Calling a tool

crab call notes echo '{"query": "hi"}'

crab call <service> <tool> [arguments] calls a tool of a service on the account the context is signed in to, and prints its result. On Cloud, <owner>/<service> calls another account's public tool, as crab call crabtalk/hackernews top; the local host has no accounts and refuses it. The arguments are a JSON object, read from stdin when omitted. It exits 2 when the tool ran and failed. It also prints what the run used and cost to run; see Cost of a call.

crab inspect <owner>/<service> prints how to call one before calling it: its README, its public tools, their arguments and what one call may use.

An agent calls it over MCP, with an API key.

What a call answers

• A tool that ran answers whether it succeeded or failed: done with its result, or failed with the message it gave. done holds the result as JSON text, parsed a second time.
• A call to a private service, or to a tool that is not public, from outside its account finds nothing.
• The response says what the call used, and a deprecated tool's response says it is deprecated, and when it is sunset once that is set. See Changing an interface.

Query tools

A #[tool(query)] tool only reads, and is also called with GET, its arguments in the URL.

• Its answer carries an ETag. Sent back, it answers "not modified" without running the tool while the service's data is unchanged.

Limits

• Cloud applies shared account quotas to REST and MCP, plus concurrency limits. Nodes also enforce per-caller limits. See Request limits for both layers and how to handle refusals.
• An account's services hold at most the memory its plan allows at once: a call is refused while they hold it, and one growing past it fails.