Anjin

CI License: MIT

MCP server for anjin.tech, a Japan tech-jobs aggregator that pulls from Findy, Green, HERP and Levtech. Roughly 7,700 listings with parsed salaries across ~3,500 companies, machine-translated to English. Gives an agent a way to search the Japanese job market, pull up a listing, and match a profile against open roles.

Things you can ask for:

"Find me remote Rust jobs in Japan paying over ¥10M"

"What's the salary distribution for backend roles?"

"Which companies are hiring TypeScript engineers in Osaka?"

Tools

Install

Requires Node.js 20+ and git.

npm install -g github:egegungordu/anjin-mcp

This puts an anjin-mcp command on your PATH. Then register it with your client:

Claude Code

claude mcp add anjin -- anjin-mcp

Codex

codex mcp add anjin -- anjin-mcp

opencode, in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "anjin": { "type": "local", "command": ["anjin-mcp"], "enabled": true }
  }
}

VS Code

code --add-mcp "{\"name\":\"anjin\",\"command\":\"anjin-mcp\"}"

{
  "mcpServers": {
    "anjin": { "command": "anjin-mcp" }
  }
}

git clone https://github.com/egegungordu/anjin-mcp.git
cd anjin-mcp && npm install     # builds via the prepare script
claude mcp add anjin -- node /absolute/path/to/anjin-mcp/dist/index.js

No API key or account needed. anjin.tech is public.

How it works

anjin.tech has no API, no sitemap and no feeds, so this server fetches the site's Next.js pages and reads the JSON that's already sitting inside them: the RSC flight payload on list pages, streamed Suspense HTML on detail pages. Parsing the streaming protocol instead of CSS selectors means a redesign of the site won't break anything here.

Requests go out one at a time with a 400 ms gap between them. Pages are cached for 5 minutes, the company dataset for 12 hours. Everything times out at 20 s and retries once if the failure looks transient.

The tool surface follows Anthropic's tool-design guidance. Six tools shaped around a whole task each, instead of one tool per endpoint. Responses are short by default, with a detailed flag when you want the rest. Ids are readable. Enums and defaults cover the common case so the model only has to spell out what it actually cares about, and romaji prefecture names map to Japanese on the way in. Errors say what to try next instead of printing a stack trace. Scraped listing text is labeled as data, not instructions, so a job post can't steer the model.

Known limits, all of them upstream. There's no free-text keyword search over jobs, only the filter vocabulary. Wantedly listings can't be browsed. About 68% of the headline job count has detail data that parses. The company index is mostly stub records (name, HQ, a few metrics, categories on maybe 7%), so a question like "which companies use Go" has to go through search_jobs instead. Report figures come from the site's daily snapshot, so they lag live search counts.

Development

npm install          # installs deps and builds
npm test             # offline parser tests against captured fixtures
npm run test:handshake   # offline MCP protocol handshake + tool surface
npm run test:smoke       # live end-to-end suite (hits anjin.tech)
npm run test:protocol    # live MCP round-trip

src/index.ts          server entry: tool registration (stdio transport)
src/lib/http.ts       fetch layer: serialization, rate gap, TTL cache, retry
src/lib/flight.ts     RSC flight payload + React streamed-HTML resolution
src/lib/jobs.ts       job search params/parsing + job detail parsing
src/lib/companies.ts  company dataset cache, search, resolution, detail
src/lib/report.ts     market report section extraction
src/lib/match.ts      multi-search candidate matching + scoring
src/lib/format.ts     concise/detailed renderers shared by tools

License

MIT. This is an unofficial client for a public site and isn't affiliated with anjin.tech.