Skip to main content

​
Any MCP client, by hand

Merv is a plain HTTP MCP server. Register it the way your client registers any remote server:
  • URL: https://experiments.rapidreview.io/mcp (or your self-hosted brain’s /mcp)
  • Auth: if the client implements MCP OAuth discovery (RFC 9728/8414), dynamic client registration, PKCE, and refresh, leave headers empty and let the 401 start the browser flow. Otherwise create a project key at rapidreview.io/merv and send Authorization: Bearer <key>.
  • Version header: clients send X-RP-Client-Version; anything below the floor published at /api/meta gets HTTP 426.
Then paste the Merv skill text the agent needs (the research-workflow skill at minimum) into its instructions, and use a second session for reviews. merv-client env prints a ready-made header-based MCP entry for headless clients.

​
Becoming a native adapter

A first-class client gets: one-command install, OAuth by default, skills delivered natively, reviewer subagents, and a long-run wake recipe. Concretely that means a thin adapter under merv/clients/<id>/ (or a marketplace manifest), reusing the canonical skills/ and agents/ trees, plus:
  1. A URL-only MCP registration — no credential header in anything committed (tests enforce this).
  2. A documented sign-in command or UI path.
  3. Skills delivery: auto-discovery, manifest, or the hosted content-versioned catalog.
  4. Reviewer handoff: auto-discovered agents, plugin-injected subagents, or a delegate mechanism.
  5. A merv_run wake recipe (see Long runs).
  6. A row in the client matrix and an entry in the UI’s Connect list.
Open an issue or PR on rapidreview-io/Merv — the adapter table in merv/docs/CLIENTS.md is the current contract.