Unified Jobs MCP Server
# Unified Jobs MCP Server
A zero-dependency [Model Context Protocol](https://modelcontextprotocol.io) server that gives an AI
agent one thing no single-source jobs tool does: a **cross-source, deduplicated jobs feed**. It
queries applicant-tracking systems, the LinkedIn public guest search, and the Hacker News
_"Who is hiring?"_ thread, then collapses the **same role seen on several sources into one canonical
record** — with a `sources[]` array and a union of the richest salary, longest description, and
earliest post date. The agent gets a clean list instead of three overlapping ones plus a dedup
problem of its own.
- **No API key, no account, no login.** All data comes from keyless public endpoints.
- **Zero dependencies.** Hand-rolled spec-conformant JSON-RPC 2.0 over stdio — no SDK, no transitive
package tree in your agent's trust boundary. Installs and boots instantly under `npx`.
- **Public postings only.** No member data, no PII beyond public job fields.
## Install
Claude Desktop (`claude_desktop_config.json`) / Cursor / Cline / Continue:
```json
{
"mcpServers": {
"unified-jobs": {
"command": "npx",
"args": ["-y", "github:rezearcher/unified-jobs-mcp"]
}
}
}
```
Or run it directly:
```bash
npx -y github:rezearcher/unified-jobs-mcp
```
## Tools
### `search_unified_jobs`
The cross-source feed. Pull one or more companies' ATS boards (`companies[]`), search LinkedIn
(`keywords` + `location`), and/or the current HN thread — in a single call — and get back **one
deduplicated list**. The same role on several sources merges into one record:
```jsonc
{
"title": "Backend Software Engineer - Infrastructure",
"company": "Palantir",
"location": "United States",
"salary": "$135,000 - $200,000 / yr",
"posted_at": "2026-10-14T00:00:00.000Z",
"sources": [
{ "source_platform": "lever", "url": "https://jobs.lever.co/palantir/..." },
{ "source_platform": "linkedin", "url": "https://www.linkedin.com/jobs/view/..." }
],
"primary_url": "https://jobs.lever.co/palantir/...",
"merged": true
}
```
Every result set reports dedup stats: `raw_source_hits`, `canonical_jobs`, `cross_source_merges`,
`collapsed`, `per_source_hits`.
| field | notes |
|-------|-------|
| `companies` | ATS companies: bare slug (`"stripe"`), domain (`"stripe.com"`), or `"platform:token"` (`"lever:palantir"`, `"workday:tenant.wd5:Site"`). ATS + token auto-detected for slugs/domains. |
| `keywords`, `location` | LinkedIn search (and `keywords` also filters HN). |
| `sources` | subset of `["ats","linkedin","hn"]`. Default: all three. |
| `maxPerSource` | cap rows per source before merge (default 25, max 200). |
| `maxResults` | cap canonical records after merge (optional). |
| `remoteOnly`, `dedupe`, `includeDescription` | booleans; `dedupe` defaults on. |
### `search_hn_hiring`
Just the current month's Hacker News _"Ask HN: Who is hiring?"_ thread, normalized into the same
schema. The monthly thread id is **auto-discovered** (it rotates on the 1st of every month) and each
free-text post header is parsed into structured fields. Optional `query` / `remoteOnly` filters.
### `list_sources`
The source families, how to address ATS companies, and exactly how the cross-source dedup behaves.
No network call.
### `get_premium_access`
The metered depth endpoints beyond the free reads above — a normalized cross-ATS **compensation**
layer and a per-company **hiring-intent signal** — served pay-per-call over [x402](https://x402.org)
(USDC on Base, no account/API key; the caller's wallet pays the 402 inline, a few tenths of a cent
per call). Returns the base URL, each endpoint's path + price + what it adds, and the payment
handshake. No network call, no payment. Free preview first:
`https://assetfactory-jobs-x402.sigrunner.workers.dev/jobs/preview?companies=stripe.com`.
## Sources
| family | coverage | keyless |
|--------|----------|---------|
| **ats** | Greenhouse, Lever, Ashby, SmartRecruiters, Workday, Recruitee, Workable | ✅ |
| **linkedin** | public guest job search (keywords + location) | ✅ (rate-limited on cloud IPs → degrades open) |
| **hn** | _"Who is hiring?"_ current-month thread | ✅ |
Every source is wrapped so a single source failing (e.g. LinkedIn's guest endpoint rate-limiting a
cloud IP) **never fails the whole call** — you get the sources that did return. The hosted metered
version adds Indeed and rotating residential proxies for volume.
## How the dedup works
- **Canonical key:** `company | title | city`, with legal suffixes stripped but seniority and role
qualifiers kept — so `Backend - Infrastructure` and `Backend - Defense` stay distinct.
- **Merge:** scalar fields are first-non-null in source-rank order; salary is taken whole from the
single richest source (kept internally consistent); `posted_at` is the earliest; description is the
longest; `sources[]` accumulates every origin.
- **Degrade open:** a row with no derivable key is emitted as its own record with `merged: false` —
never silently dropped.
## Maintenance
A GitHub Actions smoke workflow runs a real MCP client handshake plus a live cross-source pull on
every push and daily. The ATS and HN families gate it; LinkedIn is advisory (cloud-IP rate limits).
A genuine upstream shape change fails the roundtrip and opens a tracking issue automatically — the
feed is kept working, not left to rot.
## License
MIT © Grey Ridge Signals Group LLC
TDQS
Scored across 4 tools
search_unified_jobs and search_hn_hiring overlap substantially: the unified tool already covers the HN thread as a source, so the dedicated HN tool is largely a subset of the general one. The descriptions do explain the difference (HN-only, no company list, whole thread), which keeps it usable, but an agent must read carefully to avoid redundant calls. list_sources and get_premium_access are clearly distinct from the search tools and from each other.
All four names follow a clean verb_noun snake_case pattern: search_unified_jobs, search_hn_hiring, list_sources, get_premium_access. Verbs (search/list/get) accurately signal the operation type, and there are no style deviations.
Four tools is on the thin side but defensible for a focused read-only job aggregator: two search entry points plus two metadata helpers. Nothing feels gratuitous, though the surface is small enough that a single search endpoint with source filters could arguably replace two.
The surface covers discovery well (cross-source search, HN search, source documentation), but there is no way to fetch a single canonical job record by id, no pagination/cursor continuation tool, and no detail endpoint for a merged record. get_premium_access is an informational upsell rather than a functional capability, which leaves the surface feeling search-only with dead ends after results are returned.