Driving capsule-corp from an MCP client¶

capsule-corp speaks the Model Context Protocol, so any MCP client can browse and drive the catalogue — not only the agent that writes the capsules.

capsule mcp

That serves the catalogue over stdio. Point a client at it by adding a server entry; in Claude Code, for example:

{
  "mcpServers": {
    "capsule-corp": {
      "command": "capsule",
      "args": ["mcp"],
      "cwd": "/path/to/your/catalogue"
    }
  }
}

Tools¶

Tool	What it does
`capsule_list`	List capsules, optionally filtered by folder or status
`capsule_search`	Full-text search over titles, questions, and reports
`capsule_get`	One capsule in full: pre-registration, results, verification, report
`capsule_new`	Create a skeleton (does not call a model)
`capsule_design`	Run the design phase — spends money
`capsule_freeze`	Lock the pre-registration — irreversible
`capsule_implement`	Run the implementation phase — spends money
`capsule_run`	Execute on `local`, `slurm`, `ssh`, or `modal`
`capsule_verify`	Deterministic checks, then the blinded judge

Resources¶

capsule://index — every capsule with its current status.
capsule://<id> — one capsule in full.

The resources are the part worth caring about. They let an agent ask "what have I already tested here, and what did it show?" before spending compute on a question the catalogue has already answered — including the ones whose hypotheses were refuted, which are exactly the results that never get written up anywhere else.

Going the other way: giving capsules MCP tools¶

pi has no built-in MCP support; it is deliberately minimal and pushes MCP, sandboxing and sub-agents out to extensions. The third-party pi-mcp-adapter provides it:

npm install -g pi-mcp-adapter

capsule doctor reports whether it is installed. Track earendil-works/pi#563 for an official path.