Offline hybrid code-search MCP server — one Rust binary, no model, no database. BM25 + vector retrieval over local repositories via the Model Context Protocol.
Two tools with complete input schemas and descriptions, but quality varies significantly. search_code has a detailed, LLM-optimized description (280+ chars) with well-typed parameters including enums and numeric constraints. reindex is minimal (67 chars) and lacks context on when/why to use it. Both tools lack documented output schemas, critical for agent planning. search_code's description is verbose but informative; reindex's is dangerously terse. No error handling guidance, no recovery hints. Parameter descriptions are present but some lack constraints (e.g., 'path' accepts any string with no validation hints). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Reindex a repository, optionally forcing a full wipe-and-rebuild instead of an incremental diff.
Search a code repository with hybrid BM25 + vector retrieval (RRF-fused). Lazily indexes the repo on first call. Returns the top-k hydrated chunks with file, line range, kind, signature, snippet, score, and the file's imports/exports. Set adaptive=true (opt-in, default off) to pick fusion weights from the query shape (identifier/snake_case/camelCase up-weight BM25; multi-word NL up-weights vector); explicit bm25_weight/vector_weight always override it. LIMITATION: adaptive is lexical-only (no corpus signal) and the vector arm is near-noise with the default feature-hash embedder.
reindex description is 67 characters, below the 10 - 1024 baseline and lacks WHEN/WHY context. LLMs cannot determine when to call reindex vs relying on lazy indexing in search_code.
No output schemas documented for either tool. Agents cannot plan downstream calls or extract required fields (e.g., does search_code return file paths, line numbers, scores?). Critical for tool chaining.
No error handling guidance. If search_code fails on an invalid path or reindex times out, the LLM has no recovery hint. Missing pattern:recovery-guide.
No tool annotations despite clear risk profiles. reindex is WRITE (destructive); search_code is READ_ONLY. Missing destructiveHint/readOnlyHint annotations prevent agents from reasoning about side effects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
'path' parameter in both tools accepts any string with no validation hints. Description should state: 'Absolute filesystem path (must exist and be readable)' to guide LLM input and enable early validation.