Skip to content
fajarhidePublic

About

AI teammates for your whole team: give them a goal, they do the work, ask before anything leaves and keep every change with Undo. Self-hosted on the smallest VPS, any model provider.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Kamebox

AI teammates that get the work done, wait for your yes, and keep the receipts.

Most AI agents work for one person. Kamebox is built for a team.
Your people and a team of bots share a roombox and work toward the project's goal.
Bots read on their own, anything that leaves waits for a person to tap Allow,
and every change lands in a record with Undo.

An owner starts the SaaS Builder pack and its bots join the roombox. Investigator reads web pages on its own, Builder asks before creating a code repo, and the owner taps Allow on a phone.

Try the live demo · Run it yourself · Read the code

The demo is shared, resets every 2 hours and allows 10 messages.

Kamebox is early. It runs end to end on your own server today. Not yet lists what is missing.

What it does

A personal assistant answers one person in one tab, and the work stays in that chat. Kamebox puts the bots where the whole team talks, and what they do ends up in records the team can open next month.

  • Start from a Pack. Every roombox starts from one. Office ships with Kamebox (Operations, Finance and Sales), and so does SaaS Builder, a nine bot team with milestones, a budget and seven boxes. More packs and box templates come from a Pack registry in the Pack Library.
  • Talk to them like a team. Bots reply in the chat. Reply to any message, @mention a bot or a person, and see what is waiting for you under Mentions, behind the bell.
  • Plan first when it matters. Plan mode (Shift+Tab) gets you a plan to reword and approve before any bot acts. The Tasks tab shows milestones and tasks ticking off live.
  • Routines on a schedule, such as a daily brief, in the org's time zone. A routine runs with nobody watching, so it reads on its own and every write it wants waits as an approval.
  • Boxes are the record. Bots write into boxes (Findings, Decisions, Leads, invoices, stock). Each change shows in the roombox as a card with Undo.
  • Your currency and language. Owners pick the org's currency (IDR included) and language (English or Bahasa Indonesia). Bots answer in the language you write in.
  • Your model provider. ANTHROPIC_API_KEY by default, or KAMEBOX_PROVIDER=openai for any OpenAI-compatible endpoint.
  • Bug reports with evidence. Anyone, viewers included, files a report with screenshots from inside the app. Owners get a prefilled GitHub issue link that leaves the files and roombox messages out.
  • Check a running install. bun run verify <url> signs in as an owner and checks your live install from outside, behind its proxy: sessions, a viewer that cannot post, a member that cannot invite or change a role, setup that stays closed and no secret in a response.

The shell

Three promises, each pinned by a test.

Ask first. A connector tool runs on its own only when its name reads (get, list, search) and its server marks it read-only. Everything else becomes an approval card: Allow once, Always allow this tool or Deny. Sending email, publishing, deploying and spending always ask. Once a bot has read outside data, its first call to a new connector asks too, so text planted in a web page cannot steer the team somewhere new. Pinned by tests/shell.test.ts and tests/mcp.test.ts.

Keep a record. Every approval keeps who asked, who decided, what ran and what came back. An approval runs once, even with two approvers tapping at the same moment. Pinned by tests/approvals.test.ts.

Undo. Every box change a bot makes is a card with Undo. Undo refuses instead of guessing when someone edited the field since. A call that already left through a connector cannot be pulled back, which is why it asks first. Pinned by tests/records.test.ts.

Run it yourself

Tip

Runs on the smallest VPS. 1 vCPU, 512 MB RAM and 1 GB of disk is enough. On a live install it uses about 55 MB idle and under 100 MB at peak, from a 212 MB image, with all data in one SQLite file. Give it 1 GB of RAM if you run connectors.

On a box with Docker:

git clone --depth 1 --branch <latest tag> https://github.com/fajarhide/kamebox.git /opt/kamebox
install -d -m 700 /etc/kamebox && printf 'NODE_ENV=production\nPORT=3000\nKAMEBOX_SETUP_TOKEN=%s\n' "$(head -c 24 /dev/urandom | base64)" > /etc/kamebox/env
cd /opt/kamebox/deploy && docker compose up -d --build

Kamebox then listens on 127.0.0.1:3000. Whoever finishes setup first becomes the owner, so setup asks for the KAMEBOX_SETUP_TOKEN above (grep SETUP /etc/kamebox/env). deploy/README.md covers HTTPS with the included Caddy, your own proxy, backups, updates and bun run verify.

To hack on it, install Bun 1.4 or newer, then bun install, cp .env.example .env and bun run dev, and open http://localhost:3000. bun run seed adds a demo team (Maya, Alex, Sam and Lee) to look around with. It has no passwords, so keep it off the internet.

Connectors

Bots reach outside tools through MCP servers, and every tool a server offers goes through the shell above. A server starts only when its command is listed in KAMEBOX_MCP_COMMANDS (comma separated, empty starts nothing). List the full command line, such as KAMEBOX_MCP_COMMANDS="bun /opt/mcp/github-server.ts", and an entry with a space must match the whole line.

The built-in web reader starts with no allowlist entry: give a connector the launch builtin:web-read and its bots can read a web page as plain text, with private and internal addresses refused. After a bot has read outside data, any further web read asks first.

Never list a bare interpreter (npx, bunx, bun, node, python): it lets any owner run any code on the server as the app user, and that user can read the database.

Meet the bots

Bot avatars. Top row: Operations, Finance and Sales. Below: Strategist, Investigator, Builder, Growth, Guardian, Operator, Designer, SEO and Social.

Bots are hexagon shell creatures and people are circles, so you can tell them apart in a busy roombox. Owners rename any bot and give it a soul on the Bots page.

How Kamebox compares

The best known AI agents are built for one person. Kamebox is built for a team.

Built for Works toward a goal Who approves Work kept as records with Undo Self-host, your own data Source Price
Kamebox A team ✅ Milestones, budget, tasks ✅ The project's starter and the owners ✅ In boxes ✅ One SQLite file FSL-1.1-ALv2, Apache 2.0 after two years Free to self-host
Claude Duet One person, one ongoing thread Not listed You, before it sends, pays, signs or deletes Not listed No, hosted No, hosted In Pro and Max, US waitlist
Claude Cowork One person, on a folder of files Not listed Not listed Not listed No, hosted No, hosted Paid plans
Grok Bot Named bots that use a computer Not listed Pauses for approval Not listed No, hosted No, hosted SuperGrok at $30/month, or a paid Cursor plan
OpenAI Dots One always-on agent Not listed Your rules: may, ask first, never Not listed No, hosted No, hosted Eligible Pro and Business Premium plans
Meta Muse One person, in the Muse app or WhatsApp Not listed Checks before sensitive actions Not listed No, runs on a machine Meta operates No, hosted Not listed
OpenClaw An agent you run yourself Not listed Not listed Not listed Yes MIT No software fee
Hermes Agent An agent on your own machine Not listed Not listed Not listed Yes MIT No software fee, optional Nous Portal credits

What the table cannot show is where the work ends up. In Kamebox every bot change lands in a box (Findings, Decisions, Product development) with who asked, who allowed and Undo, so next month the team opens the record instead of scrolling a chat. A pack arrives ready: nine bots with their boxes and their ask lists, nothing to configure but the connectors. In a busy roombox a coordinator picks at most two bots to answer, and often none.

Why not a personal agent?

The others win on computer and browser use (Grok Bot, Dots and Duet run one for you), thousands of connectors (Duet has 6,000+ in the Claude Marketplace), a chat app you already have (Muse in WhatsApp), any model on your own machine (OpenClaw, Hermes Agent) and no setup; Kamebox reaches a browser only through an MCP server you wire up.

Use Kamebox when the work belongs to a team. Everyone sees the bots' work in one roombox, so Maya reads Investigator's findings without asking Alex. Sam can ask Builder to create the code repo, but it waits for Alex to allow it. And the findings, the decisions and the backlog are still there next month.

If the work is yours alone, use a personal agent. If it belongs to a team, use Kamebox.

Details as of October 2026.

Not yet

  • Real email sending. Approving an email records the decision; nothing is sent.
  • Spending against the money budget. Nothing counts spend against it yet (#21). The token budget does stop new bot turns.
  • Running code. Builder writes files in the project's workspace, and nothing in Kamebox runs them.
  • More Packs.
  • WhatsApp.
  • A translated UI. Bots write in Bahasa Indonesia; the screens are still English.

Contributing

The most useful contribution needs no code: tell us a project your team would hand to a team of bots. Open an issue with the goal, what you would type in the roombox and what you would expect back.

Code is welcome too: start with CONTRIBUTING.md, and point your AI coding agent at AGENTS.md.

License

FSL-1.1-ALv2. Use it, change it and self-host it for free, for your own team or for the teams you work with. The one thing it does not allow is offering Kamebox itself as a competing product. Each version becomes Apache 2.0 two years after its release. Versions up to v0.1.7 were released under AGPL-3.0 and stay under it.

Packs carry their own license, Apache 2.0 or MIT, set in each Pack file.

The name "Kamebox" and its logo are not covered by the license; forks need their own name.

About

AI teammates for your whole team: give them a goal, they do the work, ask before anything leaves and keep every change with Undo. Self-hosted on the smallest VPS, any model provider.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages