Harris County MCP server

harris-county — Harris County (TX) District Clerk civil dockets

Status: incubator. Not wired into MCPPACKS, so nothing routes here. Fleet #1085. This pack holds the two halves that are finished and measured*; the tools are not written yet because they read a mirror table that does not exist. See ../README.md for the promotion checklist.

Part of Pipeworx — an MCP gateway connecting AI agents to 1576+ live data sources.

What the corpus answers

"Is Company X being sued in Harris County state court, and by whom." That is the question court-listener structurally cannot answer — it covers federal dockets and appellate opinions, not Texas state civil filings. Harris County is the largest single-county state civil corpus we have a sanctioned bulk path to.

Measured against the 20260829 weekly index: AMAZON LOGISTICS INC 12 cases, KROGER 21, WALMART INC 15, HALLIBURTON COMPANY 3, SYSCO HOUSTON INC 3.

What is here

Verified, so nobody re-derives it

All of this was measured on 2026-09-02 against the real files, not inferred.

Reachability from a deployed Cloudflare Worker — PASS. GET /admin/harris_probe on the deployed registry-api (not wrangler dev, not a laptop):

egress: cloudflare-worker, elapsed 2,244ms, 164 datasets advertised, 144 civil
daily   200  content-disposition: attachment;filename=2026-08-31 CivilHistoricalDaily.txt   207,352 B  1,385 lines
weekly  200  content-disposition: attachment;filename=Civil Case Index JWEB 20260829.txt  20,783,741 B  31,540 lines

Weekly fixed-width layout — 42 fields, 0 misaligned rows out of 31,539. Offsets came from the header's label positions and are asserted on every parse by assertWeeklyHeader(). 228 case numbers legitimately end in a letter (severed/ancillary matters).

Field decoding — 17 of 17 status codes present in the file are mapped, none unmapped: A ACTIVE 12,861 · D DISPOSED (FINAL) 12,454 · E READY DOCKET 3,342 · H HOLD FOR JUDGMENT 1,317 · O POST JUDGMENT 868 · IP INACTIVE (PENDING) 63 · … Raw D is not a case status anyone can read; that is what FIELD_CODES.xlsx is for.

Three traps, all of which return HTTP 200

and the server answers 200 with the page re-rendered — a 224 KB HTML document where the file should be. curl --data-urlencode leaves $ literal; URLSearchParams encodes it. downloadHarrisFile asserts on Content-Disposition: attachment, never on the status code.

1. The $ in the postback control names must arrive as %24. Get it wrong

slices the wrong characters, and you ingest half a street name as a defendant. assertWeeklyHeader() is the only thing between that and the database — call summarizeWeekly() at ingest and require misaligned === 0.

1. A shifted column does not error. Every field still slices, it just

787 rows carry it. Ingested literally it becomes the third most-sued party in Harris County, ahead of the City of Houston. parseWeeklyLine blanks it.

1. NO DEFENDANT RECORD found is a sentinel, not a name.

What is left

Everything except the load itself is now written and typechecked: the schema, the ingest, its manual trigger, and the four tools. What remains needs database credentials the machine that wrote it does not carry, so it stopped at the last step that could be honestly verified from there rather than at a half-loaded table.

Shortened. The full README is on GitHub.