@pipeworx/h1b

US H-1B visa sponsorship and wage data from DOL Labor Condition Application (LCA) disclosures — "does company X sponsor H-1B", "what does role Y pay at company X", "who sponsors the most data engineers in Austin".

Part of Pipeworx — an MCP gateway connecting AI agents to 1865+ live data sources. This is an independent, unofficial integration — not affiliated with, endorsed by, or published by the upstream provider.

Tools

• h1b_employer_sponsorship(employer, year?) — filing count, salary range, top titles/locations for an employer.

• h1b_salary(job_title, employer?, city?, year?, limit?) — real disclosed base-salary stats + a sample of records for a role.

• h1b_top_sponsors(job_title, city?, year?, limit?) — employers ranked by filing count for a role, with median salary.

Argument names/shapes are unchanged from before fleet #2514 — only where the answer is served from changed.

Related MCP server: h1b-mcp

Two-tier source (fleet #2514)

Primary: DOL's own LCA disclosure files for recent federal fiscal years. Older years come from h1bdata.info. Every DOL-sourced response carries data_as_of.

Fallback: h1bdata.info, a third-party aggregator of the same DOL data. Used only when the requested year is outside the local mirror's loaded window, or the DB is unreachable — this pack must never regress below its pre-#2514 behavior while the mirror is partially backfilled. Every response carries a source field saying which one actually answered it; the pack never silently swaps sources without saying so.

A fallback zero is never presented as a confident "no" (fleet #3198). If the DB read itself failed this call (no credentials, a timeout, or a transient error — source_status: "primary_unreachable"), a zero-row h1bdata.info answer returns sponsors_h1b: null and a message saying the primary source was never actually consulted, not disproven. Only when the primary source answered fine but the year is genuinely outside its loaded window (source_status: "out_of_window") does a fallback zero stand as a real (if third-party, lag/rename-prone) "no". dbRpc retries once immediately on a transient failure before falling back at all, so a single network blip doesn't by itself demote a mirror-covered query to the weaker source. h1b_employer_sponsorship's matched_employers is also now ranked exact-match first (then name-contains variants), each tagged name_match, since the employer-name match is a plain substring and can legitimately pull back more than one distinct legal name.

Why a mirror at all: the live h1bdata.info scrape caps at 20,000 raw HTML table rows per query (fetchLcaLive's safety cap) — for any employer or job title with more filings than that, salary stats and top_sponsors rankings were being computed from a silently truncated, non-representative slice. For DOL-sourced years, top_sponsors aggregates over every matching row, with no such cap (see total_all_filings in 222_h1b_lca_disclosures.sql's h1b_top_sponsors RPC, a window-function sum computed before the row-limit).

Fiscal year, not calendar year, on the local mirror. DOL's files are organized by federal fiscal year (Oct 1 - Sep 30); year is passed straight through as that fiscal year when served locally. h1bdata.info's own "year" filter is a different, unverified semantic (most likely calendar year) — a query answered by each path for the same year value is not guaranteed to cover the identical 12 months. The response's note/source fields say which happened; this is a real, disclosed difference, not papered over.

PERSONAL-DATA RULE (mandatory, local-copy rule — task #2514)

DOL's own record layout (LCA_Record_Layout_FY2024_Q4.pdf, read in full 2026-09-29) carries named-individual contact fields that are never read into the mirror — scripts/h1b-lca-transform.py uses an ALLOWLIST of source columns (not a blocklist), so a column DOL adds later is excluded by default:

• EMPLOYER_POC_{LAST,FIRST,MIDDLE}_NAME, EMPLOYER_POC_EMAIL, EMPLOYER_POC_PHONE — the employer's named point of contact.

• AGENT_ATTORNEY_{LAST,FIRST,MIDDLE}_NAME, AGENT_ATTORNEY_EMAIL_ADDRESS, AGENT_ATTORNEY_PHONE — the attorney/agent representing the employer.

• PREPARER_{LAST,FIRST,MIDDLE}_NAME, PREPARER_EMAIL — whoever filled out the form.

DOL's file already excludes the foreign worker's own name/address (DOL's own cover note names only "Attorney's FEIN and Attorney's State Bar Number" as withheld — no worker PII was ever in this dataset). Kept: EMPLOYER_NAME / TRADE_NAME_DBA and employer address down to city/state/country (no street), plus job/wage/worksite fields — this identifies the employer making a mandatory public disclosure, not a private individual.

Data source

• https://www.dol.gov/agencies/eta/foreign-labor/performance — DOL OFLC performance-data page, links each fiscal year's disclosure file.

• Verified live 2026-09-29 (not documented by DOL, reverse-engineered from decision_date distributions — the record-layout PDF's cover text is WRONG about this for closed years):

  • The current, still-open fiscal year's latest quarterly file is cumulative from Oct 1 of that FY through the end of the named quarter (FY2026_Q3.xlsx spans Oct 2025-Jun 2026). Lives at https://www.dol.gov/media/LCA_Disclosure_Data_<label>.xlsx. Load only the single latest file for an open year.

  • Once a fiscal year closes, DOL's archived quarterly files for that year are each discrete — decision_date falls only within that one quarter (verified: FY2025_Q3.xlsx = Apr-Jun 2025 only, FY2025_Q4.xlsx = Jul-Sep 2025 only — NOT the full year, despite the record-layout PDF literally stating "Reporting Period: October 1 ... through September 30" on its cover page). Lives at https://www.dol.gov/sites/dolgov/files/ETA/oflc/pdfs/LCA_Disclosure_Data_<label>.xlsx. A closed year needs all four of its quarter files for full coverage; they do not overlap.

• No auth, no rate limit observed. US federal administrative data — public domain.

Ingest harness

• supabase/migrations/222_h1b_lca_disclosures.sql — h1b_lca_disclosures table + h1b_lca_ingest_runs bookkeeping + three RPCs (h1b_lca_search, h1b_top_sponsors, h1b_lca_coverage). RLS on; anon/authenticated have no grants; service_role only.

• scripts/h1b-lca-transform.py — streams one .xlsx file row-by-row via openpyxl(read_only=True) (never buffers a whole workbook — a full fiscal year runs 80-250MB uncompressed), allowlist-selects the non-personal columns, and computes base_salary_annual from WAGE_RATE_OF_PAY_FROM/WAGE_UNIT_OF_PAY (Hour×2080, Week×52, Bi-Weekly×26, Month×12, Year×1).

• scripts/h1b-lca-upsert.sh — per file: download (tries the current-year URL, falls back to the closed-year archive URL), transform, \copy into a staging table, then INSERT ... ON CONFLICT (case_number) DO UPDATE (CASE_NUMBER is DOL's own stable per-application id — no synthetic key needed).

• .github/workflows/h1b-lca-refresh.yml — workflow_dispatch only (deliberately no schedule: — Bruce's "supa stagger" rule queues big Supabase loads one at a time, overnight, on purpose; GitHub's schedule: trigger has also been unreliable fleet-wide since ~2026-09-09). Default files input is the "latest 8 quarters" the task asked for, with verified zero overlap and zero gap: 2026:FY2026_Q3 (cumulative, = FY2026 Q1+Q2+Q3) + 2025:FY2025_Q1,2025:FY2025_Q2,2025:FY2025_Q3,2025:FY2025_Q4 (four discrete quarters) + 2024:FY2024_Q4 (one discrete quarter) — 6 files, Jul 2024-Jun 2026.

Observed sizes/rows (live, 2026-09-29 — proof-of-transform run, not a full load)

Estimated total for the default 6-file load: ~1.2-1.3M rows, ~650MB of xlsx downloads. Every .xlsx sheet carries a large trailing block of genuinely-empty phantom rows past the real data (e.g. FY2026_Q3's declared sheet dimension is 1,032,736 rows for 437,496 real ones) — this is a DOL export artifact, not a bug in the transform script; those rows are dropped and counted as "without CASE_NUMBER".

Known source data-quality trap

DOL's raw WAGE_RATE_OF_PAY_FROM/WAGE_UNIT_OF_PAY occasionally disagree in a way that produces an absurd base_salary_annual (an observed FY2026_Q3 row annualizes to ~$453M — almost certainly a unit-entry error by the filer, kept verbatim rather than "corrected", same discipline as sba_loans keeping literal duplicate/conflicting source rows). median-based stats are resistant to this; a max in a response can occasionally show one of these outliers — that is the source data, not a mirror defect.

Gotchas

• The live DOL .xlsx header spells the H-1B-dependent-employer column H_1B_DEPENDENT (underscore) — the record-layout PDF spells it H-1B_DEPENDENT (hyphen). Trust the live file; verified against FY2026_Q3.xlsx 2026-09-29.

• Do not load more than one file per still-open fiscal year, and do not load a closed year's file without all four of its quarters — see "Data source" above. Getting this wrong either duplicates rows (double-counting an overlapping cumulative + discrete file) or leaves silent gaps.

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "h1b": {
      "url": "https://gateway.pipeworx.io/h1b/mcp"
    }
  }
}

What this endpoint actually serves

tools/list at https://gateway.pipeworx.io/h1b/mcp returns the tools in the table above plus the shared Pipeworx meta-tools — ask_pipeworx, discover_tools, search_within, remember/recall and the rest of the gateway-wide set. So the tool count you see is larger than this table: a single-pack endpoint currently lists roughly 30 shared tools alongside the pack's own. The connection's initialize response states its exact scope, and is the authoritative answer for a given day.

This is deliberate, not multiplexing by accident. The meta-tools are what let a scoped connection answer a question this pack does not cover — via ask_pipeworx, which routes across the whole catalog — without you adding a second MCP server. There is currently no way to mount a pack endpoint without them; if the extra schemas cost you more context than the routing is worth, connect to the full gateway once rather than to several pack endpoints.

Or connect to the full Pipeworx gateway to get every pack's tools listed directly, instead of just this one's:

{
  "mcpServers": {
    "pipeworx": {
      "url": "https://gateway.pipeworx.io/mcp"
    }
  }
}

Both URLs reach the same gateway and the same 1865+ data sources. The only difference is which pack's tools are listed directly; ask_pipeworx reaches all of them from either one.

No MCP client? Call it over HTTP

curl -X POST https://gateway.pipeworx.io/v1/tools/h1b_employer_sponsorship \
  -H 'Content-Type: application/json' \
  -d '{"employer":"Google","year":2024}'

No account needed for the first calls. Inspect any tool: GET https://gateway.pipeworx.io/v1/tools/h1b_employer_sponsorship. Find one: POST https://gateway.pipeworx.io/v1/tools/search_packs with {"query":"..."}.

Run it from npm

@pipeworx/mcp-h1b is a local stdio bridge to the Pipeworx gateway: your MCP client starts it with npx, and it forwards every call to the same https://gateway.pipeworx.io/h1b/mcp endpoint shown above. It needs network access; there is no offline mode.

{
  "mcpServers": {
    "h1b": {
      "command": "npx",
      "args": ["-y", "@pipeworx/mcp-h1b"]
    }
  }
}

Without a key it calls anonymously, capped at 50 calls a day. A free account raises that to 200 a day: get a key at https://pipeworx.io/signup and add "env": { "PIPEWORX_API_KEY": "<your key>" } to the server entry. The key is sent as a Bearer token, the same way a hosted connection sends it, so calls are metered and billed exactly as they are there.

It lists only this pack's tools by default. Set PIPEWORX_META_TOOLS=1 to also list ask_pipeworx and the other gateway-wide tools.

Behind an HTTPS proxy: Node's built-in fetch does not read HTTPS_PROXY by default. Newer Node releases honour it when NODE_USE_ENV_PROXY=1 is set; check your Node version's documentation.

Using with ask_pipeworx

Instead of calling tools directly, you can ask questions in plain English — this works on the pack endpoint above as well as on the full gateway:

ask_pipeworx({ question: "your question about H1b data" })

The gateway picks the right tool and fills the arguments automatically.

More

• Docs and guides

• pipeworx.io

License

MIT