
Every screen follows (needs you, runs, checks, machines, history, routines, settings and their forms,
dialogs and the command palette), and web/DESIGN.md is its rule book. 0.4.0's environment screens are
drawn in it: a closed environment's replay in its screen's place on the night band, Live and Replay so far
on an open one, how it opened, checkpoints, the network and its refused hosts, Resume on a save, who is
watching, and the prepared machines. Dawn's values come through 0.4.0's generated theme and pass its
design rules. - The live view page. The screen is a card, as large as the window allows, with nothing above it. One
pill under it says who has control (the agent's violet dot and its name, or you, with a white ring round
the card and since when) and has one button, Take over or Hand back. The screen's
details, Stop agent, the activity list and the shortcuts are under More. On a phone the pill sits at the
thumb and a keyboard button types on the screen. - A new mark. The openajar mark is "Window a" everywhere: the README, the web app's logo, favicon and app
icons, and the Mac companion's app and menu bar. - Commands start in the repository.
terminal_execon an environment's machine starts in the repository's
folder, as the file tools do, and a relativecwdstarts there too. A command that runs out of time says
so with the next step (timeout_sup to 200, orterminal_job_start), without the straysetsidline. - Every argument of the everyday tools says what it is (
terminal_exec'scmd,file_edit'sold
andnew,terminal_job_read's offsets and waits, and the rest), within 0.4.0's budget for the
default list (7,141 of 7,150 tokens).
Added
- A desktop on every machine. From the moment an environment is ready until it closes, its screen is
a small desktop in the Dawn look: a top bar with the environment's title and what the agent is doing, a
dock with Browser, Terminal and Files, a terminal window where the agent's commands show as it runs them,
and the browser as a window that comes and goes on it. Apps the agent starts open there too.
desktop = falseinenvironments.tomlturns it off. A new machine's desktop is up in about half a
second. - Screen tools for desktop apps that are not web pages:
screen_snapshot(a picture of the screen and
its windows),screen_act(click, drag, scroll, type and keys, paced like a person's hand) and
screen_open(start an app), in thescreentoolset. A take over stops the agent's input on the machine
within about a second and lets go of every key and button, andscreen_snapshotis refused while a person
has control, as the browser tools are. - The agent's pointer. A large violet pointer with the agent's name ("Claude Code") shows where the
agent clicks, types and scrolls, in its browser and with the screen tools, in the live view and on the web
app's environment page. It glides from point to point, fades after a few quiet seconds, and hides while
you have control. It adds no time to the agent's work. - Record a clip. A record button next to full screen in the live view (or R) saves what you see
as a video on your computer: the screen as a rounded card on a dark stage, the agent's pointer in it, a
small mark under it, 1920 by 1080 at 30 frames a second, MP4 where the browser can record it. It is made
in your browser and never uploaded, shows the time it has run, and stops by itself after three minutes.
View-only links have it too. - Still pictures of each screen. The environments list and Home show a small, real picture of each
desktop, taken only while someone looks at them, at most once a minute, never while a person has control
and never for a machine that sleeps. Taking one does not count as the agent using the machine. - Screen recovery. A screen that dies comes back at the node's next check (within about a minute) and
the live view follows by itself: when only the desktop's parts died (its helper, the window manager)
every window stays; when the display server died, the screen starts again empty. A desktop that
finishes installing late starts under the running browser. - The eval suite grows (
make eval, docs/EVALS.md):make eval-smoke(run.py smoke) holds the error
catalog, the skill and the everyday tool list to budgets with no node or model, in CI too; a model's
repeats are compared with Fisher's test (real, likely or the same); and four tasks: probes for what
testers found, a build longer than the default wait, a checkout form with every kind of field read back
by the shop, and a command typed into the desktop's Terminal with the screen tools.
Removed
- The old tool and argument names 0.4.0 announced are gone.
lane_exec,lanes_list,files_putand
every other name 0.4.0 renamed, the three merged tools (browser_tab,browser_dialog_answer,
approval_status), and the old argument names (wait_msonterminal_job_read,timeout_swhere it
becamewait_s) no longer run. A call by one fails as an unknown tool or argument, and the error names
what to call instead ("lane_exec is terminal_exec", "browser_tab is browser_tabs(select=...)", "wait_ms is
wait_s, in seconds"), so an agent on old instructions recovers in one step. docs/TOOLS.md, "Removed tool
names", has the table. A job no longer says where it runs aslanebesideterminal. The CLI still
calls a node from before 0.4.0 by the old names.
Fixed
- What the commands window shows is the agent's real record. A command's output can no longer clear the
window or draw a fake command or exit line, also when an escape sequence is split across reads. - On a server node a control link alone no longer drives a screen with a desktop. The dock opens a root
terminal there, so a control link needs the owner's panel session in the same browser or an owner link. In
local mode the live view listens on 127.0.0.1 only, as before. - Said plainly: the agent's commands run as root on its own machine, so root can read that screen and
what a person types there. The take over keeps the agent's tools out of your way; it does not keep a
hostile agent out of its own machine. The live view says so under Take over, and docs/THREAT-MODEL.md has
a section on the desktop. Sign in on that screen only to accounts the agent is meant to use. - An environment with a desktop still closes, or saves, when it is idle: the node's own screen checks and
still pictures are not counted as use. - The dock's Browser never opens the browser past a stop or a freeze; after a node restart the desktop's top
bar and dock know the browser at once; a resumed save never signals a process from an old pid file; a...
agentos 0.4.0
A computer for your agent, with a browser, a terminal and a disk that stay. This release makes
saving when idle the default, signs the agent's browser in with logins from your vault, and opens a
prepared computer in seconds.
Highlights
- Opens in seconds: a repository's
.agentos/environment.tomlgives it a prepared machine (setup
already run) and the app running when the open returns; an optional warm pool. - Signed in: bind logins from your vault to a computer; the agent's
logintypes them in, again after
a logout, and the model never sees them. - Saves when idle is the default. On Morph (new provider) a save pauses the machine with its memory.
- Network policy (
full,limited,none) held outside the machine, and app secrets that stay on the
node (the machine holds a placeholder). - Checkpoints and rollback, replay of the browser after the machine is gone, video proofs, and
proofs in the pull request. - A smaller tool list: 30 everyday tools (about 7,000 tokens, from 41 and 10,150), no tool named
"lane". Old names keep working for this release and say their new name. - The web app is part of the install, as an option.
- An evals suite (
make eval) that measures agent experience, speed and tokens. - Security fixes from an independent review.
Upgrading: run the install line again. Agents connected before need agentos connect once. Logins
bound by title before this release must be bound again. Before 1.0 a minor version may change tools and
settings; renamed tools keep their old names until the next minor release.
Added
-
web/scripts/codemod-legacy-tokens.mjsmoves the web app's pre-Mono class names (bg-bg,text-13) to
the Mono ones, and has run on all ofweb/src: no old name is left, and the design check refuses one.
Two type roles join the system for the sizes no role had,text-heading(20/28) andtext-headline
(32/40).web/scripts/visual-compare.mjsphotographs the catalog and the main screens on the mock and
compares two runs pixel by pixel. The catalog's screenshot baselines are no longer in git: each machine
records its own on the first run. -
The web app's design system for agents:
web/DESIGN-SYSTEM.md(a screen in five steps, the tokens, the
component groups, the checks), an index per group inweb/design-system/, andweb/AGENTS.mdand
web/CLAUDE.md. The catalog at/$org/design(dev only) draws every component by group, and agentos's
own composites from the mock; the e2e suite photographs it in both themes at 1440 and 390. -
Design rules in
pnpm test(pnpm design:check): no raw colour, arbitrary design value, class that is
not a token, unknownvar(), design value instyle, hand-drawn component or replaced word in a screen,
and a catalog entry for every component. What was there is counted in an allowlist that only shrinks. -
The Tailwind theme is generated from the token script (
web/tokens/theme.py), and the app's own colours
(board wires, Site API fields, the night band, the inverse button's edges) are tokens with light and dark values. -
An environment's agent signs in to accounts, with logins from your vault.
agentos environments login add <site> [--user NAME] [--totp] --title "<title>" | --repo <repo> | --allbinds a vault login
to computers (the password, and a TOTP seed, from a hidden prompt or--stdin, never the command
line;--entrybinds a login already in the vault;list,remove). A binding by title follows the
title across saves and resumes. The agent'slogin(on=...)opens the site's login page when the
browser is elsewhere and fills it with the vault's own checks, again whenever a site logs it out, and
never asks anyone;environment_openandenvironment()list the computer'sloginsby name and
site.agentos environments login "<title>" <site>signs it in now from the node and says what needs
you (a CAPTCHA, a code by email or text) with the live view's link. Withbrowser = "vm"the value is
typed into the VM's browser, whose session cookies stay in it and in its saves (encrypted): see
docs/THREAT-MODEL.md. No replay frame is taken while a fill runs. The local node now serveslogin,
vault_fillandvault_list, and theenvironmentsspecial includessecretonenv/*lanes: an
agent connected before needsagentos connectagain. -
The one-line install includes the web app, as an option. Each release carries the web app
prebuilt (agentos-web-<version>.tar.gzand its.sha256, from the newweb-releaseworkflow and
scripts/web_asset.py, ormake web-assetlocally).agentos setupgains a web step that offers it
(yes by default,--no-webskips it): it downloads the copy for the installed version, refuses one
whose sha256 differs or whose archive holds a link, an absolute path or a.., and unpacks it into
~/.agentos/web/appwithout sudo.agentos web install|update|remove|statuschange it later.
Running the installer again updates the web app to the same version (without setup, the installer
runsagentos web update;AGENTOS_NO_WEB=1skips it), and a copy from another version shows in
agentos setup checkandagentos doctor. The service serves a checkout's ownweb/distfirst and
the installed copy second, looked up on each request, so an install shows on the next reload. -
What the evals found, fixed. A lane's name passed as
on(<id>-browser, the open's old
lanes.browser) reaches the machine for the chat whose session holds it, and anyone else hears it is a
lane's name and to pass theon; the open's result names no lane any more.lane_execadds a one-line
hintnaminglane_file_read(path=...)when a command only read a file (nothing on any other command).
A command refused a host on alimitednetwork says so in its own result (network.blocked_nowwith the
owner's command), and the CONNECT 403's status line carries the same words.NODE_USE_ENV_PROXY=1joins
the proxy variables, so Node'sfetchgets the 403 instead of ENOTFOUND. The environment's browser starts
during the open, beside the clone (a close cancels a start nobody used), so the firstbrowser_navigate
no longer waits 10 to 15 s for it. Abrowser = "vm"environment on alimitednetwork is no longer
quarantined by the exit IP self-check (an environment's tunnel is not read), so a video proof works there.
The evals gaint07_video_proof, thefirst_navigate_smetric and a wrong turn for a lane's name ason. -
An environment's page in the web app tells what happened, on the real node. A closed one leads
with its replay (frame by frame, the agent's calls, a person's take-overs and the proofs on one track
that fits a phone), then its proofs and pull request; an open one with its live browser. How it opened
is one line that unfolds to each step and the work it skipped. Checkpoints are drawn as a line of work
with the rollback on a side line, and each opens what changed since it, file by file. The owner resumes
a save from the app (a refusal by the node's limits stays in the dialog with the fix) and watches it
wake. Prepared machines has a page of its own with the warm pool's worst month. A video proof shows its
preview and opens a player. -
The app sees how an environment came up, and what the node's own machines cost. The detail has
bringup: each step of the bring-up with when it began and its seconds (the ones running while it
starts), and the work it skipped (a prepared machine's setup and the seconds saved, a warm machine's rent
and boot).GET /api/orgs/{org}/environments/preparedgives the owner the prepared machines, the warm pool
with its hourly price and the most it can cost a month, and this month's opens by where they started. The
saves name the machine they wake on, its price and their services; the detail adds the replay summary and
the resume (resumed,resumed_by,owner_woken,resumed_at). -
Checkpoints and rollback inside an environment.
environment_checkpoint(note)keeps every
repository under/workas it is (tracked and new files git does not ignore) as a git commit under
refs/agentos/checkpoints/<n>, made with a temporary index: seconds, no provider call, and the
agent's index, stash and branches untouched.environment_rollback(checkpoint)checkpoints the work
first (so it can be undone), then puts the branch back at the checkpoint's HEAD and the files as they
were, removing new ones; ignored files (node_modules,.env) stay. Refused mid merge or rebase.
About 20 are kept;environment()lists them, and the web API gives what changed since each one and
one checkpoint's patch against now. They never leave the machine: the git gateway refuses a push of
refs/agentos/*, and a files-only resume does not keep them. Intools_more("terminals"). -
Watching keeps an environment open. A live view of its browser, or a connection through an
environment_portlink, counts as use while it is open, so the reaper never saves or closes it under
you;environment()and the web API saywatched. The web app polling the page no longer counts. -
You resume a saved environment without a chat.
agentos environments resume <title or save id>,
Resume in the web app's Saved list, or the Resume button on the saved environment's live-view link
(which now shows the title, when it was saved and the note instead of an error). It resumes as
environment_opendoes, within the node's limits, held by no chat: you watch it at once, and the next
agent that opens its title takes it (opened: "resumed"). An agent's call with the handle of a saved
one says when it was saved and how to come back. -
Video proofs.
browser_record(action="start"), the flow, thenbrowser_record(action="stop", proof="guest checkout, end to end")records the real screen of...
agentos 0.3.0
Agents now work on their cloud computer the way they work locally, one chat per machine, with a lot less waiting. This release comes from three rounds of agent testers and two field reports from a real project.
Install or update
curl -fsSL https://raw.githubusercontent.com/sabalys-ramunas/agentos/main/scripts/install.sh | shAlready on 0.2.0? Run the same line: it updates agentos in place. Then run agentos setup ssh once to let your agents use ssh agentos-<id>.
Highlights
- The agent's own shell.
ssh agentos-<id>,scpandrsyncfrom your laptop. A one-word fix that took 41 tool calls in testing now takes about a dozen. - About three times faster to open. 81 s instead of 255 s on a real run. The local service is no longer throttled by macOS, the clone through your laptop moves a quarter of the bytes, and public repositories clone straight from GitHub.
- Dev servers survive agentos. Jobs run detached on the VM. A restart or an update of agentos reattaches to them, with the same job ids.
- One chat per machine.
environment_openreturnson, and only calls that carry it reach the machine, so two chats never share a computer by accident. - Commits are yours. With your GitHub login saved, commits carry your GitHub name and noreply address.
- No more polling.
lane_job_read(until_port=3000)waits until a dev server listens.environment(on=...)waits for a machine that is still starting.environment_openwaits for a free slot.
- Memory you can see. Every VM gets swap. A process the kernel kills for memory is named in the result, with the next size to try.
- No system updates mid-session. They are applied when a VM is set up and when the dev snapshot is built.
Fixes worth knowing
- A git command run while the repository was still cloning could break the clone. Calls now wait until the machine is ready.
- A push could reuse a mirror fetch from before it, which let a force push over a colleague's commit through. Pushes now always see the remote as it is.
- A save could be undone by the background reaper and leave its title held.
Breaking
- Clients that send no MCP session id (Claude Code, for example) must pass
onon every call to a machine. The skill and the server's instructions teach this.
Full list: CHANGELOG.md.
agentos 0.2.0
The first tagged release of agentos: your coding agent's own computer in the cloud.
agentos runs on your laptop next to Claude Code, Codex or T3 Code. When the agent needs a machine, it gets a fresh VM with a terminal, a real browser and your repo, works there, shows you screenshots as proof and opens a pull request. Then the VM is deleted, or saved for tomorrow for cents a month.
Install
curl -fsSL https://raw.githubusercontent.com/sabalys-ramunas/agentos/main/scripts/install.sh | shThe installer now installs the newest release (this one), then starts a guided setup in the terminal. Nothing bills until an agent opens a VM.
Highlights
- One command to install, then a guided setup: VMs (a Hetzner token, or Docker), your GitHub login, the limits, and Claude Code and Codex connected to a small service on your laptop. No server needed.
- Three ways an environment stays: single use (deleted when done), saves when idle (the whole machine is saved and the VM deleted, resumed by its title from any chat), and kept running (only when you allow it).
- Private repositories through your GitHub login, or a card where you paste a token. The agent never sees one.
- Moving files:
environment_uploadsends a local checkout up with a one-use link, andenvironment_downloadgives you a file from the VM.environment_portgives you a link to a dev server, andenvironment_pull_requestopens the pull request. - The browser runs on the VM, ready at boot with the
devsnapshot that setup offers to build. - A new README and logo.
Also
- The engagement engine, the day plan and the classic panel are removed: agentos is about agent environments now.
- Alpha: tools and settings may still change between minor versions before 1.0.
Full list: CHANGELOG.md.