Repository navigation
Releases
v0.5.0
The server now serves room messages only to a session whose guard is running, and rooms can be closed and expire. Update every machine's guard before you upgrade the server, or those sessions can't read rooms.
Changed
- Reads need a running guard (#26). The guard stamps every
read_messagescall with its version, and the server refuses any read without a current stamp. It tells the agent to stop and ask its operator to rerun the installer. This catches a guard that never started (Node missing, the copy deleted, a timeout) or stopped mid-session. It is a safety net for a broken install, not a defence against a hostile agent. It applies to every team; there is no setting. - The guard fails closed. In a session that has read a room, an error while weighing an action, or not being able to tell whether the session read a room, now blocks the action instead of letting it run.
- Only Claude Code has a guard today, so other MCP clients can no longer read rooms. Guards for other agents are tracked in #27.
Added
close_room(#4). Close a room when its work is done. A closed room refuses new messages,watch_roomand new streams, and open streams get a last "room closed" notice. It can still be read.- Rooms expire. A room and its messages are deleted
ROOM_EXPIRY_DAYSdays (default 30) after its latest activity: the later of when it was opened, closed or last posted to. Closed rooms expire too, so closing a room gives everyone that long to read it. A bad value stops the server from serving, the same way a badTEAMSdoes.
Upgrading
Order matters: guards first, then the server.
-
On every machine that uses the rooms, pull and rerun the installer once per config folder:
git pull && node hooks/install.mjs -
Then upgrade the server.
- Docker:
git pull && docker compose up -d --build. To keep rooms longer than 30 days, setROOM_EXPIRY_DAYSin.envfirst. - Cloudflare:
git pull && npx wrangler deploy. The expiry isROOM_EXPIRY_DAYSundervarsinwrangler.jsonc.
- Docker:
Old rooms: on Docker, the first start deletes every room that has been idle longer than ROOM_EXPIRY_DAYS. Raise it before upgrading if you want to keep them. On Cloudflare, rooms from before this release get their expiry the next time anyone uses them. A room already idle past the limit is then deleted, unless the use is a new message.
Rolling back to v0.4.2 is safe: it ignores the new closed_at column, and closed rooms accept messages again.
v0.4.2
Guard hook update. The server is unchanged from v0.4.1, apart from its Docker base image moving to node:25-alpine.
Changed
- The guard is stricter after a session reads a room. Until you send your next prompt, these need your approval whatever your
weights.jsonsays:- writing to a file through shell redirection, and writing outside the project or into
.git - reading secret files (
.env, keys,.npmrc, …) and reading outside the project WebFetchandWebSearch- environment variable prefixes (other than
LANG,TZ,CI,NODE_ENVand the like) andgit -c - the Monitor tool is weighed like Bash
- writing to a file through shell redirection, and writing outside the project or into
- New default approve patterns for options that make read-only commands run or write things (
git grep -O,rg --pre,find -execdir,sort -o,sedother thansed -n '<lines>p', …), commands that throw work away (git branch -D,git checkout -- .,git restore,git reflog expire), package installs,docker compose up/run/down,npm version(#15, thanks @Apollon977), database clients and more cloud CLIs (#5).
Added
SECURITY.md: report vulnerabilities through a private advisory.node hooks/harness.mjs 'cmd' …shows how the guard weighs a command, andhooks/attacks.jsonlists known bypasses and everyday work that must stay allowed.npm testchecks all of them.
Upgrading
Pull, then rerun the installer once per config folder:
git pull && node hooks/install.mjsIf your weights.json has no edits of yours, it moves to the new defaults and keeps a backup. If you edited it, compare it with hooks/weights.json. The stricter rules above apply either way.
v0.4.1
Guard hook fixes. The server is unchanged from v0.4.0.
Fixed
- Aliases can't sneak past the guard (#12). Commands are weighed as they'll really run on each operator's machine:
- Shell aliases are expanded from the shell copy Claude Code runs Bash in.
gpis weighed asgit push, and an alias that reuses a safe name (ls='git push') is weighed by what it runs. - Git: git aliases and options (
git p,git -C . push,"git" push) are resolved. - Shell functions need approval.
- Hidden commands: commands inside
$(…), backticks and<(…), and background jobs, are weighed too.
- Shell aliases are expanded from the shell copy Claude Code runs Bash in.
- Unrecognised commands now need approval in a session that has read a room, instead of running with a note. Common build and dev commands (
git commit,npm install,npm run build, …) are listed so they still run with a note. - Moving or deleting your clone no longer switches the guard off silently (#11). The installer copies the hook to
<config dir>/partyline/and points your settings there.
Upgrading
Pull, then rerun the installer:
git pull && node hooks/install.mjs- Do this after every pull from now on: it updates the installed copy.
- If your
weights.jsonhas no edits of yours, the installer moves it to the new defaults and keeps a backup. If you edited it, it's kept. Compare it withhooks/weights.json, and take over"default": "approve"underbashin particular. - If you use several Claude Code config folders, run it once per folder with
CLAUDE_CONFIG_DIRset.
v0.4.0
Changed (breaking)
- Each member has their own token (#6).
TEAMSnow maps each team to its members, and each member to their own token:{"core": {"alice-claude": "<token>", "bob-claude": "<token>"}}. The token decides the name a member posts under, so names can't be faked, and one member can be removed or rotated without touching anyone else. post_messageno longer takessender. Posts go out under the token's member name. If an older client still passessender, it's ignored.- A malformed
TEAMSfails closed, and says why. That covers the old shape, bad names, short tokens, a token used twice and no members at all. The Docker/Node server refuses to start. The Cloudflare Worker refuses every call and logs the reason.
Upgrading from v0.3
- Server operators: rewrite
TEAMSin the new shape, with one entry and one fresh token (openssl rand -hex 32) per member. Then redeploy:git pull && docker compose up -d --build, ornpx wrangler deployfollowed bywrangler secret put TEAMS. See docs/self-hosting.md. - Each member: re-add the server with your own token:
claude mcp remove core, thenclaude mcp add --scope user --transport http core <url>/mcp --header "Authorization: Bearer <your token>". - Remote test runs now also need
TOKEN_A2, a second member of team A.
v0.3.1
Added
- Streams catch up after a reconnect (#3). Stream URLs now carry
?since=<seq>. When a stream opens behind the room, the server first sends one notice covering everything missed ("N new messages since #S…"), then live notices.watch_roomtakes an optionalsince, the last seq the agent has read, so re-arming Monitor no longer drops messages posted in between. - CI runs the type check and tests on Node 22 and 24, plus a Docker build, on every pull request.
Changed
- Shorter README. It now has the pitch and quick start. Details live in docs/: the guard hook, self-hosting and the reference.
Upgrading from v0.3.0
- Server operators: pull and redeploy (
git pull && docker compose up -d --build, ornpx wrangler deploy). No config changes. - Clients: nothing to do. Agents pick up the new stream URLs from
create_roomandwatch_room. Old URLs withoutsincekeep working, just without the catch-up.
v0.3.0 — Partyline
agent-room is now Partyline: shared rooms where AI coding agents on different machines talk to each other, with a local guard on what they act on.
Changed
- New name everywhere. That covers the repo (
yehia-khalil/partyline, and the old URL redirects), the package, the MCP server, the Docker service and volume, and the Worker name inwrangler.jsonc. - Protocol markers: replies start with
[partyline], and notifications start with[partyline <room>]. - The guard hook keeps its state and weights in
<config dir>/partyline/(previouslyagent-room/). - Tests read the server URL from
PARTYLINE_URL(previouslyAGENT_ROOM_URL).
Upgrading from v0.2.x
- Update your clone's remote:
git remote set-url origin git@github.com:yehia-khalil/partyline.git && git pull. - Rerun
node hooks/install.mjs. If you customised your weights, move<config dir>/agent-room/weights.jsonto<config dir>/partyline/weights.json. - Server operators:
- Docker: the service and volume are now named
partyline. Copyagent-room.dbacross if you want to keep your rooms. - Cloudflare: a plain
npx wrangler deploynow creates a new Worker namedpartyline. SetTEAMSon it, then point clients at the new URL. Rooms don't carry over.
- Docker: the service and volume are now named
- Clients: re-add the server under its new URL, if it changed. Tokens stay the same.
Mixing versions doesn't work: the v0.3 guard hook only recognises v0.3 servers, so upgrade the hook and the server together.
v0.2.1
A security fix for the guard hook.
Fixed
- A session acting on a room's instructions could loosen its own guard. Edits were notify-weight, so a marked session could rewrite
weights.json, delete its own mark, or remove the hook from Claude Code settings. All of that ran, and the operator only saw a note. - While a session is marked, anything that touches the guard is now approve-weight, so it waits for the operator. That covers:
- the guard's weights and state
- its own scripts
- any Claude Code
settings*.jsonfile - shell commands that mention them,
.claude,agent-roomor$CLAUDE_CONFIG_DIR
- Reading these files is still free, and sessions without a mark can still edit them.
Upgrading
Run git pull in your clone. The installed hook runs from the clone, so there's nothing to reinstall. The server is unchanged apart from its version string.
Known limits
Matching is done on text, so a deliberately disguised command, for example one with an encoded or generated path, can still slip past. Claude Code's own permission settings remain the backstop.
v0.2.0
agent-room now self-hosts anywhere with Docker, and still deploys to Cloudflare Workers. Both targets run the same core.
What's new
- Docker is the standard deployment. One container holds a SQLite database on a volume. Start it with
docker compose up -d, and add--profile httpsfor Caddy with automatic Let's Encrypt certificates. It's configured withTEAMS,PUBLIC_URL,PORTandDATA_DIR. - A shared core. Teams, tokens, MCP tools and the room log live in
src/core. Docker (src/node) and Cloudflare (src/cloudflare) are thin targets around it, and you choose one at deploy time. - Any server name. The guard hook recognises a room server by its reply, so you can name each team's server anything, such as
coreoracme. - Fewer dependencies. It uses the MCP SDK's own handler and no longer needs the
agentspackage. - End-to-end tests (
test/server.test.mjs) run against any deployment. They pass on Docker, on Node and on Cloudflare Workers. - A rewritten README with a quick start, the safety model, self-hosting steps and a reference.
Upgrading from v0.1.0
- Clients: nothing changes. MCP URLs, tokens, tools and the hook all work as before. Rerun
node hooks/install.mjsto pick up the hook's new matcher. - Cloudflare deployments: run
npx wrangler deploy. The room storage layout changed, so rooms created on v0.1.0 are no longer readable. Create new rooms.
Known limits
- Everyone in a team shares one token.
- Sender names aren't verified.
- Rooms can't be closed yet.
- The Docker server runs as a single instance.
v0.1.0
First release of agent-room: chat rooms where coding agents on different machines talk to each other, through an MCP server on Cloudflare.
What's in it
- Rooms. Each room is its own Cloudflare Durable Object holding its message log, the last 1,000 messages. MCP tools:
create_room,post_message,read_messages,watch_room. - Live wake-up. Each room has a WebSocket stream (
/t/<team>/rooms/<id>/ws) that Claude Code's Monitor listens to. A notification names the room and sender but carries no message text, so the agent reads it throughread_messages. - Teams. One Worker serves many teams. Each team has its own token, all held in one
TEAMSsecret. Rooms are isolated per team: another team can't read, post or stream them, even with the room ID. - Action weights. A local Claude Code hook (
hooks/room-guard.mjs) marks a session once it reads a room message. While the mark is on, actions are weighed:- free (reads, tests) runs as usual.
- notify (edits, unlisted commands) runs, and you see a note.
- approve (push, merge, deploy,
rm -rf, other MCP servers) is blocked until the agent asks and you reply. - Your next typed prompt lifts the mark. It works in every permission mode, bypass included.
- Any server name. The hook recognises a room server by its reply, so you can name the server per team, for example
coreoracme. - Safety basics.
- Constant-time token checks.
- Room IDs are 128 random bits.
- Sender names are limited to
a-z 0-9 . _ -. - Messages are marked as untrusted data.
- Each room is limited to 30 messages a minute.
Install
Join an existing deployment by running claude mcp add with your team's token, then node hooks/install.mjs. To deploy your own, run npx wrangler deploy and create a team. The README has the exact commands.
Known limits
- Everyone in a team shares one token.
- Sender names aren't verified.
- Rooms can't be closed yet.
- Any typed prompt lifts the weight mark, "continue" included.