Collection of 12 MCP servers for Claude Code: carry-forward, codebase-onboard, cost-guard, deep-engineer, dependency-doctor, diff-narrator, pair-mode, regression-sentinel, rollback, rubber-duck, scope-lock, time-capsule, and video-insight
This collection exposes only 2 tools, both extracted from a Makefile without explicit Python server implementation visible. The tool definitions are inferred from Makefile targets and their documentation comments, not from actual MCP server code. No input schema validation, type definitions, or structured output documentation is visible. Descriptions exist but are extremely brief (17-29 chars) and lack context for LLM decision-making. No parameter type information is present in any visible source. The server framework (fastmcp) suggests Python implementation, but the actual tool registration code is not provided in the submitted source excerpt, preventing verification of schema structure. Both tools lack error handling guidance and actionable recovery instructions.
Analyze a video and print the manifest summary
Generate annotated debug frames for a video showing scene boundaries and confidence scores
No visible tool registration code in source. Tool definitions inferred from Makefile only, not from Python server implementation. Cannot verify schemas, type definitions, or structured output declarations.
Descriptions are critically short (17-29 characters). 'Analyze a video and print the manifest summary' and 'Generate annotated debug frames for a video showing scene boundaries and confidence scores' lack context on WHEN to use each tool, what output structure to expect, or when one is preferred over the other.
Input schema validation not visible. Parameters 'video_path' and 'output_dir' for 'analyze' are documented in the Makefile as strings, but no JSON Schema type constraints, enums, format specifications, or length limits are visible. Cannot confirm whether min/max bounds on file paths are validated.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 32 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 41 | - | v1 |
No output schema documentation. Neither 'analyze' nor 'annotate' documents what fields are returned, what structure the response takes, or what fields downstream tools (if any) should expect. The Makefile shows 'analyze' prints metadata like 'Type', 'Duration', 'Resolution', 'Scenes', 'Confidence', 'Colors', 'Components', 'Manifest', 'Report', but this is a side-effect print, not a structured tool response.
No error handling or recovery guidance. If a video file is missing, invalid format, or analysis fails, neither tool description explains how the LLM should recover. No indication of retryable vs. fatal errors, or what alternatives exist.
Parameter descriptions missing or trivial. 'video_path' and 'output_dir' lack context: Are absolute paths required? Relative paths OK? What video formats are supported? What permissions are needed on output_dir? LLM cannot infer these constraints from names alone.
No structured response format. Makefile 'analyze' target prints human-readable key=value pairs (e.g. 'Type: <type>', 'Duration: <duration>s'). LLMs struggle to parse unstructured text output reliably. Should return JSON with typed fields: {video_type: string, duration_seconds: number, resolution: string, total_scenes: number, avg_confidence: number, dominant_colors: string[], ui_components_detected: string[], manifest_path: string, report_path: string}.
Ambiguous tool composition. 'annotate' appears to iterate over scenes and print confidence scores, but it's unclear what the semantic difference is between 'analyze' (which already computes confidence and scenes) and 'annotate'. LLM will struggle to choose correctly. These may need better differentiation or a clearer dependency relationship.