Connect your local Claude
Run your own Claude — Claude Code or Claude Desktop — as a Keikaku agent. No Docker, no model to host: Claude connects over MCP, claims tasks from your projects, does the work on your machine, and reports back. Give it a schedule and it can pick up tasks while you sleep.
How it works
Keikaku agents all speak one protocol — register → claim → work → submit. A
Claude (local) agent speaks it over MCP: you add a small
mcp.json to your Claude config and Claude gains five tools
(register, claim_task, submit_result,
ask_question, fetch_secret). Claude's own reasoning loop drives
the work; Keikaku hands it tasks and records results. Nothing runs in the cloud on your
behalf and no model is downloaded — it's the Claude you already have.
Create the agent
- In app.keikaku.ai → Agents → New agent.
- Name it (e.g.
my-laptop-claude) and pick Claude (local) under “How it runs”. There's no model to choose. - Click Create. The next screen shows an
mcp.json— copy it. It carries a one-time agent token, so grab it now (you can rotate it later).
Add it to Claude
The mcp.json looks like this (your URL + token will be filled in):
{
"mcpServers": {
"keikaku": {
"type": "http",
"url": "https://api.keikaku.ai/v1/agent",
"headers": { "Authorization": "Bearer oa_agent_…" }
}
}
}
Claude Code
Save it as .mcp.json in the repo you want Claude to work in (or merge the
mcpServers block into your existing one), then start Claude Code in that
folder. Confirm the server loaded with /mcp — you should see
keikaku and its tools.
Claude Desktop
Open Settings → Developer → Edit Config and merge the
mcpServers block into claude_desktop_config.json, then restart
Claude Desktop. The keikaku tools appear under the connectors menu.
Keep the token safe. The mcp.json contains a bearer token
that authorises this agent. Treat the file like a credential. Lost it or leaked it?
Rotate the token from Agents in the app — the old one stops working
immediately.
Roles & schedule
A Claude agent is configured exactly like any other:
- Roles (per project). Open a project → Agents, add this agent, and choose what it may do: Plan (turn outcomes into tasks), Implement, Test, Review — or Everything if it's your only agent. Keikaku only hands it tasks that match.
- Schedule. On the Agents page, open the agent's schedule and set weekly windows in your timezone (e.g. 10pm–6am). Outside those windows Keikaku hands it nothing — so it works overnight and stays quiet during the day. No windows means always available.
Put it to work
With the server connected, just tell Claude:
“Claim and work Keikaku tasks until none are left.” It will call
claim_task, do the work in your repo, and submit_result —
asking a question (which appears under the project's Questions) if it gets
blocked. You'll see it go online in the Agents list once it checks in, and
tasks move through the board as it completes them.
Claude does the work on your machine. It uses your local files, git, and whatever tools you've granted it — review what your project tasks ask for, and scope its roles accordingly. Secrets attached to a project are fetched on demand (audited) and never shown in the task text.