Open Roles

A free, remote MCP server for live job postings. It reads the public job-board APIs of Greenhouse, Lever, Ashby and Workable, the applicant-tracking systems behind thousands of company careers pages, so an AI assistant can find a company's board, browse and filter its openings, and read full job descriptions with pay ranges and apply links.

• No accounts, no API keys, no auth.

• Read-only. Every tool is annotated readOnlyHint.

• No personal data collected or logged (see Privacy).

• Runs on Cloudflare Workers: TypeScript, Cloudflare's agents McpAgent and @modelcontextprotocol/sdk, streamable HTTP at /mcp.

Connect

Server URL (replace <subdomain> with your workers.dev subdomain once deployed):

https://open-roles-mcp.<subdomain>.workers.dev/mcp

• Claude (web / desktop): Settings → Connectors → Add custom connector → paste the URL.

• Claude Code: claude mcp add --transport http open-roles https://open-roles-mcp.<subdomain>.workers.dev/mcp

• Other clients that support remote MCP servers over streamable HTTP: { "mcpServers": { "open-roles": { "url": "https://open-roles-mcp.<subdomain>.workers.dev/mcp" } } }. Clients that only speak stdio can use a bridge such as npx mcp-remote <url>.

• MCP Inspector: npx @modelcontextprotocol/inspector, transport Streamable HTTP, URL as above.

Try asking: "What remote engineering roles is Anthropic hiring for?", "Compare open ML roles at OpenAI, Anthropic and Hugging Face in Europe", "What did Stripe and Linear post this week?"

Related MCP server: io.github.groundtruthtools/ats-jobs-mcp

Tools

Boards are identified as provider:slug, e.g. greenhouse:anthropic, ashby:openai, lever:palantir, workable:huggingface. Anywhere a board is expected you can also pass a careers-page URL (https://jobs.ashbyhq.com/openai), a domain or a company name. It is then auto-detected.

Filter semantics (all case-insensitive and accent-insensitive):

• keyword matches the job title. Every word must match the start of a title word (engineer matches "Engineering", while ai does not match "Air"). Use | or OR for alternatives: frontend | front-end.

• department matches department or team names (engineering → "AI Research & Engineering").

• location matches posted locations, offices and countries. It understands common aliases: US/USA, UK, NYC, SF/Bay Area, LA, DC, Europe/EMEA, APAC, LATAM, US state codes (Austin, TX also matches "Austin, Texas") and remote.

• remote: true keeps roles marked remote or remote-friendly. remote: false excludes them. Hybrid counts as not remote. The provider's workplace field wins when it exists (Lever, Ashby, Workable). Otherwise location text ("Remote-Friendly, United States") and Greenhouse custom fields ("Location Type: Remote") are used.

Results are compact JSON with stable ids to pass to get_job.

Pay data

If there are no structured pay fields, get_job returns pay_mentions: lines from the description that state an amount. These are heuristic, so treat them as hints.

How it works

• Cache: upstream responses are normalised to compact job summaries (descriptions dropped) and cached in Workers KV (binding CACHE) for CACHE_TTL_SECONDS (default 600 s). Job details are cached the same way. Company → board detections are cached for 6 h (1 h when nothing was found). Caching is best-effort: KV errors or quota limits never fail a tool call.

• Request budget: each tool call makes at most MAX_UPSTREAM_REQUESTS (default 40) upstream requests, below the Workers Free limit of 50 subrequests per invocation. Probes for one slug run in parallel, identical requests within a call are shared, and detection stops at the first good match. If the budget runs out, the affected companies are reported as not finished rather than not found. Each upstream request times out after 25 s, and a whole tool call stops starting new requests after 45 s, so multi-company searches stay within typical client timeouts.

• Sessions: McpAgent keeps one Durable Object per MCP session. The tools themselves are stateless.

Run locally

npm install
npm run dev            # wrangler dev on http://localhost:8787 (local KV and Durable Objects, no login needed)
npm run test:e2e       # in another terminal: end-to-end test, writes test/e2e-output.txt
npm run typecheck

test/e2e.mjs uses the MCP SDK Client with StreamableHTTPClientTransport. It lists the tools, runs find_job_board("Anthropic"), runs list_jobs (keyword "engineer", department "engineering", remote), and calls get_job on the first result. It then runs search_jobs, whats_new and the error paths, and does a smoke test of Lever, Ashby and Workable boards. Pass another URL as the first argument to test a deployment.

Deploy

npx wrangler login
npx wrangler kv namespace create CACHE   # copy the returned id ...
# ... into wrangler.jsonc → kv_namespaces[0].id (replacing REPLACE_WITH_KV_NAMESPACE_ID)
npm run deploy                            # → https://open-roles-mcp.<subdomain>.workers.dev

The Workers Free plan is enough. Optional settings live in wrangler.jsonc → vars: CACHE_TTL_SECONDS and MAX_UPSTREAM_REQUESTS (raise it on a paid plan, where the subrequest limit is much higher). After changing bindings or vars, run npm run cf-typegen.

Privacy

• No authentication, accounts or cookies. The server never asks who you are.

• No request logging: Workers observability is disabled in wrangler.jsonc, and the code logs nothing.

• KV holds only public job-board data and company-name → board lookups, all with short expirations.

• The per-session Durable Object holds MCP protocol state (e.g. the client's name/version from initialize) for the session's lifetime. It holds no user data.

• Upstream requests go to the four job-board APIs with a generic User-Agent. Nothing from the client is forwarded apart from the company or board being looked up.

Caveats

• Board detection is a heuristic. Boards are found by guessing slugs, so expect misses and the occasional wrong match. Check confidence and careers_url.

  • Greenhouse: the most reliable. The API returns the company name, which is checked against your query, and slugs usually match the brand. Some companies use a different token (e.g. scaleai).

  • Ashby and Lever: the APIs expose no company name, so a match is only "medium" confidence and slug collisions can't be detected. Stale empty boards are common after a company switches ATS (e.g. lever:mistral has 0 jobs).

  • Workable: the least reliable. Many famous names (openai, stripe, figma, linear, …) are registered by unrelated or squatting accounts, mostly with no jobs, and the account name is whatever the owner typed. Empty boards are ranked last and hidden when a better board exists. If a company isn't on any of the four systems, though, a look-alike Workable board with postings could be returned.

  • Companies on other ATSs (Workday, SmartRecruiters, iCIMS, Personio, in-house, …) aren't covered. Neither are EU-hosted Lever (jobs.eu.lever.co) or Greenhouse EU boards.

• Timestamps differ by provider. Greenhouse has first-published and updated times. Lever has only creation time. Ashby has only publish time. Workable has only a publish date. whats_new can therefore report "updated" jobs only for Greenhouse, and reposted jobs may appear new.

• Remote/hybrid labels are inferred. Classification depends on what each company fills in. The Workable list endpoint only says remote or not; get_job shows hybrid/on-site.

• Payloads can be large. Big boards return several MB because list endpoints include descriptions (OpenAI on Ashby is ~14 MB, Anthropic on Greenhouse ~10 MB). The first request for such a board takes a second or two and uses noticeable memory and CPU; later requests hit the cache. On the Free plan, KV allows 1,000 writes per day. Beyond that, results are simply not cached.

• Live data: postings change constantly and may close between list_jobs and applying. Always apply via the official apply_url. This project is not affiliated with Greenhouse, Lever, Ashby or Workable. Respect their terms and rate limits.

• McpAgent is marked deprecated (feature-frozen) in agents 0.30 in favour of createMcpHandler. It still works and is used here as specified, but a future major version may require migrating.

License

MIT © 2026 Shanto Mathew