MCP server for Claude — audit open-source dependencies for known vulnerabilities
OSV MCP has well-structured tool definitions with clear naming (verb_noun pattern) and comprehensive descriptions (avg 180 chars). All 6 tools have input schemas with typed parameters. However, parameter descriptions lack depth, many omit constraints, ranges, or format guidance. Output schemas are not documented. Error handling is minimal (only check_batch and scan_dependencies return error objects). No tool annotations (readOnlyHint, idempotentHint). Composition is strong: tools chain well (list_ecosystems → check_package), batch variants exist, and results include necessary IDs. Security is solid: no secrets in params, all tools are read-only. Naming is consistent and action-oriented.
Check multiple packages for vulnerabilities in one call. More efficient than calling check_package repeatedly — uses the OSV batch API internally. Args: packages: List of packages, each with "name", "ecosystem", "version" keys. Max 100.
Check a single package for known vulnerabilities. Args: name: Package name (e.g., "requests", "lodash"). ecosystem: Package ecosystem (e.g., "PyPI", "npm"). Use list_ecosystems to see options. version: Exact package version (e.g., "2.28.0").
Get a quick severity breakdown for a list of packages. Returns only the summary statistics (no per-package vulnerability details). Faster to read than check_batch for large dependency lists. Args: packages: List of packages, each with "name", "ecosystem", "version" keys.
Get full details of a specific vulnerability by its ID. Supports OSV, CVE, and GHSA identifiers (e.g., "CVE-2024-3094", "GHSA-xxxx-xxxx-xxxx", "OSV-2024-1234"). Args: vuln_id: The vulnerability identifier.
List all supported package ecosystems with example package names. Use this to check which ecosystem string to pass to other tools.
Parameter descriptions lack format constraints and ranges. 'version' param in check_package has no guidance on format (e.g., semver, exact match). 'format' param in scan_dependencies lists options in description but not as enum constraint. LLMs cannot parse free-text constraints reliably.
Output schemas are not documented. Tools return dicts (model_dump()) but LLMs cannot see the structure of PackageResult, Vulnerability, or SeveritySummary objects. Downstream tool selection and field extraction are guesswork.
No tool annotations (readOnlyHint, idempotentHint). All tools are read-only and safe to retry, but this is not declared in the schema. Agents cannot infer safety properties without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Parse a dependency file and scan all packages for vulnerabilities. Paste the contents of your dependency file and specify the format. Supported formats: requirements.txt, package.json, go.mod, Cargo.toml. Args: content: Raw text content of the dependency file. format: File format — one of "requirements.txt", "package.json", "go.mod", "Cargo.toml".
Error handling is minimal. Only check_batch and scan_dependencies return error objects; others may fail silently or raise exceptions. No recovery guidance (e.g., 'Try list_ecosystems() to see valid ecosystem names'). Errors do not categorize as retryable vs. user-fixable.
Parameter 'format' in scan_dependencies lists valid values in description ('requirements.txt', 'package.json', 'go.mod', 'Cargo.toml') but not as an enum in the schema. LLMs may hallucinate unsupported formats.