OpenHire · 哨兵

OpenHire · 开聘

A job-search radar for your AI assistant — first-party listings, every posting's real age, and your résumé never touches our servers. 让 AI 助手替你盯岗的求职雷达 —— 一手职位、岗位在架时长打分，简历不经过我们的服务器。

Every number here is checkable. Open-duration report (per-employer medians, named, monthly) · numbers.json (the provenance of every figure we quote, regenerated each refresh) · raw report data. 我们引用的每个数字都能核：岗位在架时长月报 · numbers.json

Is your company in here, and you did not put it here? Your postings are in this index because your own careers page serves them publicly. We can measure how long a role has been open; we cannot see why, and a long-open role is a question, not a verdict. Claim your company — free, no payment, ever — and say why in your own words, or ask us to remove you and we will, without arguing. 贵司被收录了、而且不是贵司提交的？点这里认领，免费， 可以用自己的话解释，也可以直接要求我们移除。

三不原则 · Three things OpenHire never does

1. 不存简历 · No résumé, ever. 简历留在你的电脑上。过网的只有一个你自己生成的匿名指纹和岗位号；协议里没有简历字段，多传一个 resume 参数会被拒绝，CI 测试钉死。 Your résumé stays on your machine. Only an anonymous, client-generated fingerprint and a job id ever cross the wire; the protocol has no résumé field, an extra resume argument is refused, and CI pins it.
2. 不刷假单 · Never applies for you. 它只把你送到雇主自己的投递页，从不向任何 ATS 提交申请；连表单长什么样，我们的测试也只用一次普通 GET 看，到此为止。 It hands you the employer's own apply page and never submits anything to any ATS. Even our own tests only GET the form once and stop.
3. 不追踪 · No tracking. 不埋点、不统计谁看了什么。服务器只记一条匿名授权（指纹 + 岗位号），没有你是谁；索引本身跑在你自己的机器上。 No analytics, no view counting, no telemetry. The server keeps one anonymous authorization record (fingerprint + job id) and nothing about who you are; the index itself runs on your machine.

还有一条不在这三条里，但同样锁死：排序买不到，它是（匹配度，新鲜度）的纯函数，见下文三条隐私红线。 One more that is locked the same way: ranking cannot be bought; it is a pure function of (match, freshness). See the three red lines below.

What your agent actually sees

You ask your assistant a question in plain language. It calls search_jobs, and every row comes back carrying the employer's real posting date — so the agent can reason about staleness instead of guessing.

You: Any senior Python roles that are actually still open? Skip the stale ones.

// one row from search_jobs — trimmed to the fields that matter here
{
  "title":        "Senior Python Engineer",
  "company":      "MongoDB",
  "datePosted":   "2026-03-31",   // from the employer's ATS, not a board's refreshed label
  "days_open":    166,
  "ghost_score":  0.61,           // pure f(relist_count, datePosted) — frozen by a test
  "apply_channel":"https://boards.greenhouse.io/…",   // straight to the employer
  "verified_at":  "2026-09-02T09:47:10Z"
}

Assistant: This one has been open 166 days with a ghost_score of 0.61 — I'd deprioritise it. Here are four posted in the last three weeks instead…

ghost_score measures how long a posting has been open, not whether the employer still intends to hire. A long-open role can equally mean "hard to fill". Treat it as a reason to ask, not a verdict.

An MCP server that turns any MCP-speaking assistant — Claude, Cursor, Windsurf, Cline, ChatGPT via connectors — into a private radar for AI / Infra, autonomous-driving and embodied-AI jobs — pulled straight from 143 employers' own career sites and public ATS APIs (Greenhouse / Lever / Ashby / 北森 Beisen / Moka, plus first-party employer career sites such as Li Auto's), across the US, Europe and China (Waymo, Figure, Zoox — and Unitree, XPeng, UBTECH, Mech-Mind…). No account. No signup. No résumé upload. Ever.

Three things a job board won't do for you:

• Surfaces how long each role has really been open. Every listing carries a ghost_score aged off the employer's real posting date where the employer's system reports one (otherwise the day this index first saw it, flagged date_signal): the "2 days ago" a board shows you can be 300 days old in the ATS.
• Structural privacy, not a pinky-promise. There is no résumé field in the protocol; a CI test fails the build if anyone adds one. Matching runs on your machine — only an anonymous fingerprint reaches the server.
• Ranking you can't buy. Order is a locked pure function of (match, freshness). No sponsored slots, no bidding — the signature is frozen by a test.

This is the 「哨兵 / Sentinel」 reference implementation — see design_handoff_openhire_v01/README.md for the full protocol spec.

Quickstart — under a minute

Nothing to install, and no crawl to sit through. Add this to your MCP client's config:

{ "mcpServers": { "openhire": { "command": "uvx", "args": ["openhire@latest", "serve"] } } }

Then ask your assistant for a job. That is the whole setup. The server downloads the ~30 MB public index by itself in the background on first start, so searches fill in within a couple of minutes while you are already talking to it. No account, no signup, no résumé upload.

The one prerequisite is uv (it provides uvx). Without it the config above fails with nothing but "server failed to start", so install it first:

curl -LsSf https://astral.sh/uv/install.sh | sh     # macOS / Linux
irm https://astral.sh/uv/install.ps1 | iex          # Windows PowerShell

On Claude Desktop you can skip even that — download openhire-0.6.8.mcpb and double-click it. No terminal, no Python, no uv.

On Cursor, one click (it still needs uv on your PATH):

Per-client config paths and the trade-offs of uvx vs a one-time install are in Works with below.

pipx install openhire      # keeps it isolated and puts `ohp` on your PATH
ohp bootstrap              # 143 employers · ~17k live postings · no account

ohp search --required-skills rust,k8s --remote --role-family engineering
ohp search --currency CNY --role-family engineering   # e.g. CN 智驾 / robotics roles

Works with

All clients use the same MCP entry. The config below works in every MCP client and pulls the package on demand — but it does need uv present first.

New to MCP? Two shortcuts before the config below. Claude Desktop: download openhire-0.6.8.mcpb and double-click it. No terminal, no Python. Cursor — (needs uv installed). Cursor / Claude Code — or paste this to your agent: "Install the MCP server at github.com/gzchenhao/openhire. Install uv first if it is missing, then add uvx openhire@latest serve to my MCP config and tell me which file you changed."

Prerequisite: uvx ships with uv. Without it the config below fails with nothing but "server failed to start" in your client — install uv first:

curl -LsSf https://astral.sh/uv/install.sh | sh     # macOS / Linux
irm https://astral.sh/uv/install.ps1 | iex          # Windows PowerShell

{ "mcpServers": { "openhire": { "command": "uvx", "args": ["openhire@latest", "serve"] } } }

First start downloads a ~30 MB index in the background; searches fill in within a few minutes.

Stuck? Run ohp doctor. It checks the three things that all look identical from the chat window — uv missing, no index yet, server configured but not enabled — and reads every client config it can find. It runs in your terminal, which matters: if the client never started our server, nothing we wrote inside it can reach you.

Editing the config may not be the last step. Several clients require you to enable or trust a newly added server before its tools load — the config is saved, the server never starts, and the only symptom is that your assistant does not seem to know about the tools. If a search does nothing, open your client's MCP/connectors panel and check that openhire is listed and switched on. (Claude Desktop needs a full quit and reopen; Cursor and Windsurf pick it up on reload; some clients show a per-server toggle.) 改完配置不一定就完事：部分客户端需要你在设置里手动「信任 / 启用」这个 server， 工具才会加载。症状是配置明明在、助手却完全不知道有这些工具。

What uvx costs you, every time. uvx resolves the package on each invocation — measured at 7–8 s per call even with a warm cache. That is paid on every MCP session start and every CLI command. It buys you never having to manage an install. If you would rather pay once:

pipx install openhire     # then use "command": "ohp", "args": ["serve"] — process-start latency

@latest also means your tool surface can change under you without warning. Pin it when that matters: "args": ["openhire==0.6.8", "serve"].

The server auto-downloads the public job snapshot on first run if the index is empty, so ohp bootstrap is optional. If you ran pipx install openhire, "command": "ohp" works too.

Claude Desktop — %APPDATA%\Claude\claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/); quit & reopen after editing:

{ "mcpServers": { "openhire": { "command": "ohp", "args": ["serve"] } } }

Cursor — ~/.cursor/mcp.json (or a project .cursor/mcp.json):

{ "mcpServers": { "openhire": { "command": "uvx", "args": ["openhire", "serve"] } } }

Windsurf — ~/.codeium/windsurf/mcp_config.json:

{ "mcpServers": { "openhire": { "command": "uvx", "args": ["openhire", "serve"] } } }

Claude Code — this repo is also a plugin marketplace, so two commands inside Claude Code do the whole setup (still needs uv on PATH):

/plugin marketplace add gzchenhao/openhire
/plugin install openhire@openhire

Cursor plugin / Agent Plugins — the repo carries .cursor-plugin/plugin.json and the vendor-neutral plugin.json + mcp.json (agent-plugins.org), so any client that loads Agent Plugins can install it from the repo URL.

TRAE（字节） — one-click import links (TRAE asks you to confirm the config before adding it): TraeCode 国内版 · TRAE international. Or paste the uvx config above into TRAE's MCP settings by hand.

腾讯 WorkBuddy — ~/.workbuddy/mcp.json, same mcpServers shape. WorkBuddy reads the file when it starts, so after editing it quit and reopen the app (or add the server through its MCP panel's 「添加 MCP」 instead); a hand-written entry does not appear in the list until then. No uv on the machine? pip install openhire into any Python ≥ 3.11 environment and point command at that environment's ohp executable with "args": ["serve"].

First start downloads the ~30 MB public snapshot (jobs/companies only) — give it a moment. To refresh later run ohp bootstrap --force or ohp ingest. On Windows Claude Desktop from the Microsoft Store, the config is under …\Packages\<Claude package>\LocalCache\Roaming\Claude\.

Hosted / remote: ohp serve --transport streamable-http --host 0.0.0.0 --port 8000 exposes http://host:8000/mcp (also --transport sse). A Dockerfile is included.

GitHub unreachable from your network (mainland China without a proxy, some sandboxes)? The snapshot is a GitHub Release asset and the download now resumes and retries, but a blocked host stays blocked. Fetch openhire-index.db.gz through a mirror you trust or on another machine, then ohp bootstrap --snapshot-url <url-or-local-path>, or set OPENHIRE_SNAPSHOT_URL in the server's env so the auto-download uses it. ohp bootstrap --fresh skips the snapshot entirely and crawls the employers' own ATS (20+ minutes, no GitHub involved). An empty index caused by a failed download says so in the tool result (bootstrap_error) instead of looking like "no matches". 国内网络 GitHub 不通：经你信任的镜像拿到快照后 ohp bootstrap --snapshot-url <地址或本地文件>， 或在 MCP 配置的 env 里设 OPENHIRE_SNAPSHOT_URL；ohp bootstrap --fresh 不经 GitHub 直接抓。

What it does

Optional, entirely local: ohp init --scan <dir> derives a skill fingerprint from your own repos. You never write a résumé; the code never leaves your machine — only an anonymous vector does.

For employers: claim your tenant

If your company is in this index, the listings came from your own public ATS — we did not ask, because we did not need to. What we cannot know is your side of it: whether a role is an evergreen talent pool rather than a stale req, or how fast you actually reply.

Claim it （中文表单）. Free, verified by corporate identity — a GitHub org membership or a reply from a corporate domain — and never by payment. We answer within 3 business days, claiming leads to no paid follow-up of any kind, and your proof is used to verify and then nothing else.

No GitHub account? Email gdchenhao@qq.com with "Employer claim" and your company name in the subject — sending from your corporate domain is itself the verification — or have anyone file the form on your behalf, since what we verify is the company and not the filer. 没有 GitHub 账号？直接发邮件到 gdchenhao@qq.com，用贵司企业邮箱发出来即完成身份核验。

A claim gets you:

• Your own note, attributed to you, beside the roles you name. We can see how long a role has been open; we cannot see why. An evergreen talent pool, a genuinely hard-to-fill role, and a neglected one look identical from outside, and only you can tell them apart.
• evergreen / hard to fill / closed status on specific titles. A closed role stays visible while your ATS still serves it — we do not hide what your own site returns — but it is marked closed on your word so nobody else applies.
• A correction if your ATS's date field means "requisition opened", not "went live". That one misreading makes every role you have look years old, and it is not your fault.
• response_sla_days on every one of your postings, including ones you post later
• claimed: true on your company, with the date

It does not get you rank, and it does not lower your ghost_score. Ordering is a locked pure function of (match, freshness) and the score is a pure function of (relist count, posting age); tests freeze both and assert the claim path touches neither. Your note sits beside the score and explains it. We would rather show a high score with your explanation than a quiet score somebody paid for.

Asking to be removed entirely is also fine, and we will not argue about it.

Verified claims live in src/openhire/seed/claims.py — in the repo, not in a private database — so each one is a reviewable diff, and the weekly rebuild re-applies them instead of quietly dropping them.

Keeping it current

The index refreshes weekly, so a search can be up to seven days behind. When the user is about to act on one employer and wants today's truth:

ohp refresh unitree          # ~1 minute · at most one crawl per employer per 6 hours

Over MCP this is the refresh_index tool. Three rules are built in, not advisory:

• One employer per call. A full crawl is 20+ minutes and no client waits that long. An ambiguous word (robot → 11 matches) is refused with the candidate list, never fanned out.
• Six-hour throttle per employer, checked before any network call, so a too-soon request costs the ATS nothing and returns last_refreshed_at instead of an error.
• Not for speculative or looped calls. Each one hits somebody else's public endpoint.

That last point is the whole design constraint: our own crawl is a weekly batch we control, and handing refresh to callers turns it into our users hitting their endpoint on our behalf. The throttle is what keeps that boundary ours to keep rather than ours to spend.

Reading the three dates

Every row carries three timestamps that answer three different questions. Read together they separate an abandoned requisition from one somebody is still tending:

ghost_score = 1.0 alone is not a verdict. Open 367 days and untouched for 367 days reads as abandoned; open 327 days but touched 13 days ago reads as a tended evergreen req. Among the rows where the ATS actually reports a last-touched date, 67% of ghost_score >= 0.99 postings were touched by the employer within the last 30 days (measured 2026-09-19; the live figure is pct_ghost_hi_touched_within_30d in docs/numbers.json, which is regenerated every refresh). ghost_reason spells out which input drove the score ("age only: open 367d, never relisted").

None of these measure intent. A long-open role can equally mean hard-to-fill — treat the numbers as a reason to ask, not a verdict.

Asking about one employer

search_jobs(company=...) takes whatever the user actually said — an id (unitree), or any part of the name in either language (宇树, Unitree, XPeng). A name this index does not carry comes back as the empty-result object naming the company, never as an unfiltered search.

ohp search --company 宇树 --role-family engineering
ohp search --company waymo --distinct          # one row per role, not one per city

--distinct (collapse_role_group over MCP) keeps one row per role_group and adds role_group_size. About 20% of a page is the same role listed once per city; folding is opt-in because each city row has its own job_id and apply_channel, which matters under a location or visa constraint.

The five protocol fields

Every listing is valid schema.org/JobPosting, plus:

• verified_at — last moment confirmed live on the employer's own site
• source — employer_site | ats_public_api (never a job board)
• ghost_score — 0–1 listing-activity signal, aged off the real posting date (lower = fresher). A noise filter, not an accusation: long-open listings are often evergreen talent pools or slow pipelines — the score simply lets agents down-rank low-activity noise
• response_sla_days — the employer's OWN committed reply window. Null on almost every row, and that null is meaningful: it is set only when an employer claims their tenant. We never infer or estimate it, because we cannot observe a reply even in principle — the application deep-links to the employer and never touches this server. Read null as "no employer has claimed this tenant", not as "missing" or "slow". Employers: claim yours — free, verified by corporate identity, never by payment. It buys a verified badge, the ability to correct listing status (an evergreen pool carrying an unfair staleness score, say), and this field. It does not buy rank: ranking is a locked pure function of (match, freshness), and a test freezes that signature.
• apply_channel — always the employer's own application URL, deep-linked to the specific job

Privacy Policy

Short version: there is no résumé field in the protocol, matching runs on your machine, and the only user-originated value the server ever stores is an anonymous client-generated fingerprint. No analytics, no telemetry, no third-party sharing. Full policy: docs/PRIVACY.md.

Privacy model

First-run data — the snapshot vs. fresh

ohp bootstrap (default) downloads a small public index snapshot (a GitHub Release asset — companies + jobs only, zero user data) and then runs one incremental crawl to refresh verified_at / delisting. --fresh skips the snapshot and crawls the public ATS from scratch with the free offline heuristic extractor. Either way: no account, no PII.

Two things that surprise people:

• The incremental crawl is slow and quiet. On a cold index it can run for 20+ minutes with no output. It is working, not hung. If you only want the data, ohp serve skips it entirely — the server downloads the snapshot on first start and is answering in seconds.
• The snapshot URL is pinned to the v0.1.0 tag on purpose. It looks stale; it is not. That asset is overwritten in place every Monday by a scheduled workflow, so the URL is a stable address for always-current data. Pinning it to the newest tag would break every client the moment a release is cut.

Three rules this project will never break

1. Your résumé stays on your machine — it never transits the server, and we never store it.
2. Ranking is not for sale — it is only f(match_quality, freshness), a locked pure function.
3. Employers pay only for authorized, delivered outcomes — never for exposure. (v0.1 has no billing at all.)

These are enforced by CI (tests/test_privacy.py, tests/test_ranking.py, tests/test_snapshot.py).

Development

python -m venv .venv && . .venv/Scripts/activate   # Windows
pip install -e ".[dev]"
pytest        # privacy red lines + ranking + snapshot must be green

Set OPENHIRE_DATABASE_URL=postgresql+psycopg://… to run against Postgres instead of the default local SQLite file (~/.openhire/openhire.db).

Roadmap

• v0.2 – v0.3 (shipped) — CN ATS adapters (北森 Beisen + Moka) · weekly auto-refreshed public snapshot · ghost_score public beta · 143 employers across US / EU / China
• next — Employer claim + verified badges — employers can reserve their claim today via a corporate-identity GitHub issue (zero-cost now; badges + listing-status control ship next) · response-SLA enforcement (7-day auto-delist) · redacted proof-of-fit — an anonymous, candidate-authorized match summary that travels with an application (skills overlap only; identity never included, résumés still never transit the server)
• v1.0 — Open, vendor-neutral schema extension for AI-readable job postings

FAQ

Where does the job data come from? Directly from 143 employers' own public ATS APIs (Greenhouse, Lever, Ashby, 北森 Beisen, Moka, plus first-party employer career sites such as Li Auto's), the same endpoints that power their careers pages. No third-party job boards. source is always ats_public_api, and verified_at records the last time we confirmed each posting live. The public index is auto-refreshed weekly, so a fresh ohp bootstrap starts from recent data.

Why should I trust ghost_score? It's a pure, open, unpurchasable function — min(1, 0.15·relist_count + staleness) aged off the real ATS posting date, not our crawl date. The formula lives in pipeline/ghost_score.py, is unit-tested, and takes no money as input (red line #2). Long-open, repeatedly-relisted postings score higher; you can always re-rank client-side. Read it as signal-to-noise, not bad faith: plenty of high-scoring listings are legitimate evergreen talent pools. Employers who want their listing activity represented accurately can claim their tenant (see Roadmap).

Does my résumé actually go through the server — really? No. There is no résumé anywhere in the protocol. authorize_application has no résumé/file parameter (it structurally cannot accept one), matching runs on your machine, and the only thing that ever transits the server is an anonymous fingerprint the client generates, like #a3f9-k2p7-x8q1. This is enforced by tests/test_privacy.py, and the published snapshot carries zero user data (tests/test_snapshot.py).

Does it support China (中国区)? Yes — this is what sets OpenHire apart. Employers on 北森 Beisen (<tenant>.zhiye.com) and Moka (app.mokahr.com) are indexed: 20+ autonomous-driving / robotics / embodied-AI companies including 宇树 Unitree, 小鹏 XPeng, 优必选 UBTECH, 梅卡曼德 Mech-Mind, 速腾聚创 RoboSense, 元戎启行 DeepRoute, 星海图 Galaxea, 傅利叶 Fourier, 普渡 Pudu. Pay published as 月薪 keeps its real period (salary_period), so a salary floor no longer silently drops Chinese roles.

飞书招聘 (Feishu Hire) is not supported and won't be: it signs its job-list requests with a ByteDance _signature and gates them behind a captcha SDK, so its listings are not publicly readable. We don't break anti-bot measures.

How do I get a company added? Open a Company inclusion request issue (title it with the company + its ATS URL) — this is the best way to contribute. If you code, add it to src/openhire/seed/candidates.py (company slug + ATS vendor/tenant) and open a PR; the seeder validates tenants against the live API.

License

MIT © OpenHire Protocol · PRs welcome.

Built by a deep-tech headhunter who does not write code, pair-programming with Claude Code. Full acceptance reports, including the mistakes, in reports/.