
Exec Room: Agents working together in rooms to help run your business. Create a local team of agents that run on Claude Code and research, review, suggest, create drafts for you. Anything that leaves a room waits for your approval. Open Source.
Repo ANDRS-Projects/exec-room, read from GitHub 26 min ago. ANDRS-Projects has not joined Take My Repo.
ANDRS-Projects/exec-room is a public repo on GitHub by ANDRS-Projects. Exec Room: Agents working together in rooms to help run your business. Create a local team of agents that run on Claude Code and research, review, suggest, create drafts for you. Anything that leaves a room waits for your approval. Open Source. It declares the AGPL-3.0 license on GitHub. It is written mainly in JavaScript, has 1 stars and last changed on 2026-11-21. Nobody has reviewed it on Take My Repo yet, which says nothing about its quality. Read from GitHub on 2026-11-21.
In the maintainer's words. Shown from GitHub; Take My Repo does not keep a copy.
Exec Room is an open-source agent room for running a business: a local web room where you, the exec, work with a team of agents that run on Claude Code. They research, draft and propose actions (GitHub issues, Notion pages, email drafts, files to save, jobs for other agents). They have no tool that runs commands on your computer and write only to drafts/; anything that leaves the room (an issue, a Notion page, an email draft, a saved file, a job for another agent) waits for your approval. These are Claude Code permission rules, not a sandbox: see Security.

The room starts with one agent: Coen, the room lead. Every message goes to him first. You build the rest of the team yourself, in the room:
Each agent gets a name, a job description, instructions, the tools it may use and an avatar from the preset set.
npm run room:key to pay with an Anthropic API key instead, with a spend cap per run and per daydrafts/, secret files blockedThe pictures show a demo install with a made-up bike repair shop; every name and number in them is invented.
Approve before act. Coen asked two advisors, then proposed a job for two makers and a GitHub issue. Nothing happens until you approve; you can also send a proposal back with a note or dismiss it.

A brief per room. What every agent in the room should know about the business, written by you.

What a room can use. Folders, sites, web search and connectors are set per room; anything not listed is refused.

Decisions. Every proposal you approved, sent back or dismissed in the room, across its chats.

Narrow windows. The same room in a phone-width window; the chat list opens as a drawer.
The room runs on your own Claude plan. Every message, every advisor Coen consults and every maker run is a Claude Code run that spends your plan's usage. Two jobs can run at the same time (in different rooms), which uses it up faster. The room keeps the cost Claude Code reports per message and per chat, but it sets no budget and no limit of its own.
Started on an API key instead (see Running on an API key), every run is billed to that key, pay-as-you-go. There the room does set limits: at most $2 per run and $10 per day unless you change them, and the cost it shows is real spend.
claude), signed in to your Claude plan (tested with version 2.1.316) — see the Claude Code docs. By default the room runs agents on your plan and never uses an Anthropic API key; to pay with a key instead, see Running on an API key (no plan sign-in is needed for that). It starts every run with --agents, --tools, --setting-sources and --disable-slash-commands, so it needs a recent version that has all four.gh), signed in, for GitHub actionsop), signed in, for tokens: the settings page takes an MCP server's token only as a 1Password reference, and the API key and Notion token can be references too. Without it, use Connect, or "tokenEnv" in room.config.json (see Configuration).git clone https://github.com/ANDRS-Projects/exec-room.git
cd exec-room
./setup.sh # checks the requirements, installs dependencies, creates room.config.json, offers to start the room
Then open http://127.0.0.1:4179. Stop the room with Ctrl-C; start it again later with npm run room.
Click Add agent to build your first advisor or maker, and Room settings to choose what the room may read, which sites it may open, and what it may propose. Its Decisions tab lists every proposal you approved, sent back or dismissed in that room, across all its chats: who decided, which agent proposed it, in which chat, when, and what came of it. The Chats panel lists the room's chats; picking one opens it with its history. Coen picks up where he left off as long as Claude Code still has its own log of that chat (about 30 days after the last message, unless you changed Claude Code's cleanupPeriodDays); an older chat is marked "history only" and Coen starts fresh there.
By default every run uses your Claude plan. To pay with an Anthropic API key instead (pay-as-you-go), give the room the key and start it in key mode:
export ROOM_ANTHROPIC_API_KEY=sk-ant-... # or: export ROOM_ANTHROPIC_API_KEY_REF="op://vault/item/field" (read with 1Password's `op read`)
npm run room:key
The choice is made when the room starts and holds until you stop it: the plan and a key never run side by side. npm run room starts it on your plan again. A key that merely sits in your environment (ANTHROPIC_API_KEY) is never used.
| Variable | What it's for |
|---|---|
ROOM_ANTHROPIC_API_KEY |
The key. Read once at start-up, kept in memory, handed to the agent runs only. It is never written to data/, the job log or the page, and the page has no field for it. |
ROOM_ANTHROPIC_API_KEY_REF |
Instead of the key itself: a 1Password reference the room reads once with op read. |
ROOM_RUN_CAP_USD |
The most one run may spend (default 2). Claude Code checks it after each turn, so a run can pass it by one turn. |
ROOM_DAY_CAP_USD |
The most all runs may spend in a day (default 10). Once it is reached nothing starts until the next day, including the turns the room starts itself. |
What is different in key mode:
data/claude-key), without your plan sign-in. With a plan sign-in present Claude Code uses the plan and not the key, so this is what makes the key the one that pays.Without a key the room does not start in key mode. See SECURITY.md for how the key is handled.
Most settings live in the room itself (Room settings), per room: folders agents may read, web search, sites they may open, Notion and Microsoft 365 access, GitHub repos, Notion pages and email recipients. A new room starts with nothing beyond drafts/.
room.config.json (created from room.config.example.json, gitignored) only holds what the whole install needs:
| Key | What it's for |
|---|---|
notionTokenRef |
A 1Password reference (op://vault/item/field) read once at start-up with op read. Or set ROOM_NOTION_TOKEN instead. Without a token, Notion actions are off. |
notionDatabases |
Only for installs from before Notion databases were set per room: the databases listed here move, at start-up, into the rooms that allowed them, with Status = draft. Add databases in Room settings → Can propose → Notion databases instead. |
notionChannels |
The channel names a post may have (default: X, LinkedIn, Substack Note, Substack Post, Instagram, Blog). |
hiddenConnectors |
Other claude.ai connectors on your account to keep out of the agents' view, by the name in their tool names (mcp__claude_ai_<name>__…), e.g. ["My_Connector"]. Agents can only call the tools a room allows; connectors that aren't hidden are still visible to them, so this keeps the rest out of their context. Claude Docs, Gmail, Slack, Google Drive and Google Calendar are always hidden. Notion and Microsoft_365 can't be listed, as those are switched per room: such an entry is skipped with a warning at start-up, like an entry that isn't a connector name. |
agentsGetClaudeFiles |
Whether agents are given Claude Code's own files, as any Claude Code session started in the room's folder would be: the CLAUDE.md files of that folder and the folders above it (this repo's notes for contributors, your personal ~/.claude/CLAUDE.md, anything other tools write into such files) and the memory Claude Code keeps for the folder. Default false: agents get only what the room gives them (their own description, the room's note and brief, your message), and the room checks at every start that this still holds, with one small run that costs about a cent; ROOM_CONTEXT_CHECK=off skips that check. Set it to true if you keep instructions in CLAUDE.md that the agents should follow too; every agent in every room then gets all of those files, whatever else they say. |
mcpServers |
MCP servers set up by hand, next to the ones set up in Exec Room settings: an entry from the catalog ({"catalog": "<id>", "tools": [...], "tokenRef": "op://…"}) or a server by its web address ({"name": "…", "url": "https://…", "readOnlyTools": [...], "tokenRef": "op://…"}, or "tokenEnv" with the name of a variable that holds the token). You name each tool agents may use. A program on your computer can't be added yet. See MCP servers. |
slack |
Approving proposals from Slack: botTokenRef and appTokenRef (1Password references to the Slack app's two tokens; or set ROOM_SLACK_BOT_TOKEN and ROOM_SLACK_APP_TOKEN) and approver (your Slack member ID). Leave it out to keep Slack off. See Approving from Slack. |
ownerName |
Your name as decisions are recorded under it: "approved by …" on a proposal card, the Decisions tab in Room settings, and the credit line on what an approved proposal creates (default: "the owner"). One line, up to 60 characters. It is a label, not a login. |
For Notion, the databases and pages a room may save to are set in the room, not here: see Notion.
Agent runs and gh are started with a short, fixed set of environment variables (what a program needs to run and reach the network, and their own login settings), not with everything your shell has. If your setup needs another variable passed on, for example one a version manager uses to find node, start the room with its name in ROOM_PASS_ENV (names separated by commas).
Environment variables are listed in .env.example. The server reads them from your shell; it does not load a .env file.
Exec Room settings (the room menu → Exec Room settings) list the MCP servers the room uses, for every room at once:
mcp-catalog.json), set up with Set Up: Money Magician (reading your books: invoices, customers, vendors, taxes, profitability; with Connect or an API token; made by ANDRS, the company behind Exec Room), GitHub (reading code, issues, pull requests, commits and releases; with a read-only token), Notion (with Connect) and Resend (reading your inboxes; with Connect, see Resend inboxes).Agents get the ticked tools of a server and nothing else of it. For a service in the catalog the room knows which tools write, and refuses those. For a custom server it can't tell: you decide when you tick its tools, and the page shows the server's own read-only claim, which the room doesn't check. What agents send to a server's tools (a search, say) goes to whoever runs that server, so add a custom server only if you trust them with what the agents in those rooms can read.
GitHub. Set it up with a fine-grained token (GitHub: Settings → Developer settings → Fine-grained tokens), read-only, with the repos agents may read; the room uses GitHub's read-only address. The token decides which private repos agents can read, in every room that switches GitHub on; a room's own repo list only limits which repos it may propose issues for. Every fine-grained token can also read all public repos on GitHub, so agents can search text anyone wrote there.
Notion is set up here too (Popular MCPs → Notion → Set Up → Connect); what agents read in it and what they can save there is in Notion.
Agents can read your Notion and propose to save three kinds of things there: rows in a database, posts in a post database, and pages. Nothing is written until you approve the card, and then the Exec Room app writes it, not the agent.
notionTokenRef (a 1Password reference) or ROOM_NOTION_TOKEN. The token is used by the room for saving and checking, never by agents. It also works for saving while Connect is paused or needs a new sign-in.In a room with Notion switched on (Room settings → Can use), agents can search, open and query Notion through the Connect sign-in: three read tools, nothing else. Notion's server offers more (68 tools in November 2026); the settings page lists the rest as blocked for agents. This works the same on a Claude plan and on an API key. While Notion is paused or needs a new sign-in, agents read no Notion.
Not connected? On a Claude plan, agents read through the claude.ai Notion connector on your account, if you have it. On an API key, put the link of a Notion page or database in your message and the room reads it with your integration token for them (see Running on an API key).
All three are set per room in Room settings → Can propose, and every room starts with none.
Rows in a database. Under Notion databases, paste a database's link. The Exec Room reads its columns from Notion. Then:
Posts. A database with Channel and Post Content columns also takes posts: a whole package at once (say LinkedIn, X and a Substack note, one row each), with an image prompt and the insight it came from. Only the post columns are the agent's by default there, so columns a posting tool reads (an image link, a "posted" mark) stay empty. The channel names a post may have are set with notionChannels.
Pages. Under Notion places for pages, add the pages new pages may go under, each with a name and what it's for (meeting notes, research, a blog backlog). Agents propose a new page (an article, a review, notes) under one of these places, or under a page inside one; where it lands is checked when you approve. A page the room made can later be updated: the new version goes on top and the previous one is kept under it, so nothing is lost. Pages written by hand, or made in another room, can't be updated by agents.
Agents never get a tool that writes to Notion. The room only creates pages and rows, and updates pages it made itself; it never moves, deletes or comments, and the only change it makes to a database's columns is the Status column you add with the button. Details are in SECURITY.md.
If you use Resend Inboxes (an address that receives mail, often with Resend's own inbox agent drafting replies), agents can read them and Coen can help you answer.
Set it up. Exec Room settings → Popular MCPs → Resend → Set Up → Connect, and approve "Exec Room" in Resend. Then switch Resend on in each room that should use it (Room settings → Can use). Resend only offers full access when you sign in, so the limits below come from the Exec Room app, not from Resend.
What agents can do:
Who mail may go to. A reply only to people already in its thread. A new email only to the room's Email recipients for drafts (Room settings → Can propose). Drafts with attachments can't be sent from the room.
Every new proposal can also come to you as a Slack message, with Approve and Dismiss buttons, so you can decide from your phone while the room runs on your computer. It is off until you set it up.
approver count. A tap from anyone else does nothing and is noted in the terminal.You need a Slack workspace where you can add an app, and 1Password with its command-line tool (op), as for the Notion token.
Create the Slack app. Go to api.slack.com/apps, choose Create New App, then From a manifest, and pick your workspace. Paste the contents of docs/slack-app-manifest.json and click Create. If Slack says the app needs an admin's approval, your workspace restricts apps: approve it as an admin, or ask one.
Make the app-level token. In the app's Basic Information, under App-Level Tokens, choose Generate Token and Scopes. Give it a name (for example exec-room), add the scope connections:write, and generate it. Copy the token: it starts with xapp-.
Install the app. Under Install App, choose Install to Workspace and allow it. Then, under OAuth & Permissions, copy the Bot User OAuth Token: it starts with xoxb-.
Put both tokens in 1Password and copy their references (in 1Password: the field's menu → Copy Secret Reference; they look like op://Private/Exec Room Slack/bot-token). The room never needs the tokens themselves. If an item's title has a space at the end or other unusual characters, use the item's ID in the reference instead of its title.
Find your Slack member ID. In Slack, open your profile, choose ⋯, then Copy member ID. It starts with U.
Add it to room.config.json:
"slack": {
"botTokenRef": "op://Private/Exec Room Slack/bot-token",
"appTokenRef": "op://Private/Exec Room Slack/app-token",
"approver": "U0123ABCDEF"
}
Restart the room. 1Password may ask you to allow it to read the two tokens. The terminal then says Slack: connected. New cards are sent to you as a Slack DM with Approve and Dismiss. If 1Password didn't hand a token over in time, or a setting is wrong, the room starts without Slack and says why, in the terminal and on the page; fix it and restart.
Using a Slack app you already have (Slack's free plan limits how many apps a workspace can have). This only works for an app you created yourself at api.slack.com/apps, not one from Slack's app directory. Don't paste the manifest over it: that would replace its settings. Instead:
chat:write and im:write under OAuth & Permissions, and under App Home turn on the Messages Tab with users not allowed to send messages. Reinstall the app when Slack asks, and carry on from step 3.ROOM_DATA_DIR, as for a test copy) stays off Slack unless you start it with ROOM_SLACK=on. ROOM_SLACK=off keeps any room off Slack.Import. Already have agents for Claude Code in ~/.claude/agents (as <name>.md or <name>/AGENT.md)? Click Add agent → Import from ~/.claude/agents and pick one. The room copies its name, description, model and instructions into the build form; you choose its role, tools and avatar there. It's a one-time copy: later changes to the file don't reach the room, and a tools: line in the file is ignored, so a file can never give an agent more access than you chose.
Export. Agents you build in the room are stored in data/agents.json. To use one outside the room too, open it and choose Export as AGENT.md: the room saves it in drafts/ and shows the command to copy it into your own ~/.claude/agents folder. The room never writes there itself.
Readable in every room. Runs start in the folder you cloned the room into, so every agent in every room can read that folder, whatever folders the room has: the room's own code and docs, room.config.json (your Notion database ids and the 1Password reference to the token, not the token itself), and drafts/, including drafts written in other rooms. Only data/, secret-looking files (.env, keys and the like) and .git are excluded. Every agent can also read ~/.claude/agents, ~/.claude/skills and ~/.claude/knowledge (and what those folders link to). Keep nothing in any of these places that an agent in another room shouldn't see.
The room is built to be usable with a keyboard and a screen reader. It has not been audited, and it makes no claim of conformance to WCAG.
Keyboard. The controls are buttons, links and fields in the page's own order, so Tab reaches them and each shows a focus ring. Enter sends a message and Shift+Enter starts a new line. The tabs in Room settings move with the arrow keys; Escape closes the room switcher, the chat list on a narrow window, the dialogs and the Dismiss form on a proposal card. When a turn starts, focus stays in the message box.
Screen reader. While a turn runs, a screen reader is told only what matters, in a short sentence each time: that your message went out, that a focus agent took over, how the turn ended ("Coen replied.", or who was consulted and how many replied), that a proposal is waiting for your decision, how an approved action ended, and any notice or error the room shows in the chat (a proposal that couldn't be used, a run that failed, a message that wasn't sent). The text of a reply is not read out as it streams in: read it in the chat when the turn is over. Opening a saved chat announces nothing. The agents in the rail are two groups, advisors and makers, and each agent's button is labelled with its name and what it does.
Reading. The state of a proposal, a consulted agent or a chat is written out in a label ("waiting for you", "done", "failed"), not shown by colour alone. Those labels are 12px or larger, the fields you type in have an edge at 3:1 against white, and the pulsing "working" dots stand still when your system asks for reduced motion. There is a dark theme: it follows your system by default, and the half-moon button in the header lets you choose Auto, Light or Dark and one of three dark styles (navy, black, warm), remembered in your browser. Text and field edges keep their contrast in each.
Known gaps. The spoken updates were checked by reading what the page hands to a screen reader, not yet in a full session with VoiceOver or NVDA. The outlines of cards, buttons and agent chips are lighter than 3:1; buttons and chips carry a text label, so the outline is not the only cue. Some secondary text is 11px. A room that is working elsewhere shows a dot in the room switcher, with "Working now" only as hover text. There is no high-contrast theme. If something doesn't work with your setup, please open an issue.
npm test
908 checks run the server against fake claude, gh and Notion stand-ins in tests/fakes/, so they need no network, accounts or credits. Expected output: ok: all 908 checks match tests/expected.txt.
See SECURITY.md for the full model. In short: the room binds to localhost and rejects other hosts and origins; agents get read-only tools (never Bash, the tool that runs commands), write only in drafts/, and can't read secret files (.env, keys, .git) or the room's own data; every action is a proposal you approve. Agents are given only what the room gives them: your own CLAUDE.md files and the memory Claude Code keeps for a folder are not loaded for them unless you turn that on (agentsGetClaudeFiles, see Configuration), and the room checks that at every start.
These are permission rules, not a sandbox. The blocks work by name: tool names, folder paths and file-name patterns, passed to Claude Code as allow and deny rules. Nothing isolates the agents from your machine at the operating-system level, so a secret in a file with a name that isn't on the list, inside a folder you gave a room, is readable. Everything agents read can also carry prompt injection: files, web pages, Notion pages, GitHub issues and pull requests, mail previews and what the room's MCP servers return can contain text that tries to steer an agent. Give each room only what its job needs, and read a proposal before you approve it.
To report a vulnerability, use Report a vulnerability on the repository's Security tab rather than a public issue.
See CONTRIBUTING.md. Working on the room with Claude Code? CLAUDE.md has the commands and architecture.
Exec Room is licensed under the GNU Affero General Public License, version 3 only (SPDX: AGPL-3.0-only) — see LICENSE.
Copyright (C) 2026 ANDRS Projects Europe B.V.
Commercial license. Organisations that cannot, or prefer not to, use Exec Room under the AGPL can get it under a separate commercial license. See COMMERCIAL.md or write to legal@andrs.nu.
The AGPL covers Exec Room and works based on it; agent instructions, room settings and data you write yourself, without copying from Exec Room's own files, do not come under the AGPL by being used with it (this explains the license; it is not an additional permission).
Claude and Anthropic. Exec Room contains no AI model. Agents run through Claude Code, a separate program from Anthropic that you install yourself; your use of Claude Code and the Claude models is under your own agreement with Anthropic, and the Exec Room license gives no rights to them. Exec Room is an independent project, not affiliated with or endorsed by Anthropic; "Claude" and "Anthropic" are trademarks of Anthropic, PBC.
Third-party parts keep their own licenses: the Inter font in public/vendor/inter/ (SIL Open Font License 1.1, see LICENSE.txt), and the marked (MIT) and DOMPurify (MPL-2.0 OR Apache-2.0) builds in public/vendor/; the DOMPurify build also contains regenerator-runtime (MIT). Their notices and license texts are in THIRD-PARTY-NOTICES.md.
ANDRS-Projects/exec-room is a public repo on GitHub by ANDRS-Projects. Exec Room: Agents working together in rooms to help run your business. Create a local team of agents that run on Claude Code and research, review, suggest, create drafts for you. Anything that leaves a room waits for your approval. Open Source.
It declares the AGPL-3.0 license on GitHub.
Its maker has not tagged a system, so it is unknown here. Its README says what it needs.
Open "Take it" on its page: Take My Repo writes an install plan for your machine and hands it to your AI agent (Claude Code, Codex or Cursor), which asks you before each step. Take My Repo has not tested it on any machine.
Its last change on GitHub was on 2026-11-21. Take My Repo does not judge from that whether it is maintained.

The open-source app everyone uses to manage agents at work

Nobody has reviewed it on Take My Repo yet, which says nothing about its quality.
Open-source, self-hostable social media management platform. Schedule, publish, and manage content across 10+ platforms from a single dashboard. Free alternative to Buffer, Sendible, and SocialPilot.

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Reverse engineer anything with agents, from app behavior down to native binaries.
🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
MR Office: open-source document, spreadsheet, presentation, PDF, handwriting and code editors. https://mroffice.ai
Self-hosted uptime monitoring on Cloudflare Workers with public status pages
A lightweight domain knowledge accumulation system for Claude Code
Claude Code skill to create and refresh the Claude Capabilities Dashboard
A quiet corner before the busy work — sticky notes, a breathing exercise, and a thoughtful quote, tucked at the edge of your screen
Type once, get instant Unicode look-alike text styles plus a tap-to-copy kaomoji library
This page is also available as structured data, with the same sources and dates.
/api/v1/resources/gh/ANDRS-Projects/exec-roomTake My Repo has not tested this on your machine. Requirements come from the project's own documentation.