dependency-fitness-mcp — npm dependency fitness verdicts for coding agents

Is this npm package safe to depend on — and if not, what do I move to?

An MCP server that gives a coding agent a cross-validated fitness verdict for an npm package before it writes or upgrades a dependency:

{
  "deprecated": true,
  "yanked": false,
  "malicious": false,
  "superseded_by": { "latest": "14.0.0", "majors_behind": 13, "breaking_boundary": true },
  "safe_migration_target": {
    "package": "uuid",
    "version": "14.0.0",
    "rationale": "Maintainer's deprecation notice recommends 'uuid' (verified present and not deprecated).",
    "confidence": "high"
  },
  "confidence": "high",
  "last_verified": "2026-06-05T18:10:31Z"
}

It reconciles four free, sanctioned sources — the npm registry, Google's deps.dev, OSV.dev, and GitHub — into one confidence-scored answer, and infers a safe migration target when a package is deprecated or superseded.

Why this exists (and what it deliberately isn't)

"Is it deprecated?" is already free — deps.dev serves that flag, and several free MCP servers already answer "what's the latest version?". This tool does the part nobody serves as data:

• Migration-target inference. When a package is deprecated, it parses the

maintainer's own deprecation notice for a named successor, then verifies that successor actually exists and isn't itself deprecated before recommending it.

• Cross-validation, not a guess. It reconciles deprecation across the npm

registry and deps.dev, catches "deceptive deprecation" (registry says active but the GitHub repo is archived), and flags disagreement with a confidence level instead of inventing an answer.

• It refuses to guess. If a package is deprecated but no successor can be

established, it says exactly that (low confidence) rather than recommending a plausible-but-wrong replacement. A wrong "use X instead" ships broken code.

• Anti-slopsquatting. A non-existent / hallucinated package name returns a

clear "not found" verdict (with an OSV malicious-record check), so an agent won't silently install a hallucinated dependency.

This is intentionally a narrow tool: the deprecation / yank / supersede / migration middle, where the free incumbents sit on either side but leave the seam open.

Tools

check_package_fitness

Single-package verdict. Input: package (e.g. request, @babel/core), optional version (exact, semver range, or dist-tag — omit for latest). Output: the full Verdict (structured) plus a human-readable summary.

audit_dependencies

Batch verdict for a CI / pre-merge gate. Input: packages (e.g. ["[email protected]", "request"]) and/or the raw contents of a package.json. Output: a per-package verdict array plus a summary (how many deprecated / malicious / vulnerable / behind). Capped at 50 packages per call.

Install / connect

Requires Node ≥ 18. Run via npx (no install) or install globally.

Claude Code:

claude mcp add dependency-fitness -- npx -y dependency-fitness-mcp

Claude Desktop / Cursor / any MCP client (mcp.json / claude_desktop_config.json):

{
  "mcpServers": {
    "dependency-fitness": { "command": "npx", "args": ["-y", "dependency-fitness-mcp"] }
  }
}

Optional env: GITHUB_TOKEN raises the GitHub rate limit (used only for the archived-repo cross-check); everything else needs no key.

Run locally / develop

npm install
npm run build        # tsc -> dist/
npm test             # vitest (offline, deterministic synthesis tests)
npm run smoke        # live: hits the real registries, prints verdicts
npm run dev          # run the server from source over stdio

How a verdict is built

        ┌─ npm registry ── per-version `deprecated` string, dist-tags, repo URL  (authoritative)
query ──┼─ deps.dev ────── isDeprecated / deprecatedReason / advisoryKeys        (corroborator)
        ├─ OSV.dev ─────── advisories + MAL-* malicious markers + "fixed in"      (corroborator)
        └─ GitHub ──────── archived flag + last-push recency                       (deceptive-deprecation check)
                    │
                    ▼
   cross-validate deprecation ─→ infer + verify migration target ─→ confidence + warnings ─→ Verdict

npm is the source of truth; the others corroborate. A corroborator being unreachable lowers confidence and adds a warning — it never fabricates a signal.

Status

v0.1 — thin, working, npm-only. This is a fast public validation of whether a narrow "agent-data endpoint via MCP directory" can find its users organically. Roadmap and the explicit kill criterion live in KILL_CRITERION.md. Next layers (documented, not yet built): PyPI, and de-facto-successor inference by mining what high-trust packages actually depend on now.

License

MIT © Christo Wilken / 9592 Solutions UG. Built in public.