RiffCast MCP server — exposes plan/render/review as MCP tools over stdio.
RiffCast MCP exposes 4 tools via STDIO. Tool definitions are registered in server.ts with names, descriptions, and input shapes (Zod schemas). However, source code for individual tool files (render.ts, review.ts, propose-directions.ts, skill.ts) is not provided, only package.json and server.ts are visible. This means tool schemas, parameter descriptions, and error handling cannot be verified. Additionally, no output schemas are documented, parameter descriptions are not visible, and error handling patterns are absent from the visible code. The server appears functional but lacks the transparency needed for confident LLM integration.
Tool implementation files not provided in source code review. Cannot verify parameter descriptions, input schemas, output schemas, or error handling for any of the 4 tools.
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, breaking downstream tool chaining and forcing agents to guess at response structure.
Tool names lack clear action verbs. 'propose-directions' and 'skill' are vague, does 'skill' retrieve, create, or validate a skill? Does 'propose-directions' suggest next steps or return available options? Ambiguous names reduce LLM selection confidence.
No error handling guidance visible. If render fails, does the LLM retry, ask the user, or abort? No recovery patterns documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 27 | 2026-07-28+ | v2 |
render tool marked as WRITE risk but no confirmation/dry-run pattern visible. Destructive operations should support confirmation before execution to prevent accidental data loss.