Loading ToolBench
MCP server for Verdoc - deterministic AGENTS.md and repository analysis for coding agents, paid per call in USDC over x402. No account, no API key.
Two tools with clear, domain-specific purposes and well-documented descriptions (194 - 250 chars, within baseline 34 - 392). Input schemas present with required fields and type definitions. However, parameter descriptions are minimal or absent, and output schemas are not documented. Tool names follow verb_noun convention but lack clarity on what distinguishes them (both are 'verdoc_*' prefixed analysis tools). Error handling is present but generic. No tool annotations (readOnlyHint, idempotentHint) despite both being read-only, deterministic operations.
Generate an AGENTS.md for a public GitHub repository: build and test commands quoted from the repository's own manifests with the source file named, a directory map, entry points and edit boundaries. Every path is checked against the git tree and anything ambiguous is marked unverified rather than guessed. Use this before working in a repository you do not already know. Costs $0.02 in USDC from the configured wallet.
Structured facts about a public GitHub repository: toolchain commands with the manifest each was quoted from, directory layout, language split, CI and test presence, a 0-100 documentation-health score, and phantom_paths - paths the README cites that do not exist in the git tree. Deterministic; no language model. Costs $0.01 in USDC.
Parameter descriptions missing or incomplete. 'repo' parameter has a description, but 'format' enum in verdoc_agents_md lacks explanation of what 'json' vs 'markdown' output contains or when to use each.
Output schemas not documented. Tool descriptions state what is returned (AGENTS.md, structured JSON) but do not specify the schema of returned fields, making it impossible for LLMs to plan downstream operations or extract specific data.
Tool annotations missing. Both tools are read-only and deterministic (idempotent), but lack readOnlyHint and idempotentHint annotations. This forces LLMs to infer safety properties rather than relying on explicit metadata.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Error handling is generic. callVerdoc() returns {error: true, text: '...'} but does not categorize errors as retryable, user-fixable, or fatal. LLMs cannot determine whether to retry, ask the user, or abandon the operation.
Tool naming lacks distinction. Both tools are prefixed 'verdoc_' and perform similar analysis (one returns markdown, one returns JSON). Names do not clearly signal the difference to LLMs, consider 'generate_agents_md' vs 'analyze_repository' or similar verb-based distinction.