MCP server for RepoTracer - a repository scout engine for AI coding agents providing read-only repository investigation tools
RepoTracer MCP server defines 6 read-only tools with generally clear names and descriptions. Most tools have reasonable input schemas with parameter descriptions. However, significant gaps exist: output schemas are not documented anywhere in the source code provided, error handling guidance is minimal, and several tools lack clear guidance on when to use them relative to similar tools. The `repo_scout` tool is the centerpiece but lacks explicit pagination/limits documentation. Parameter descriptions are present but often generic (e.g., 'Optional' without explaining consequences). No evidence of tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite all tools being read-only. Tool names follow verb_noun convention adequately, but descriptions could be more LLM-optimized per the 50-200 char baseline.
Find files matching a glob pattern. Supports wildcards and glob syntax. Results limited to 100 files and 32 KiB output.
Search files by regex pattern with optional path filtering. Supports output modes: count, content, or truncated_content.
Read file contents with optional byte range limiting. Supports offset and limit parameters for large files.
Query an on-demand, content-hash refreshed syntax index for Rust, Python, JavaScript, TypeScript/TSX, or Go. Modes: definitions, references, outline. References are same-name syntactic occurrences, NOT resolved callers. Use exact symbol names, optional path relative to the current target or absolute for related evidence, and next_offset for pagination. Read returned ranges to establish behavior; use Grep for unsupported languages, aliases, generated code, and dynamic dispatch. Output reports parse errors, exclusions, and truncation. No source is injected until this tool is called.
Execute a read-only shell command in the repository context. Returns command output as text. Available only in specific configurations.
Output schemas completely undocumented. No tool defines what its response contains, field names, types, or structure. LLMs cannot plan downstream tool calls or extract required data.
Tool descriptions lack LLM-optimized guidance on WHEN to use each tool vs similar alternatives. E.g., Grep vs Symbols (both search code); repo_scout vs Symbols (both find code). Descriptions don't disambiguate for agent decision-making.
No tool annotations (readOnlyHint, idempotentHint, destructiveHint) despite all tools being read-only and idempotent. Annotations are current MCP spec features (2026-07-28) that help agents reason about safety and retry semantics.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2025-06-18+ | v2 |
Scout a repository to find relevant code implementing a feature or fixing a bug. Returns citations to specific file ranges with line-by-line evidence.
Error handling guidance is missing. No documentation of failure modes, recovery steps, or how agents should react to errors (e.g., 'file not found', 'invalid regex', 'timeout'). Errors likely return raw messages with no actionable recovery.
Parameter descriptions are often minimal and lack constraints. E.g., 'offset' in Symbols has no guidance on valid range. 'cmd' in exec_command has no list of allowed commands.
Pagination documentation unclear. Glob mentions 100 file limit; Symbols mentions next_offset pagination. But do all paginated tools return consistent structures? What's the total count? What's the next_cursor format?
Tool descriptions are under typical baseline. E.g., Grep (38 chars), exec_command (50 chars), Glob (50 chars) vs rubric baseline average 194 chars. Short descriptions lack context for LLM tool selection.