blind-mcp

Find out what a role actually pays — including in markets where the employer publishes nothing.

Colorado, California, New York, Washington and Illinois require a salary range on covered job postings. India, Singapore and most of the EU require none. So the same title, at the same company, in the same week, is posted with a band in Denver and without one in Bengaluru.

The band is still there. It is just attached to a different listing.

pay_bands(company="Databricks", role="forward deployed")

  101 matching openings · band width ratio x1.37 · 33 publish no range

  LEVEL            TYPICAL BAND        BANDS  POSTS  PRECISION
  mid              152,900–210,155         2      3  tight
  senior           182,000–250,208         1     58  tight
  lead             178,800–245,850         1      1  tight
  manager          211,800–291,300         4      5  tight
  senior_manager   232,900–320,200         1      1  tight

  steps: mid→senior +19.1% · lead→manager +18.5% · manager→sr_mgr +9.9%
  silent: Remote-India, Seoul, London, Berlin, Amsterdam

Per level, because one range across seniorities is a number nobody is offered — undivided, that role reads as 140,400–320,200, a 2.3x spread covering five different jobs.

typical is the modal band. distinct_bands vs postings is the honest measure of evidence: 58 postings sharing one band is one data point advertised 58 times. precision flags how much a band actually narrows things — Databricks posts tight per-level bands, Figma posts a single x2.5 band spanning its whole ladder, and both are real.

That is the same employer, the same title, the same moment — a far better anchor for an unpublished number than any salary survey, and it takes one call.

Reads the public job-board APIs (Greenhouse, Ashby, Lever). Documented, intended for machines, and stable — not scraping.

Tools

Pay

Culture — currently blocked

⚠️ Blind began returning 403 to all automated requests around September 2026. This is site-wide bot protection, not a block on this project: curl and an empty User-Agent are refused too, and only a browser User-Agent gets through. robots.txt still permits /company/, but the WAF does not.

We do not spoof a browser to get around it — that would be circumventing an access control, and it is the first thing their next escalation defeats. The Blind tools now raise BlindBlocked with an explanation rather than returning empty results that would read as "no discussion found".

The pay tools are unaffected. See #12.

Coverage

Measured, not estimated:

Not covered: employers who self-host — Google, Meta, Amazon, Apple — and most Indian-headquartered companies. BoardNotFound explains the two causes rather than just failing.

If a company is on one of these boards under a token you can't guess, read it out of their careers URL and pass it directly:

pay_bands("Some Rebranded Co", "engineer", board="greenhouse:theirslug")

Adding a Workday or SmartRecruiters adapter is #13.

Install

uv sync

Register with Claude Code:

uv tool install --editable .          # puts `blind-mcp` on PATH
claude mcp add --scope user blind -- blind-mcp

Or in claude_desktop_config.json:

{
  "mcpServers": {
    "blind": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/blind-mcp", "blind-mcp"]
    }
  }
}

Do I need to log in?

No — and you probably shouldn't. This was measured, not assumed:

• Post bodies, full comment threads, Blind's AI comment summaries, company listings, topic pages, reviews and salary pages all return HTTP 200 anonymously, with no gating.
• Driving an authenticated session with automation tripped Blind's anomaly detection after roughly six navigations: Automatic Logout — we noticed a login from a new device or location (code 2009), with a redirect to /session-out.

So a cookie buys no extra read access and costs you session stability. The only thing it would unlock is company-internal channels, which are gated to verified employees of that company.

If you still want it, set BLIND_COOKIE to your session cookie header (copy it from a logged-in browser request in DevTools). There is no login flow — the server only replays a cookie you supply, read once at startup and attached to each request. It is off by default, and the client raises immediately if Blind invalidates it rather than silently returning logged-out HTML.

The cookie itself is never written to disk, but responses fetched with it are cached under a separate auth/ directory so authenticated and anonymous results can never be served for each other. Don't commit the cookie.

Being a good citizen

Blind's robots.txt disallows /search/ for every user-agent, so this server has no search tool and refuses to fetch that path. It also:

• sends an honest User-Agent (no browser impersonation — Blind serves it a 200 anyway)
• caches every response on disk for 6h, so repeat questions cost zero requests
• spaces requests ~1.5s apart
• fetches robots.txt first and fails closed if it can't be read

Configure via BLIND_MCP_CACHE_DIR, BLIND_MCP_CACHE_TTL, BLIND_MCP_MIN_INTERVAL, BLIND_MCP_USER_AGENT.

Blind's Terms of Service restrict automated access. This reads public pages at human pace for personal research; bulk crawling is both a ToS problem and, given that Blind's value rests on anonymity, a privacy one. Don't build a dataset of posts joined to employers and nicknames.

Reading the output

Blind is anonymous and unverified. Weight claims by the commenter's employer (company on each comment) and treat a single loud voice as one data point. In testing, the Roku India RTO answer was corroborated by two independent commenters and an unrelated Glassdoor review — that's when it's worth trusting.

Contributing

Small project, easy to contribute to. The cheapest useful change is a keyword alias — one dict entry plus a test — and it's the change that most improves answers, because research only finds threads whose words match yours.

Start with good first issues — nine open, all real and reproduced, each one saying where the code is and how to test the fix — then CONTRIBUTING.md. Issues are labelled by size (size: XS is under 30 minutes) and mentored means ask questions in the thread and you'll get walked through it. First review within 48 hours.

Two hard rules, both explained in CONTRIBUTING: don't commit captured Blind pages (fixtures are generated), and don't weaken robots/cache/throttle for speed.

Publishing

mcp-name: io.github.dheerajjha/blind-mcp

See RELEASING.md.

License

MIT