JobScout MCP

JobScout MCP

A privacy-first, bring-your-own-connections MCP server for multi-source job discovery. JobScout gives AI agents one normalized pool with provenance, deterministic deduplication, and specialist AI/Web3 signals.

JobScout deliberately stops at trustworthy discovery. It does not store CVs, rank candidates, contact employers, or apply to jobs.

Why JobScout

• Search enabled providers without making one source the whole market.
• Keep discovery provenance separate from canonical employer application routes.
• Continue with partial results when an individual provider fails.
• Enforce remote and freshness filters after provider retrieval.
• Disclose which constraints a provider could not apply, rather than returning an unscoped result as though it answered the question.
• Classify AI, agentic and Web3 signals locally and deterministically.
• Keep candidate identity and career policy in a private client such as Career OS.

MCP tools

Employer ATS sources have their own tool, so a deployment that allows or denies tools by name can offer them separately (for example as a Pro-only tool) without touching the free search. The free search does not return ATS-source results. jobscout_list_sources reports each source's search_tool, and a free search that names an employer source lists it in sources_elsewhere rather than searching it. To combine both pools, pass them to jobscout_deduplicate.

Two MCP prompts guide first-run use without the server storing anything: jobscout_setup walks through enabling providers and what each one contacts; jobscout_find_jobs gathers role, location and remote preference per search. A search run with zero enabled providers returns setup_required: true with guidance instead of a misleading empty result, and results report providers_disabled, records_rejected (with a records_rejected_by_provider breakdown), undated_records, location_unfiltered and warnings so thin results are always explained. require_dated and include_descriptions let a caller trade recall for certainty, or drop the untrusted description text it does not need.

Quick start

Node.js 22.13 or newer is required.

npm exec --yes --package=@sarutobi-sasuke/jobscout-mcp -- jobscout-mcp

Codex example with the public Himalayas adapter enabled:

codex mcp add jobscout --env JOBSCOUT_ENABLE_HIMALAYAS=true -- npm exec --yes --package=@sarutobi-sasuke/jobscout-mcp -- jobscout-mcp

MCP hosts can also run the package straight from GitHub, which is useful for pinning to a commit or testing an unreleased branch:

npm exec --yes --package=github:SarutobiSasuke8/jobscout-mcp -- jobscout-mcp

See installation for Claude Desktop, Cursor, local development, JobSpy, and troubleshooting.

Install in ChatGPT, Codex and Claude

JobScout runs two ways: locally over stdio (works today), or as a hosted Streamable HTTP server. The hosted service is planned for https://jobscout.mcprack.dev with a ready deploy recipe in deploy/, but it is not deployed yet.

Local, works today

• Codex: codex mcp add jobscout --env JOBSCOUT_ENABLE_HIMALAYAS=true -- npm exec --yes --package=@sarutobi-sasuke/jobscout-mcp -- jobscout-mcp
• Claude Code: claude mcp add jobscout --env JOBSCOUT_ENABLE_HIMALAYAS=true -- npm exec --yes --package=@sarutobi-sasuke/jobscout-mcp -- jobscout-mcp
• Claude Desktop, Cursor and other JSON hosts: see installation.

ChatGPT and Codex plugin (JobScout Discover)

The directory package is in plugins/jobscout-discover and follows OpenAI's plugin format. Its manifest and mcp.json point at https://jobscout.mcprack.dev. It is not yet listed: it needs that deployment to go live, OAuth for existing-account sign-in, a demo account and a portal submission. See JobScout Discover for the review materials and the full list of gaps, and deployment for the hosting recipe and operator runbook.

Once the deployment is live, a ChatGPT developer-mode connector or a Claude custom connector can point at https://jobscout.mcprack.dev/mcp, and Claude Code can use claude mcp add --transport http jobscout https://jobscout.mcprack.dev/mcp. Until then, use the local install above.

Plans are sold only on the operator's own site. There is no in-chat checkout. See monetisation.

Providers

Providers are disabled by default and failures are isolated.

The provider contract supports future official ATS and specialist job-board adapters without coupling the core to any one vendor. See provider documentation.

AI and Web3 intelligence

Every normalized job can include deterministic signals covering AI agents, inference, evals, safety, Web3 protocols, DeFi, DePIN, wallets, exchanges, developer infrastructure, gaming, and agentic commerce. Technology detection currently includes MCP, A2A, x402, LLMs, RAG, Ethereum, Base, and Solana.

These are inspectable discovery signals, not opaque candidate scores. See job intelligence.

Optional agent operating guide

agents/jobscout-operator.md gives an MCP-capable agent a safe, reusable discovery workflow and output contract. It is intentionally one functional operator—not a bundled fictional team—and contains no private candidate profile or Career OS policy.

Trust and safety

• Provider responses are untrusted and schema-validated.
• Only HTTP(S) job URLs are accepted.
• Remote and subprocess outputs have explicit size and time limits.
• JobSpy is spawned without a shell and receives typed JSON over stdin.
• No auto-apply, login automation, CAPTCHA bypass, credential capture, or proxy evasion.
• Users remain responsible for provider terms, job freshness, location eligibility, and employer-route verification.

Read SECURITY.md before enabling third-party providers.

Development

npm install
cp .env.example .env
npm run check
npm run smoke:mcp
npm run check:plugin

The protocol smoke test uses the official MCP Inspector. Provider contributions must include fixtures, failure behavior, provenance handling, and tests; see CONTRIBUTING.md.

Verify

npm ci
npm test          # build, then every test under test/, offline
npm run check     # the full gate CI runs: lockfile, typecheck, lint, boundary, plugin, tests

The tests never touch the network. Each source adapter (Greenhouse, Ashby, Lever, RemoteOK, We Work Remotely, Lenny's Job Board, Himalayas, and the JobSpy bridge output) is exercised against a recorded-shape fixture in test/fixtures/sources/, made of invented companies and jobs, and every normalised field is checked, including provenance: source, discovery URL, source job id and captured_at. test/dedupe-edge-cases.test.ts covers the same job arriving from several sources, spelling and URL variations, re-posts and how provenance merges. npm test preloads test/network-guard.ts, which fails any test that tries a real network call; only loopback is allowed. CI runs npm run check on Ubuntu and Windows for every pull request.

Release status

The next release includes more conservative duplicate handling and streaming feed-size limits. See reliability evidence and remaining gates. These changes are unreleased until a new version is tagged and published.

Live on npm as @sarutobi-sasuke/jobscout-mcp and on the official MCP Registry as io.github.SarutobiSasuke8/jobscout-mcp. Releases are tagged vX.Y.Z on GitHub; CI publishes to npm with provenance. See RELEASING.md for the release process and rollback order.

Prior art and licence

JobSpy is used as an optional MIT-licensed dependency rather than copied. borgius/jobspy-mcp-server demonstrated demand for a JobSpy MCP wrapper; no source from it is copied here.

JobScout MCP is Apache-2.0 licensed. Third-party dependencies retain their own licences.