Hand work to other agents
Your main agent writes a numbered brief and starts another agent to do it.
You pick the provider and model tier for each job, so each task goes to the model that fits it and cheap models do the mechanical work.
Coming soon
Agents that spawn and talk.
Beehaus is one MCP server that lets the AI coding agent you are talking to hand work to other agents (“bees”) and lets agents message each other.
Two halves
Spawn bees for the work. Let agents talk.
Route each job to the provider and model tier that fits it, and give every session an addressed inbox.
Addressed agent inboxes
Every session can register an inbox with an address. Agents send each other plain-text messages, wait for replies, and acknowledge them.
Messages are data, never approvals: a message from another agent approves nothing.
How it works
Four steps from brief to verified diff
Write a numbered brief
Your main agent writes a numbered brief and picks the provider and model tier for the job.
Approve or use a trust window
By default a spawn request parks until you approve it from a link. It runs at once only while a trust window you opened is active.
Run on your machine
The bee runs locally under an allowlist (tests, builds, local git commits) using your own provider account. It can never git push.
Verify from transcript & diff
Beehaus reports liveness from the transcript and mailbox, lets the parent read transcripts, send follow-ups, check diffs, or abort.
Safety & verification
You stay in control. Verify, don’t trust.
Built for local machines with strict boundaries between human approvals, agent outputs, and execution.
Human approval by default
A spawn request is parked until a human approves it from a link. It runs immediately only while a trust window you opened is active.
Allowlist & rate limits
Bots run under an allowlist (tests, builds, local git commits) and can never git push. Rate limits apply per hour and per day.
Real liveness, not run IDs
A run ID never means the agent is running. Beehaus reports liveness from the transcript and mailbox.
Inspect, steer, or abort
The parent agent can read the bot’s transcript, send follow-up messages, check its diff, and abort it at any time.
Untrusted data boundary
Everything a bot writes is handed back marked as untrusted data, never as instructions. A message from another agent approves nothing.
Runs where you work
Bots run on your own machines (a Linux box, a Mac, a Windows machine) with your own provider accounts. Works with any MCP client.
Quota at a glance
Move work to a lane that still has budget
One call shows each provider’s remaining quota, when it resets, and a forecast of when it runs out—so the orchestrator routes around limits before jobs stall.
- Remaining quota per provider
- Exact reset schedule
- Forecast of when budget runs out
Providers
Supported coding agent providers
Use your own provider accounts or local models across any MCP client.
- Google Antigravity Gemini models
- Z.ai ZCode GLM
- OpenAI Codex CLI Codex CLI lane
- Claude Code Claude Code lane
- OpenCode Your own local models
- Xiaomi MiMo Code MiMo Code lane
- Moonshot Kimi Code Kimi Code lane
Not available yet
€10 one-time
One-time payment at launch. Runs on your own machines with your own provider accounts.
Tell me at launchFAQ
Common questions
What is Beehaus?
Beehaus is one MCP server that lets the AI coding agent you are talking to hand work to other agents (“bees”) and lets agents message each other.
Is Beehaus available now, and what will it cost?
Beehaus is not available yet. At launch, it will be €10 one-time. You can email hello@beehaus.dev via the “Tell me at launch” button to be notified.
Which providers are supported?
Beehaus supports Google Antigravity (Gemini models), Z.ai ZCode (GLM), OpenAI Codex CLI, Claude Code, OpenCode (your own local models), Xiaomi MiMo Code, and Moonshot Kimi Code.
Where do the agents run?
Bots run on your own machines—a Linux box, a Mac, or a Windows machine—using your own provider accounts, and Beehaus works with any MCP client.
How do I stay in control of spawned bots?
By default, a spawn request parks until a human approves it from a link. It runs immediately only while a trust window you opened is active. Bots run under an allowlist (tests, builds, local git commits), can never git push, and are subject to hourly and daily rate limits.
Can one agent approve another agent’s work or send commands via messages?
No. Messages are plain-text data, never approvals: a message from another agent approves nothing. Everything a bot writes is handed back marked as untrusted data, never as instructions.
How does Beehaus verify whether a bot is actually running?
A run ID never means the agent is running. Beehaus reports liveness from the transcript and mailbox, and lets the parent read the bot’s transcript, send follow-up messages, check its diff, and abort it.
How does quota tracking work?
One call shows each provider’s remaining quota, when it resets, and a forecast of when it runs out, so the orchestrator can move work to a lane that still has budget.