Toolkit tools¶
The Minecraft Agent Toolkit is implemented by mc-agent-bridge. Its live capabilities result is authoritative: the MCP surface is filtered to match the connected mod and vantage.
CLI¶
| Command | Purpose |
|---|---|
| mc-bridge discover | explain server-vantage port discovery without starting the daemon |
| mc-bridge run | own the mod connection and serve the loopback API |
| mc-bridge call |
make one JSON call to the daemon |
| mc-bridge watch --events chat,game | stream buffered/live events |
| mc-bridge mcp | expose the Toolkit over MCP for a Harness |
| mc-bridge forward --events ... | send selected events to one configured signed webhook |
Run status and capabilities first. Common server-vantage methods are:
| Method | Purpose |
|---|---|
| status, capabilities | connection, discovery and supported surface |
| state | authoritative world/tick state |
| player | resolve an online player by UUID or name and return live context/view |
| context | fetch a bounded chat-time bundle by context ID |
| entities | nearby entities |
| command, command_output | issue a command; the latter also waits for its answer |
| events | replay buffered events from a cursor |
| save | report world-save metadata |
| snapshot, snapshots | capture/list entity-order snapshots |
| fork, restore, verify, order | copy and validate a world fork |
| wait, mark, stop | timing, annotation and daemon control |
Client-only methods such as screen, chat, connect, world, lan and recording appear only after explicitly selecting client vantage and only when the mod advertises them.
MCP¶
MCP tools use the mc_ prefix: mc_status, mc_capabilities, mc_player, mc_context, mc_state, mc_entities, mc_command_output, mc_events, mc_snapshot and so on. The MCP process is an adapter that connects to the running daemon. It does not own the game connection.
A Harness should:
- call mc_status and mc_capabilities;
- resolve the relevant player when the request needs player context;
- fetch only the live world data needed;
- use mc_context promptly when an event supplied a context ID;
- treat identity as context, not command authorization.
Event forwarding¶
The optional forwarder is receiver-neutral. It signs each outbound POST, retries transient failures and reuses a stable event ID for deduplication. URL and secret come from environment or a protected config file. It does not run a model, manage a session or deliver Agent replies.
See Hermes and unattended operation and the Toolkit reference.
Repository utilities¶
| Tool | Purpose |
|---|---|
| lab_server.py | provision and control a headless Fabric lab |
| fork_verify.py | inspect, restore and compare recorded forks |
| fake_player.py | create and drive a Carpet fake player |
| game_cmd.py | run a game command and print feedback |
| mcp_probe.py | call one MCP tool against a running Toolkit |
| launch_instance.py | launch a client without a GUI launcher |
| smoke_offline.py | compatibility test for the old daemon/agent-loop path |
Legacy agent-loop¶
mc-agent-loop is not part of the Toolkit surface. It remains an optional compatibility listener with echo and temporary hermes backends. Its default chat trigger is @agent. Codex and Claude Code run as user-driven Harnesses and must not be described as loop backends.