Loading ToolBench
MCP server for Claude Architect that provides tools for delegating work, managing git operations, validating specs, and controlling autopilot workflows
14 tools with basic schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. Git read tools (1-4) have minimal descriptions (20-60 chars, below 50-char baseline). Delegation/autopilot tools (5-13) lack parameter descriptions for complex objects like 'spec' and 'responseMode'. No tool documents output schema structure. Error handling descriptions are absent. Tool names are clear (verb_noun pattern), but parameter naming inconsistencies (e.g., 'spec' vs 'delegationSpec') and missing type hints on object parameters reduce usability. No evidence of idempotency hints, permission gates, or recovery guidance in error cases.
Resume an autopilot workflow
Start an autopilot workflow for a delegation spec
Get the status of an autopilot workflow
Record a human decision on a candidate from a delegation run
Delegate work to the Claude Architect runtime by providing a checkout path and delegation spec
Delegate a pipeline execution to the Claude Architect runtime
Run diagnostic checks on the Claude Architect runtime environment
Get the list of changed files in a repository checkout against HEAD
Git read tools (git-status, git-diff, git-log, git-changed-files) have descriptions under 60 characters, below the 50-char minimum for LLM clarity. Descriptions lack context on when to use each tool and what output structure to expect.
Complex object parameters ('spec', 'delegationSpec') lack descriptions explaining their structure, required fields, or validation rules. LLMs cannot infer what to pass without explicit guidance.
No tool documents its output schema. LLMs cannot plan downstream calls or extract required fields (e.g., runId, workflowId) without knowing response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2025-06-18+ | v2 |
Get the git diff of a repository checkout against HEAD
Get the git log of a repository checkout (last 20 commits)
Get the git status of a repository checkout
Integrate a candidate from a delegation run into the repository
Review a candidate result from a delegation run
Validate a delegation spec against the schema and return its SHA256 hash
No error handling guidance. Tools return errors but do not explain recovery steps (e.g., 'spec validation failed, call validate-delegation-spec first'). Agents cannot self-correct.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (READ_ONLY vs WRITE). LLMs cannot infer safety properties without explicit hints.