Loading ToolBench
An MCP server that gives Claude live access to your GitHub workspace
Strong tool naming (all verb-first, clear actions). Descriptions are comprehensive (avg 180 chars, well within 10-1024 range) and explain WHAT/WHEN/WHY. Input schemas are complete with proper JSON Schema types, required fields, and constraints (enums, min/max). All 7 tools have descriptions and schemas. However, output schemas are NOT documented, LLMs cannot predict response structure for downstream chaining. No error handling guidance visible. No tool annotations (readOnlyHint, destructiveHint). All tools are read-only (low risk), but lack explicit marking. Parameter descriptions are solid (avg 65 chars), though some could be more prescriptive about format/range.
Return per-contributor commit statistics for a repository: total commits, total lines added / deleted, and the number of weeks with at least one commit. Results are sorted by total commits descending.
Get detailed metadata for a specific GitHub repository: description, language, stars, forks, open issues, topics, default branch, visibility, and creation / last-update timestamps.
Generate a 7-day activity digest for a repository. Returns: merged PRs, newly opened issues, closed issue count, and the top 5 contributors by commit count for the period. Great for weekly stand-ups or team summaries.
List issues for a repository. Supports filtering by state (open / closed / all). Returns issue number, title, labels, assignees, author, comment count, and a truncated body.
List repositories for a GitHub organisation or the authenticated user. Returns repo names, descriptions, primary language, star counts, and open issue counts sorted by most recently pushed.
Output schemas not documented. LLMs cannot predict response structure (field names, types, nesting) for downstream tool chaining or data extraction. Forces agents to guess or make exploratory calls.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are read-only and safe to retry, but this is not explicitly declared. Agents cannot optimize retry logic or warn users about side effects.
No error handling guidance. Tools do not return actionable recovery hints (e.g., 'Repository not found. Try search_code() to verify the repo exists'). Raw API errors will confuse agents.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Search code across GitHub using GitHub's code-search syntax. Optionally restrict the search to a specific repository. Returns file name, path, repository, and a link to each result.
Return a comprehensive summary of a pull request: title, description, state, author, base/head branches, list of changed files with line deltas, all conversation comments, and aggregate addition / deletion counts.
Pagination not fully specified. list_repos and list_issues accept 'limit' but no 'offset' or 'cursor' parameter. No indication of total count or next_cursor in response. Large result sets will be truncated silently.
Parameter descriptions lack format/constraint details. E.g., 'repo' parameter does not specify if it accepts 'owner/repo' or just 'repo'. 'query' in search_code references 'GitHub code-search syntax' but does not explain valid operators or examples.