Anthropic Tracker MCP
pete-builds/anthropic-tracker-mcp
Acesso híbrido a dados de hiring: cache + live API
Ver no GitHubServidor MCP que expõe dados de contratação da Anthropic para Claude Code e clientes MCP. Combina 7 ferramentas cached em SQLite (snapshots noturnos) com 3 ferramentas live que consultam a API pública do Greenhouse em tempo real para análise de tendências e posições em aberto.
Instalação
npx convez-ai pete-builds/anthropic-tracker-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de tendências de contratação e departamentos
- Busca de vagas abertas com dados frescos em minutos
- Monitoramento de mudanças em posições (adições/remoções)
- Consulta de faixas de compensação e alertas de RH
Veredito

Ferramenta bem arquitetada para consultas híbridas (cache + live) com design defensivo (read-only, sem mutações). Excelente para monitoramento estruturado de dados de hiring, mas nicho muito específico (Anthropic internal primarily).
README
A tradução para o português deste README ainda está na fila.
anthropic-tracker-mcp
MCP server that exposes Anthropic's hiring data to Claude Code (and any other MCP client). Two data sources, ten tools.
- 7 cached tools read from a read-only SQLite mount populated by
anthropic-tracker's nightly cron. Fast, indexed, good for trend analysis. - 3 live tools hit Greenhouse's public board API directly so you can bypass the cache when freshness matters (within minutes of a posting going up).
The DB layer is read-only by design: the volume is mounted with :ro and the SQLite driver opens it via file:...?mode=ro. The live layer is read-only by nature: Greenhouse's public API doesn't accept writes.
Architecture
+--> /data/tracker.db (read-only URI)
| (cached, nightly snapshot)
Claude Code --Streamable HTTP--> anthropic-tracker-mcp (port 3713)
|
+--> boards-api.greenhouse.io
(live, on-demand)
The same named volume anthropic-tracker-data is mounted read-write by the anthropic-tracker cron container and read-only here. The MCP server cannot mutate state, even if the code tried to.
Tools (10)
Cached DB (7)
| Tool | Description |
|---|---|
search_jobs(query, department=None, active_only=True) | Title LIKE search (comma = OR), optional department filter |
recent_changes(days=7) | Jobs added + removed in the last N days |
compensation_for(role_pattern) | Posted comp for jobs matching a title pattern (dollars) |
department_trends(name=None, days=30) | Per-department active count over time |
active_alerts(severity=None) | Unacknowledged alerts, optional severity filter |
daily_summary(date=None) | Snapshot for one date (default: latest) |
db_stats() | Row counts per table, latest snapshot, totals |
Live Greenhouse (3)
| Tool | Description |
|---|---|
live_jobs(query=None, department=None) | Fresh job list pulled directly from Greenhouse. Optional title search (comma = OR) and exact-match department filter. |
live_job_detail(job_id) | Full job posting (id, title, department, location, raw HTML content, url) for a single Greenhouse job ID. |
live_compensation(job_id) | Parse comp directly from the live HTML. Salaries in dollars. Returns {compensation: null} when no salary is published. |
All live tools return structured {error, detail} dicts on API failure (never raise — that would crash the MCP session). A 404 returns {error: "Job not found", job_id} distinctly.
All user input that flows into LIKE clauses (DB tools) is escaped via _escape_like() and parameterized — never f-stringed into the SQL. The same escape is applied to live query terms for consistency, even though matching happens in Python.
Stack
- Python 3.13 + FastMCP 3.1.0 (Streamable HTTP transport, per MCP spec 2025-03-26)
- SQLite (stdlib
sqlite3, read-only URI mode) - httpx (async, for live Greenhouse calls)
- beautifulsoup4 (HTML parsing for live compensation)
- Docker container, port 3713
Quick start
You need the anthropic-tracker-data Docker volume to already exist — that's created and populated by the anthropic-tracker project. Stand that up first.
git clone https://github.com/pete-builds/anthropic-tracker-mcp.git
cd anthropic-tracker-mcp
docker compose up -d --build
docker logs anthropic-tracker-mcp # confirm clean startup
The MCP server is now serving Streamable HTTP at http://localhost:3713/mcp.
Register with Claude Code
claude mcp add anthropic-tracker -t http -s user http://<host>:3713/mcp
Replace <host> with localhost if you registered on the same machine, or whatever address points to the container otherwise (LAN, Tailscale, reverse proxy).
File structure
anthropic-tracker-mcp/
server.py # FastMCP app — 10 @mcp.tool() definitions
healthcheck.py # Docker HEALTHCHECK
clients/
__init__.py
tracker.py # TrackerDB: read-only sqlite wrapper + queries
greenhouse.py # GreenhouseClient: async httpx + retry
parser.py # Compensation parser (synced from upstream tracker)
Dockerfile # python:3.13-slim, pinned digest
docker-compose.yml # Port 3713, RO volume mount, host network
requirements.txt
.env.example
CHANGELOG.md
LICENSE
Configuration
Environment variables (all optional):
| Var | Default | Purpose |
|---|---|---|
FASTMCP_HOST | 0.0.0.0 | Bind address (legacy alias: MCP_HOST) |
FASTMCP_PORT | 3713 | Bind port (legacy alias: MCP_PORT) |
TRACKER_DB_PATH | /data/tracker.db | SQLite path inside the container |
TZ | America/New_York | Container timezone |
Read-only by design (defense in depth)
| Layer | Mechanism |
|---|---|
| Docker | volumes: anthropic-tracker-data:/data:ro |
| SQLite driver | sqlite3.connect("file:/data/tracker.db?mode=ro&immutable=1", uri=True) |
| Code | No INSERT/UPDATE/DELETE statements anywhere; all tools return dict/list |
| Live tools | Only issue HTTP GETs against Greenhouse's public API |
If anyone tries INSERT INTO alerts ... inside the container, SQLite raises OperationalError: attempt to write a readonly database. The volume :ro mount also makes the file read-only at the kernel level. Three independent guarantees.
Why immutable=1 matters
The upstream anthropic-tracker writes the DB in WAL mode, which uses sidecar files (-shm, -wal) that SQLite needs to be able to create on the same filesystem as the DB. With our :ro mount, that creation fails. immutable=1 tells SQLite the file will not change for the duration of this connection, so it skips the WAL/journal/shm machinery entirely. That's also the correct posture for a read-only consumer that should never see in-flight writes.
Related
anthropic-tracker— the data pipeline (CLI, web dashboard, daily cron) that produces the SQLite DB this MCP serves
License
MIT — see LICENSE.