sportsclaw

Appearance

Connecting MCP Servers ​

sportsclaw can pull in tools from external Model Context Protocol servers, so your agent can reach workflows and services beyond sports data.

Add a server ​

bash
sportsclaw mcp add <url> --name <name> --token <token>   # connect an MCP server
sportsclaw mcp list                                      # see connected servers
sportsclaw mcp remove <name>                             # disconnect one

mcp add accepts:

  • --name <name> — a short name for the server (auto-derived from the URL if omitted).
  • --token <token> — a bearer token, if the server needs auth.
  • --description <text> and --timeout <ms> — optional metadata and per-call timeout.

Once connected, every tool that server exposes becomes available to the agent, alongside the built-in sports tools. There is no read/write classification: if the server offers delete_document or execute_workflow, the agent can call them.

Restricting a server to specific tools ​

Set tools on the server entry to register only the ones you name. Anything the server advertises that is not on the list is never given to the model:

~/.sportsclaw/mcp.json — the server name is the top-level key:

json
{
  "my-pod": {
    "url": "https://my-pod.org.machina.gg/mcp/sse",
    "tools": ["search_documents", "get_document", "health_check"]
  }
}

SPORTSCLAW_MCP_SERVERS takes the same object as a JSON string, so the relay is configured the same way. Run with --verbose to confirm what registered — the log reads mcp: "my-pod" has 3 tool(s) (filtered from 41).

Prompt rules are not enforcement

Telling the agent in its instructions not to call a write tool is not a control: the tool is still registered and the model can still call it. tools is the enforcement point.

The approval gate does not help here either. It is declarative — it fires when a tool spec defines needsApproval — and MCP tool specs are built from the server's advertised schema with no such field, so an MCP write runs without a confirmation prompt. The gate is also interactive-only: outside a TTY it fails closed, which is not what you want for a relay.

If a server exposes writes you do want, one workable shape is to allowlist only its read tools and have your own application perform the writes after a user confirms them — the agent proposes, your app executes.

Pod memory needs write tools

SPORTSCLAW_MEMORY_PROVIDER=pod stores memory through the pod's own create_document and update_document. A read-only allowlist denies those: leave the provider at auto (it falls back to file memory) rather than pod, or the run throws.

Where tokens live ​

Tokens are kept out of the config file. mcp add --token … writes the value to ~/.sportsclaw/.env as SPORTSCLAW_MCP_TOKEN_<NAME> (the name uppercased, hyphens to underscores), while ~/.sportsclaw/mcp.json holds only the URL and metadata. At connect time the engine injects the token as an X-Api-Token header. You can also set the env var yourself instead of passing --token.

What you can connect ​

Any MCP server — including Machina "pods," which surface workflows, agents, and connectors the agent can call as part of answering a question.

Connecting a Machina pod? Use machina connect

For Machina pods, prefer sportsclaw machina connect — it signs you in, mints a durable key, and registers the pod automatically (no URL to copy). Pods it registers are tagged provider: "machina", which is how doctor and the agent's get_agent_config recognize them. Use mcp add for arbitrary, non-Machina MCP servers.

sportsclaw connects to MCP servers

sportsclaw is an MCP client: it consumes tools from other servers. It does not run as an MCP server itself.