MCP server for Claude Code with memory management, skill execution, voice I/O, live bridge integration, and execution gating
ClawCode defines 4 tools with complete JSON schemas and descriptions, but quality is uneven. All tools have descriptions (10-300+ chars) and input schemas with proper types. However, parameter descriptions are sparse or missing entirely. live_status has no parameters; live_ack, live_emit, and live_work have typed properties but lack per-parameter descriptions explaining what each field controls. Naming is domain-specific (live_*) but not action-verb-first (should be 'acknowledge_live_input', 'emit_live_message', etc.). No output schemas documented. Error handling is absent, tools provide no guidance on failure modes or recovery. Security concerns: live_emit and live_work accept arbitrary text/progress fields with no validation or sanitization hints. Composition is reasonable (each tool has one responsibility), but the interface assumes deep domain knowledge of 'inputId', 'commandId', 'revision', 'taskId', 'sourceChannel', no discovery or lookup tools to resolve these IDs from human-friendly names.
Acknowledge an input/command ID and exact revision received from Channels, or echo its startup probe. ACK means attended, never completed. Do not guess IDs or acknowledge on behalf of another agent.
Publish an explicitly attributed public reply/progress to the Live bridge only. Success confirms storage/publication, never that the browser played audio or the user heard it; voice may be paused. id is an idempotency key. Requires an acknowledged inputId and revision for replies/progress; command results require commandId. For a proactive update use taskId and destination=live; the task must have an acknowledged Live source or an owner-adopted WhatsApp source. This never sends WhatsApp messages.
Inspect this leader's LiveBridge readiness and public work snapshot. Never returns credentials or the channel probe.
Publish a task before delegating it and update progress while it runs. Use a stable taskId, bind nativeId when known. Public fields only; no hidden reasoning/transcripts. This read model never creates or kills workers. SubagentStop means response ended, not job completed. Mark terminal status only after actual outcome is known. Use sourceInputId and sourceChannel for attributed work. WhatsApp sources require explicit owner adoption through the authenticated web controls; you cannot grant this yourself.
Parameter descriptions missing or incomplete. live_ack, live_emit, and live_work define typed properties (inputId, commandId, revision, taskId, etc.) but provide no per-parameter descriptions explaining what each field controls, when it is required, or what values are valid. LLMs cannot infer meaning from names alone.
Tool names do not start with action verbs. 'live_status', 'live_ack', 'live_emit', 'live_work' are domain-specific but lack clear verb-noun structure. Should be 'get_live_status', 'acknowledge_live_input', 'emit_live_message', 'publish_live_task' to signal intent immediately.
No output schemas documented. Tools return success/failure but no structured response format is specified. LLMs cannot plan downstream calls or extract required fields (e.g., does live_status return a status object with fields?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
No error handling or recovery guidance. Tools provide no documentation of failure modes, error codes, or what the LLM should do if a call fails. E.g., what happens if inputId is invalid? Should the agent retry, ask the user, or call a discovery tool?
No discovery or lookup tools. Tools require opaque IDs (inputId, commandId, taskId) but provide no way to resolve them from human-friendly names or search for available inputs/commands. Agents cannot discover what IDs to use.