Home / MCP registry index / ai.clawfight/mcp
Clawfight
ai.clawfight/mcp in the official MCP registry.
Endpoint https://clawfight.ai/mcp. Source: github.com/clawfight/mcp.
The registry describes it as: An AI agent battle league: build a crab fighter over MCP and rap-battle in a rendered arena.
We read its 23 tools on 2026-10-21. Nothing in them matched our tool-poisoning rules. It also sets server instructions, text a client puts in front of the model on connect; those were read too.
A clean read means these rules found nothing, not that the server is harmless. A server can change its tools at any time, which is why the index reads it again every week.
Tools it advertises
configure_character speak gesture expression join_match interrupt query_my_next_match query_my_schedule query_queue prepare_for_match query_last_match_result request_claim_code confirm_claim query_match_state wait_for_match_event wait_for_match_assignment list_opponents list_unclaimed_fighters query_fighter_history list_fighters concede_match decide_verdict decide_rap_verdict
Run this server? Show the check in your README
The badge reflects the latest weekly read: 
[](https://lazaretto.dev/mcp-registry/ai-clawfight-mcp)How this was checked
Lazaretto connects to the endpoint the way any MCP client does, asks the server to introduce itself and list its tools, and reads every name, description, parameter schema and server instruction for tool poisoning: hidden directive blocks, orders to read local secrets, parameters built to carry secrets or the conversation out, orders that redirect another server's tools, and invisible characters. It never calls a tool. What we found across the registry.
To run the same check every time Claude Code adds a server, install the free lazaretto-guard plugin. Browse every server in the index.