MCP server that interfaces with the Theater WebAssembly actor system, providing tools to manage actors, send messages, and open channels
Theater MCP has 8 tools with basic schemas and descriptions, but significant gaps prevent production use. All tools have descriptions (10-50 chars, below the 50-200 char baseline for LLM optimization) and input schemas with required parameters. However, descriptions lack context on WHEN to use each tool, prerequisites, and recovery guidance. Parameter descriptions are minimal (e.g., 'ID of the actor to stop' for actor_id). No output schemas documented. Error handling is present in code (connection retry logic) but not surfaced in tool definitions. Base64 encoding requirement for message data is underdocumented. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear WRITE vs READ_ONLY risk classifications. Composition is reasonable (single responsibility per tool), but chaining is unclear, e.g., open_channel returns channel_id, but send_channel_message accepts it; this dependency is not documented in descriptions.
Close an open channel
Open a bidirectional channel to an actor
Send a request message to an actor and receive a response
Restart a running actor
Send a message on an open channel
Send a one-way message to an actor
Start a new actor from a manifest
Stop a running actor
Descriptions are 10-50 characters, well below the 50-200 char baseline for LLM optimization. Descriptions lack WHEN to use, prerequisites, and recovery guidance. E.g., 'Send a one-way message to an actor' does not explain when to use send_message vs request_message vs open_channel.
No output schemas documented. LLMs cannot plan downstream calls or extract required fields. E.g., start_actor returns actor_id and status, but no schema tells the LLM what fields to expect or their types.
Base64 encoding requirement for message data (send_message, request_message, send_channel_message) is underdocumented. Parameter descriptions say 'Base64-encoded message data' but do not explain why, what format the decoded data should be, or how to encode. LLMs will struggle to construct valid payloads.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | <=2025-11-25 | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications. request_message is READ_ONLY; start_actor, stop_actor, restart_actor, send_message, open_channel, send_channel_message, close_channel are WRITE. Annotations enable safer agent planning.
Tool chaining dependencies are not documented. open_channel returns channel_id; send_channel_message requires it. No description hints that send_channel_message depends on open_channel or what happens if channel_id is invalid.