Loading ToolBench
MCP stdio server that serves local repo memory and org memory from GitHub PR history, code, tests, regressions, and cross-repo impact to AI coding agents.
Anchor MCP server has 17 tools with good naming (verb_noun pattern: anchor_get_*, anchor_search_*, anchor_check_*, anchor_plan_*, anchor_review_*) and clear, domain-specific descriptions (avg ~120 chars). However, critical gaps prevent higher scores: (1) Input schemas are NOT visible in the provided source code, only tool names and descriptions are shown. The server.ts file imports schema objects (AnchorGetContextSchema, etc.) but does not display their definitions, making schema validation impossible. (2) Parameter descriptions are inferred from imports but not verifiable. (3) Output schemas are undocumented, no evidence of structured return types, pagination, or field documentation. (4) Error handling patterns are absent from visible code. All 17 tools are READ_ONLY, which is good for safety, but without visible schemas and output documentation, the server cannot be confidently recommended for production use.
Review a diff against local architecture patterns and surface evidence-backed placement, import, and test guidance.
Use before API, auth/access, billing, schema, SDK, shared-package, or broad refactor changes. Surfaces deterministic cross-repo impact and anomalies.
Explain one file using local code evidence, PR history, team rules, regressions, and related tests before a larger AI coding edit.
Use when changing endpoint, schema, client, or SDK code to find allowlisted repos that consume matching API contracts.
Return deterministic local architecture patterns for a file, area, or query so AI coding agents follow existing repo structure.
Return a deterministic architecture graph from indexed imports, components, and test links.
Input schemas not visible in source code. All 17 tools import schema objects (AnchorGetContextSchema, etc.) but definitions are not provided. Cannot verify parameter types, descriptions, constraints, or validation rules.
Output schemas undocumented. No evidence of structured return types, field descriptions, pagination support, or chaining IDs. LLMs cannot plan downstream tool calls or extract required data without documented output structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Return concise, ranked, sanitized PR-history, team-rule, and codebase context before non-trivial AI coding edits. Use strict mode for non-stale high-confidence evidence only.
Use when adding repos, services, packages, or cross-repo integrations to inspect the local org architecture map.
Use before broad or cross-repo AI coding work. Returns sanitized, evidence-backed org context across allowlisted repos.
Return one committed repo playbook by id, with cited local evidence.
Infer exact local test commands for a source or test file using indexed test links and package scripts.
Return local Anchor index counts, history coverage, coverage score, team-rule count, stale evidence count, and health.
Return a concise onboarding brief with areas, important files, risk modules, tests, playbooks, and starter prompts.
Check local org memory freshness, coverage, cloned repos, cross-repo edge count, and API consumer count.
Create a deterministic edit plan with target files, likely symbols, risks, exact test commands, and cited Anchor evidence.
Review a diff against local Anchor history, team rules, regression memory, and likely tests. It surfaces evidence-backed risks only.
Manually search sanitized indexed GitHub PR history.
Error handling patterns absent. No evidence of recovery guidance, error classification (retryable vs fatal), or actionable error messages. Agents cannot self-correct or determine next steps on failure.
Parameter descriptions not verifiable. Schema imports exist but parameter-level documentation is not visible in source. Cannot confirm each parameter has a description explaining its purpose, format, and constraints.
No pagination or result-limiting evidence. Tools like anchor_search_history and anchor_get_org_context likely return large result sets but no visible limit, offset, or cursor parameters. Risk of context window exhaustion.