Contree MCP demonstrates solid tool coverage (19 tools) with mostly consistent naming conventions and present descriptions. All tools start with action verbs (list_, get_, set_, run_, import_, registry_, upload_, download_, rsync_, cancel_, wait_, grep_). Input schemas are explicitly visible in the code with proper type definitions. However, there are significant gaps: (1) Many descriptions are terse (10-30 chars), falling below the 50-200 char LLM-optimized baseline. (2) Parameter descriptions exist but are often minimal, 'Image UUID or tag' repeats across multiple tools without clarifying when to use each. (3) Output schemas are NOT documented anywhere in the provided code, there is no indication of what fields tools return. (4) Error handling guidance is absent, no recovery hints, no categorization of retryable vs fatal errors. (5) Some parameter relationships are undocumented (e.g., the relationship between 'disposable' and 'wait' in the 'run' tool). (6) The tool 'get_guide' appears to be a meta-tool that returns documentation sections rather than operational data, which is unusual and may confuse agent planning. These gaps place the server in the 'Fair' to 'Good' range, competent but not production-grade.

• Document output schemas for all 19 tools. Add a 'Returns' section to each tool's docstring listing field names, types, and meanings. Example: 'Returns: {images: [{uuid: string, tags: string[], created_at: ISO8601}], total_count: number}'.
• Expand tool descriptions from 10-45 chars to 50-150 chars. Include WHAT the tool does, WHEN to use it vs similar tools, and key side effects. Example: 'Pull an OCI image from a registry and cache it locally for later execution. Use this to stage container images before calling run(). Irreversible, canceling mid-import may leave partial state.'
• Clarify parameter descriptions with resolution logic and constraints. Example for 'image' param: 'Image identifier. Accepts UUID (exact match) or tag (prefix match; fails if ambiguous; use list_images to disambiguate).'
• Add error handling guidance to WRITE and IRREVERSIBLE tools. Example for 'run': 'If execution fails, check operation status with get_operation(). Failed runs with disposable=true auto-cleanup; with disposable=false, manual cleanup may be needed. Retryable unless the microVM was terminated.'
• Document parameter relationships in 'run' tool. Clarify semantics: 'wait=true blocks until completion (timeout: 5m); wait=false returns immediately with operation_id. disposable=true (default) discards VM state after completion. disposable=false persists state in the image.'
• Consider retiring 'get_guide' or reclassifying it as a resource (documentation) rather than a tool. If kept, rename to 'get_documentation' and make it clear this is for agent self-education, not task execution.

Inferred effective spec: 2026-07-28+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry