MCP servers

MCP servers give the agent more tools: a database, a documentation search, a lab system. In code flower they come only from config.yaml on the computer that starts the sandbox, under mcp_servers:, and every sandbox’s agent gets them. The agent cannot add servers itself.

A server at a URL

mcp_servers:
  docs:
    url: https://mcp.example.org/mcp
    transport: http                       # the default; sse works for Claude Code and OpenCode only
    headers: { Authorization: "Bearer <token>" }

A command in the sandbox

The command runs inside the sandbox (stdio), so it must be installed there:

mcp_servers:
  files:
    command: files-mcp                    # on the sandbox's PATH, or a full path
    args: [--root, /codeflower-state/workspace]
    env: { TOKEN: "<token>" }

Install it into the project’s environment, which persists: uv pip install <package> for a Python server (the command lands in the project’s venv, on PATH), or npm install --prefix /codeflower-state/env/mcp <package> for a Node one, with the full path /codeflower-state/env/mcp/node_modules/.bin/<command> as command. The agent can do this for you. Prefer an installed command to npx/uvx, which download the server at every start.

Rules

• Names are 1 to 32 letters, digits, - or _.
• A server with env or headers usually holds a token: config.yaml must then not be readable by other users (chmod 600 ~/.codeflower/config.yaml), or code flower refuses to start.
• Restart the sandbox after a change; the list is read when it starts.
• A server that fails to start is left out, and the chat says so.
• With network isolation, the host of each url server is allowed by itself. A server on a private address, and whatever a command server connects to, need lines in the allowlist file.