H-1B Visa Data MCP Server

H-1B Visa Data

H-1B, H-1B1, and E-3 Labor Condition Application (LCA) disclosure data published by the US Department of Labor's Office of Foreign Labor Certification. ~675K applications across the most recent five quarters. Updated weekly from the OFLC quarterly XLSX releases.

Free, no authentication required.

What an LCA is (and isn't)

LCAs are filings, not visa grants. An employer files an LCA with DOL before sponsoring a foreign worker; a certified LCA is a precondition for the H-1B petition that USCIS later approves or denies. So this dataset tells you who applied to sponsor whom, at what offered wage, in what occupation — not who actually got a visa, started work, or is currently employed.

Data Source

Capabilities

• Search employers by name — fuzzy match on the employer name. Optional state filter on employer state. Returns aggregate stats per employer (application counts, certification rate, average annual salary). Cap: 100 results per call.
• Search applications by job, occupation, or worksite — at least one of job_title (fuzzy), soc_code, or employer is required. Further narrow by state (worksite) and visa_class. Returns individual application records, not aggregates. Cap: 200 results per call.
• Employer detail — given an exact employer name from a search result, returns every application plus aggregate stats (certification rate, salary range, position counts, top job titles, top worksites). Cap: 500 applications.
• Dataset statistics — totals, certification counts, employer count, average salary, top sponsors, data freshness.

Field model

• SOC code — Standard Occupational Classification, e.g. 15-1252.00 (Software Developers), 15-2051.00 (Data Scientists). Self-reported by the employer at filing time; see the SOC mismatch note in the gotchas.
• Wage level — DOL prevailing-wage tier I (entry) through IV (fully competent). Wage level I roles are typically junior and often command intense scrutiny on RFEs.
• Case status — CERTIFIED, CERTIFIED-WITHDRAWN, WITHDRAWN, DENIED. "Certification rate" in aggregates excludes withdrawals.
• Wage — the offered wage at filing time, not what the worker is actually paid after the petition is approved. Returned as a string with unit suffix ($176,866/Year, $35.00/Hour, etc.) — see the wage-units gotcha.
• Visa class — the jobs-search visa_class filter accepts the short alias (H-1B, E-3, H-1B1) or the literal stored value (E-3 Australian, H-1B1 Singapore, H-1B1 Chile). The response always shows the literal stored value.

Workflows

Sponsorship lookup: Search the employer by name → review aggregate certification rate, average salary, top job titles, and top worksites. Use the exact employer name from the result for the detail call when full application history is needed.

Wage benchmarking by role: Search applications by job title (preferred — see SOC gotcha) → filter to a specific worksite state if local market matters → normalize wages to annual before averaging (units are mixed) → discard salary outliers above ~$1M and below ~$50K before computing means for general roles.

Sponsor discovery for a region/role: Search applications with job_title + state (no employer) → group results by employer to find who sponsors a given role in a given state.

Trend / volume: Get dataset statistics for top sponsors and totals; combine with per-employer detail for year-over-year comparisons within the loaded quarters.

Gotchas

• At least one of job_title, soc_code, or employer is required for the jobs search — state-only or visa_class-only queries return an error. Always pair a location/visa filter with one of those three.
• Empty results return silently with results_count: 0 and an empty applications list. Always check results_count > 0 before treating a response as positive evidence; a "no sponsors in state X" conclusion based on an empty result is only valid if at least one identifier was supplied.
• Wage units are mixed within a single result set — /Year, /Hour, and occasionally /Bi-Weekly co-exist. Normalize to annual before averaging or comparing across rows. The aggregate avg_annual_salary returned by the employer search and detail endpoints already filters to /Year rows only and is safe to use directly.
• SOC code does not always match the job title — both fields are employer-self-reported. The same job title can carry several different SOC codes across employers (e.g. "Product Manager" is filed as Software Developer, Industrial Engineer, Marketing Manager, IT Project Manager, and Operations Research Analyst across different employers; "Chief Financial Officer" is filed as Chief Executive, Treasurer/Controller, General Manager, and Financial Specialist). For role-level analysis, prefer searching by job_title and treat SOC as supplementary; for occupation-level cross-employer comparison, use SOC and accept that some titles will be misclassified.
• Salary outliers exist — most rows are within plausible market ranges, but C-suite filings legitimately show $1M-$2M offered wages, and a handful of rows look like data-entry errors (extremely high values, $30K-or-less for senior roles). Filter out values above ~$1M and below ~$50K before computing means or medians for general roles; keep them for executive analysis.
• Aggregate avg_annual_salary can be null — if an employer's filings are all /Hour or /Week (no /Year rows), the year-only average comes back as null. Detail-level wage fields are always populated; only the aggregate is missing. Handle the null explicitly.
• Fuzzy employer search is broad — short or generic names (Inc, Tech, Group) return many unrelated companies, all with match_score: 1.0. Use ≥2 distinctive words (Amazon Web, Goldman Sachs) to narrow. The match_score field on each result reflects pg_trgm similarity and decays for weaker matches; sort or filter on it when over-matching is a problem.
• Employer-name variants are not deduplicated — the same parent (e.g. Amazon) appears as Amazon.com Services LLC, Amazon Web Services, Inc., Amazon Development Center U.S., Inc., plus uppercase/lowercase duplicates per state. Aggregate stats are per matched name, not per parent. Combining variants for parent-level analysis is the caller's responsibility.
• Worksite vs. employer state are different — employer-level filters (state on the employer search) match the employer's address; application-level filters (state on the jobs search) match the worksite address. Remote work is recorded as the worksite the employer actually filed for.
• Geographic fields contain occasional bad data — a small number of rows have placeholder or fictitious worksite cities (literal Test City rows, made-up place names). Treat one-off worksite values for a single application with skepticism.
• limit is silently capped per tool — employer search at 100, jobs search at 200, employer detail at 500. Requesting larger values returns the cap with no warning.
• Five quarters of history — this dataset is rolling, not full historical. Long-range trend analysis (multi-year) is out of scope.
• DENIED is rare in raw counts — most rejected H-1B sponsorships never reach DOL denial; they fail later at USCIS. Don't treat a 99%+ certification rate as a quality signal.