clear-pricer-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@clear-pricer-mcp What does an established-patient office visit (99213) cost at each Chicago hospital?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
clear-pricer-mcp
© 2026 Trevor J. Romack — MIT-licensed (LICENSE) · tjromack@gmail.com
Ask what a procedure costs at three Chicago hospitals from any MCP client, and get back rows cited to the hospital's own price file.
A TypeScript MCP server over the public data release of clear-pricer: CMS-mandated hospital price files and the NPPES provider registry, cleaned, reconciled and published as versioned Parquet. Every tool is read-only, typed end to end, and returns provenance with every row: the hospital, its source file, that file's SHA-256 and effective date, and the release tag.
Public data only — no PHI, no keys, no accounts, no telemetry.
Demonstrates: exposing a versioned data release as typed, grain-safe MCP tools, verified by planted-bug mutations.
Try it
claude mcp add clear-pricer -- npx -y clear-pricer-mcpThen ask, for example: "What does an established-patient office visit (99213) cost at each hospital, and why is one missing?"
Claude Desktop, Cursor and other clients take the same command in their MCP config:
{ "mcpServers": { "clear-pricer": { "command": "npx", "args": ["-y", "clear-pricer-mcp"] } } }The first question about a table downloads it once from the pinned release (every table but the 117 MB charge table is under 3 MB) into your OS cache, verified against the release's manifest.

Behind a corporate proxy
MCP clients start the server with a minimal environment. If your network inspects TLS, pass your CA to Node in the server's config:
{ "mcpServers": { "clear-pricer": { "command": "npx", "args": ["-y", "clear-pricer-mcp"],
"env": { "NODE_EXTRA_CA_CERTS": "C:\\path\\to\\corporate-ca.pem" } } } }Related MCP server: RecoSearch
Who it's for
People who want an AI assistant to answer hospital price questions from the hospitals' own published files, with a citation they can check rather than a number they have to trust; and engineers who want a worked example of an MCP server that cannot quietly double-count, average across incomparable rows, or answer from data it has not verified.
Tools
Tool | Answers | Reads |
| "knee MRI" → billing codes, from the hospitals' own descriptions; flags codes that matched on one hospital's wording only |
|
| One code across the hospitals, for one rate basis; names hospitals that publish it another way, and how |
|
| One code at one hospital, by payer and plan, each charge cited by its position in the source file |
|
| Who an NPI was, as of a date; every version; which hospital discloses it |
|
| How far to trust each hospital's file: NPI reconciliation, template deviations, quarantined rows |
|
| Which release is served, what it was built from, and each file's verification status |
|
A question nothing in the release can answer is an error that says what was searched and what does exist, never an empty result.
99213 (CPT_CAT_I), rate basis dollar, release data-2026-10-07-67efd3d2:
- RUSH University Medical Center, outpatient: median $185.00 (range $88.00–$251.60, 15 payer plans)
— source 362174823_rush-university-medical-center_standardcharges.csv, updated 2026-09-25
- The University of Chicago Medical Center, outpatient: median $61.65 (range $61.65–$61.65, 1 payer plan)
— source 363488183_the-university-of-chicago-medical-center_standardcharges.json, updated 2026-04-01
Note: Northwestern Memorial Hospital publishes this code, but not under rate_basis 'dollar':
outpatient / algorithm_only: 6 charge rows (no dollar rate); outpatient / dollar_from_percent: 276 charge rows.How it's verified
All results below are from data-2026-10-07-67efd3d2; docs/results/contract-tests.md
is generated by the run itself.
Every file is verified before it is read. The SHA-256 of the release's
manifest.jsonis pinned in src/release.ts; the manifest pins every file; a mismatch refuses to serve (tested with tampered files, a corrupted cache, and a wrong pin).88 tests in four suites: unit (17), contract (46: each tool through a real MCP client over slices of the release), end-to-end (3), and release (22: the real release recomputed against its own
check_values.json).Grain, at full scale. Joining charges to codes turns 7,371,416 charges into 22,647,893 rows (3.07×); the tools never do.
get_payer_rates' charge selection reproduces all 49,404charge_rowsinagg_code_priceswith 0 mismatches, and covers exactly the 5,704,751 charges it summarises. Integer-cent checksums of the rate and gross columns match the release.24 planted bugs, 24 caught.
npm run mutateapplies each realistic bug (averaging across settings, a fanned-out join, closed date intervals, skipped hash checks, empty answers returned as success, …), requires it to compile, and runs every suite. The first run found a behaviour no test covered; it has one now.Clean clone in CI on Ubuntu and Windows, Node 20 and 22, plus the release suite and the compiled server driven over stdio by an MCP client.
What this does not let you claim
Not a complete or authoritative price index: three Chicago hospitals, one pinned release, the slice clear-pricer publishes, with its reconciliation failures stated there.
Not a price estimate for any patient. Published negotiated rates are contract terms, not what a given person pays.
Descriptions are the hospitals' own and can be wrong: UChicago describes CPT 44373 (small-bowel endoscopy) as a functional brain MRI. The tools surface such disagreements; they do not correct them.
The NPPES provider history (563 MB) is read remotely by HTTP range, so its row count is checked against the manifest but it is not hash-verified like every other file.
Not a hosted service: stdio only, running on the user's machine.
Develop
git clone https://github.com/tjromack/clear-pricer-mcp && cd clear-pricer-mcp
npm ci && npm test # offline, fixture-backed, a few seconds
npm run test:release # the real pinned release against its check_values.json (~120 MB download the first time)
npm run mutate # the 24-mutant suite; rewrites docs/results/
npm run build && npm run smoke # the compiled server over stdio, one question per toolThe case study is docs/CASE-STUDY.md. Design decisions and what was rejected are in DECISIONS.md; the build journal is docs/BUILD-LOG.md.
Available Tools
6 toolscompare_code_pricesCompare one billing code's prices across hospitalsARead-onlyIdempotent
Negotiated-rate summary (min, quartiles, max over payer plans) for one billing code at each hospital in the release, for one rate basis, with the hospital's source file cited on every row. Hospitals that publish the code differently are listed in not_included with what they do publish.
Inputs: code (required), setting, rate_basis, code_family. Returns: code, notes, prices, setting, +4 more
data_qualityHow far to trust each hospital's price fileARead-onlyIdempotent
Per hospital: whether the NPIs its price file discloses resolve to it in NPPES, how many of its NPIs it leaves undisclosed, how its file departs from the CMS template (and how often), and rows quarantined. Use it to qualify any price answer.
Inputs: hospital_id. Returns: notes, overall, hospitals, release_tag
find_codesFind billing codes from a plain-language descriptionARead-onlyIdempotent
Search the hospitals' own item descriptions (as published in their price files) for billing codes, e.g. 'mri brain' or 'office visit established'. Returns each code with the hospitals that publish it and their descriptions. Use the code with compare_code_prices or get_payer_rates.
Inputs: query (required), limit, code_family, hospital_id. Returns: notes, query, results, truncated, +2 more
get_payer_ratesNegotiated rates for one code at one hospital, by payer and planARead-onlyIdempotent
Every published charge for one billing code at one hospital: payer, plan, setting, negotiated rate (or the percentage/algorithm when no dollar figure is published), gross and cash price, with the charge's position in the hospital's source file. The first call downloads the 117 MB charge table once.
Inputs: code (required), hospital_id (required), limit, payer, +2 more. Returns: code, notes, rates, returned, +6 more
lookup_providerWho an NPI belonged to, as of a dateARead-onlyIdempotent
Look up a National Provider Identifier in the NPPES registry history: name, entity type, status, practice address and taxonomy as of a date (or now), every version on record, and whether a hospital in the release discloses it. Validates the NPI check digit first.
Inputs: npi (required), as_of. Returns: npi, name, as_of, notes, +17 more
release_infoWhich data release is being servedARead-onlyIdempotent
The pinned clear-pricer release this server answers from: its tag, the manifest hash it was verified against, the hospital price files and NPPES files it was built from, and each table's row count and status.
Inputs: no named parameters. Returns: files, origin, fingerprint, nppes_files, +4 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Each tool targets a distinct resource and action: code search (find_codes), cross-hospital price comparison (compare_code_prices), single-hospital payer detail (get_payer_rates), provider lookup (lookup_provider), data quality checks (data_quality), and release metadata (release_info). No overlapping responsibilities; the descriptions clarify the boundaries between price-related tools.
All names use snake_case and follow a verb_noun or noun_noun pattern, which is readable and consistent. Minor deviations: 'data_quality' and 'release_info' are noun phrases without a verb, unlike the others, but the overall convention is predictable.
Six tools is well-scoped for a domain focused on hospital price transparency data. Each tool serves a clear purpose: discovery, comparison, detail, provider validation, quality assurance, and provenance. No tool feels redundant or missing.
The surface covers the core lifecycle: finding codes, comparing prices across hospitals, drilling into payer rates, validating providers, assessing data quality, and retrieving release info. Minor gaps include a possible tool for bulk or geographic search, but the current set handles typical agent queries effectively.
Maintenance
Related MCP Connectors
Free hosted MCP server: US healthcare procedure cost estimates and provider pricing. No API key.
One governed MCP endpoint for 40+ business systems. Read-only, source-cited answers.
Governed data discovery, exact queries, decisions, simulations, and runtime utilities over MCP.
Query your org's data in natural language — read-only MCP access to SQL, NoSQL, files & warehouses.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceSource-provenanced US federal healthcare provider data over MCP. Resolve any NPI or CCN across NPPES, OIG LEIE, SAM.gov, state Medicaid exclusions, CMS PECOS, Care Compare, and Open Payments — every field carries a 14-field provenance contract, and an "excluded or compromised anywhere" check runs on every lookup.-

RecoSearchofficial
AlicenseNot gradedqualityCmaintenanceA deterministic MCP server that governs read-only queries across multiple data sources, returning answers with full provenance (every row cited) or a typed refusal, ensuring LLM answers are traceable and contract-enforced.1Apache 2.0- FlicenseNot gradedqualityCmaintenanceRead-only MCP server providing 5 tools for hybrid search, clause retrieval, policy versioning, code lookup, and plan rider override queries over a synthetic medical-policy corpus.-
- AlicenseAqualityCmaintenanceProvides governed retrieval over MCP with hybrid search, strict confidence gating, and access control, exposing three read-only tools.3Apache 2.0