Loading ToolBench
MCP security trust layer. Scan packages, inspect repo MCP dependencies, generate Policy Gate setup, install the workflow directly, check exposure, and query abuse data.
AgentScore MCP provides 8 tools with clear security-focused intent. Tool names follow verb_noun pattern (scan_package, get_verdict, check_exposure). Descriptions are present and substantive (100-200 chars typical), explaining WHAT and WHEN. However, input schemas are minimal, most tools accept only 1-2 string parameters with basic descriptions. Output schemas are not formally documented in the visible code; formatting is done via text templates in tools.ts. Error handling is present but generic (ApiError catch-all). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Parameter descriptions lack format constraints (e.g., npm package name format, path validation rules). Overall: solid naming and descriptions, but schemas and output documentation are underdeveloped.
Check if a package or agent has been reported to the KYA abuse database. Returns whether abuse has been reported and any details.
Check which monitored MCP servers depend on a given package. Use this during incident response to find blast radius. Example: 'which MCP servers use axios?'
Inspect the current repo for MCP dependencies, look up AgentScore verdicts for each package, and summarise what should be gated in CI. Use this when a developer wants to understand all MCP packages in a repo instead of scanning one package at a time.
Generate the exact GitHub Actions workflow needed to enforce AgentScore Policy Gate for a repo. Detects MCP dependencies locally and returns the OIDC-based YAML needed for setup. No API key or secret is required.
Get a trust verdict for an MCP package: allow, warn, or block. Based on scan findings (score and severity). Also reports monitoring status and publisher posture. Use this before installing or connecting to an MCP server.
Input schemas lack formal JSON Schema type definitions and constraints. All tools accept string parameters with minimal validation rules (e.g., npm package name format, path validation). No enum constraints, regex patterns, or min/max bounds visible.
Output schemas are not formally documented. Responses are formatted as plain text via template functions (formatScan, formatVerdict, etc.) rather than structured JSON objects with typed fields. LLMs cannot parse expected response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Write the AgentScore Policy Gate workflow file to this repo. Creates .github/workflows/agentscore-policy-gate.yml with OIDC authentication (no API key needed). Detects MCP dependencies and includes them in the workflow. The gate will auto-provision the repo on first push.
Check if an MCP package is under continuous monitoring and get its scan history. Shows current score, risk level, and recent changes.
Scan an npm package for MCP security issues. Checks install scripts, prompt injection patterns, suspicious URLs, source code patterns, dependency count, metadata completeness, and publisher provenance. Returns score (0-100), risk level, and detailed findings.
Tool annotations missing. install_policy_gate is destructive (WRITE risk) but lacks destructiveHint annotation. All READ_ONLY tools lack readOnlyHint. Agents cannot infer safety profile from tool metadata.
Parameter descriptions lack actionable constraints. 'npm package name' does not specify format (scoped vs unscoped, length limits, character restrictions). 'path' parameter lacks default behavior documentation. LLMs cannot validate inputs without explicit format rules.
Error handling is generic. handleApiError() returns plain text messages without categorization (retryable vs user-fixable vs fatal). No guidance on next steps (e.g., 'Try again in 60 seconds' for 429, or 'Check package name format' for validation errors).