Skip to content
Staves

Per client

npx @staves/cli connect registers Claude Code in .mcp.json, Cursor in .cursor/mcp.json, Codex in .codex/config.toml, and Gemini CLI in .gemini/settings.json. Each registration includes the connection reference so different projects keep their own credentials. npx @staves/cli init writes the same four files for local boards, with --dir instead of a connection.

An env table already on the staves entry — Langfuse keys, typically — is preserved when setup rewrites the rest of it.

For clients configured outside the project, use the exact snippet printed by connect, including --connection CONNECTION_ID. A typical registration runs:

Terminal window
npx -y @staves/cli mcp --hosted --connection CONNECTION_ID

The connection ID is a local reference, not a secret. The credential stays in your user directory. Client configuration and approval requirements still apply.

After setup, activate Staves in the client you use:

  • Claude Code: approve or reconnect Staves in /mcp.
  • Codex: trust the project and restart the MCP connection or session; check /mcp.
  • Gemini CLI: run /mcp reload to load the registration.
  • Cursor: enable Staves in MCP settings. Cursor CLI shares the configuration; inspect it with /mcp list.

Have the agent call staves_brief for your board through its native tools. That verifies native access; writing configuration or passing doctor alone does not. The agent should then show a short bullet list of what it can do with the board: build, discuss, connect a project, assess gaps, compare code, test scenarios, review roles, and report progress.

A plain terminal can use the same tools through the CLI. This also keeps work moving when a coding client still needs MCP activation:

Terminal window
npx @staves/cli tool staves_interview --hosted --connection CONNECTION_ID --input '{}'

No MCP reload is required for this fallback. The same tool arguments work through MCP and CLI. Run from the project directory, --connection can be left off: connect records the connection reference in .staves/config.json, which is where the command in CLAUDE.md, AGENTS.md and GEMINI.md finds it. Those three files name no connection, so they can be committed; .staves/config.json is this machine’s.

Interviewing requires only the agent conversation and the Staves tools. Describing an existing implementation additionally requires the agent to have repository access. This release uses local CLI/stdio connections; it does not add a public remote MCP endpoint.