MCP server that runs the Trustabl AI-agent safety & reliability scanner
Two tools with clear, detailed descriptions and well-structured Zod schemas. Tool names follow verb_noun pattern (trustabl_scan, trustabl_last_findings). Descriptions are 150 - 200 chars, within the 10 - 1024 baseline. All parameters have type definitions and descriptions. However, output schemas are not formally documented, the code parses JSON and returns structured data, but the response format is not declared in the tool definition. Error handling is present but generic in places. No tool annotations (readOnlyHint, destructiveHint) despite both tools being read-only operations.
Return the SARIF report written by the most recent trustabl_scan in a directory, without re-running the scan.
Run the Trustabl static scanner over a directory of AI-agent code. Detects agents, tools, and MCP servers across Claude Agent SDK, OpenAI Agents, Google ADK, LangChain, CrewAI, Pydantic AI and more, then reports risky patterns with severities and a production-readiness score. Writes trustabl.sarif (SARIF 2.1.0) and trustabl.json into the scanned directory, and returns the findings.
Output schemas not formally documented. Code returns structured objects (summary dict with readiness_score, risk_score, findings_total, by_severity, etc.) but the tool definition does not declare the response schema. LLMs cannot plan downstream operations without knowing what fields to expect.
Tool annotations missing. Both tools are read-only (risk: READ_ONLY declared in metadata) but the Zod schema does not include readOnlyHint or destructiveHint annotations. This prevents clients from optimizing caching or safety policies.
Error messages lack recovery guidance. When trustabl binary fails to install or scan fails, errors return raw messages ('Could not install the trustabl binary: ...') without suggesting next steps. Per pattern:recovery-guide, errors should tell the LLM what to do.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Result capping not documented. Code caps returned findings at MAX_FINDINGS_RETURNED=50 and report excerpt at REPORT_EXCERPT_LINES=40, but the tool description does not mention these limits. LLMs may expect all findings and be surprised by truncation.
Parameter 'detectors' lacks enum constraint. Accepts 'comma-separated SDK subset, e.g. claude_sdk,openai_sdk,mcp' but does not declare valid values as an enum. LLMs may hallucinate invalid detector names.