I open-sourced an MCP server for job discovery. It does one job well, and it knows nothing about you. Both of those are deliberate.
JobScout MCP plugs into Claude, Codex, Cursor, Hermes, Openclaw, or any other MCP-capable agent and gives it one clean pool of job listings. Multiple sources are pulled into a single format, duplicates are merged the same way every time, and every listing carries a record of where it was found. It also flags AI and Web3 signals on each listing, so an agent can tell a genuine protocol role from a nursing advert that happens to mention "base salary" and "security protocol". That distinction sounds trivial. It took real work.
The boundary is the interesting part. JobScout holds no CV. It ranks no candidates. It contacts no employers and it cannot apply to anything. Every tool is read-only. To be clear, it still acts on your preferences. Your agent tells it what to search for, where, and how fresh, on every call. What it never does is remember you. The preferences arrive with the request and leave with the response.
In a category racing towards auto-apply bots that spray 500 applications a week, I think discovery and judgment should be separate pieces of infrastructure. The public piece, the one everyone shares, should know nothing about you. Your own agent holds your CV, your preferences, and your standards, and it makes the calls. That split is enforced by a build gate in CI, not by a promise in the README.
It already works in practice. I run it through my own Hermes agent, which scans the feed against my criteria and briefs me on anything worth a look. If your agent runs a morning briefing, a job scan slots straight in as one more section.
Why build it: a lot of good people in Web3 are between roles right now, and most of them already use AI agents daily. The job hunt itself is still scattered across boards, aggregators and Telegram groups. JobScout gives your agent one honest feed to work from, including remote-first sources, plus optional connections to the wider board ecosystem that you configure yourself.
It is an early v0.2. Next up: official ATS adapters (Greenhouse, Ashby, Lever) rather than more scrapers, and published accuracy numbers for the classifier.
Install:
npx -y @sarutobi-sasuke/jobscout-mcpRepo:
github.com/SarutobiSasuke8/jobscout-mcpApache-2.0. If you connect it to something interesting, tell me what broke.