mcp server

Constat MCP — FDA Device Evidence Lifecycle

FDA and CMS evidence for AI medical devices: 510(k), postmarket, reimbursement, and compliance.

Description as published by the maintainer. Source

• version 0.5.3
• archived
• security

archived — The linked repository returns 404. It was deleted, renamed or made private. Dashed tags are derived by ZBS Index from the published description, not stated by the maintainer.

What this server can do

14 functions, named and described by the server itself. Parameter names are shown because they say more about what a function does than its name usually does.

cohort_postmarket_stats(panel)

Postmarket presence rates across the snapshotted AI/ML device cohort (optionally by panel): share with any recall in 24 months, with a rising MAUDE trend, with any drift signal, with a warning-letter match — every rate with its denominator inline, never pooled across devices.

device_evidence_lookup(k_number)

Look up the structured premarket evidence FDA accepted for a specific AI/ML-enabled device by 510(k) number (e.g. K252148). Returns parsed summary fields — validation study design, sample sizes, endpoints, reported performance, predicate chain, PCCP — each with a verbatim source quote and page. Null means the summary did not state it. Required: k_number.

device_postmarket_lookup(k_number)

Post-clearance intelligence for one AI/ML device by 510(k) number: its product code's recalls, MAUDE adverse-event level and trend, warning-letter and 483 matches for the applicant, plus per-device drift signals (adverse-event inflection, re-clearances of the same device line, software-recall patterns, predicate-cohort recall activity). Descriptive observables with sources — never a safety judgment. Required: k_number.

device_risk_lookup(product_code)

Review a medical-device category's public FDA signals by three-letter product code (e.g. FRN = infusion pump). Returns recalls, MAUDE adverse-event trend, warning-letter matches, a normalized category signal, its driver contributions, and interpretation limits. It does not predict enforcement against a firm. Required: product_code.

evidence_cohort_stats(panel)

Reporting-rate stats across the parsed AI/ML corpus (optionally by panel). Each rate is a presence figure with its denominator — 'reported in X of Y audited devices' — never a pooled performance value. Excludes not-yet-parsed devices from every denominator and discloses the parse queue separately. Predicate age (median years between a clearance and its cited predicates) is included when decision-date coverage clears a 60% floor, and withheld otherwise.

evidence_search(limit, panel, cursor, has_pccp, applicant, product_code, has_clinical_data, reports_any_sensitivity_metric)

Find AI/ML device clearances by filter — product code, panel, applicant, and whether the submission reported clinical data, any sensitivity metric, or a PCCP. Answers 'what evidence did FDA accept for devices like mine'. Returns matching records with their parsed evidence. Presence flags are descriptive: 'reports a sensitivity metric' is not 'reports a comparable sensitivity' — analysis units differ across devices.

firm_compliance_history(limit, since, firm_name, fei_numbers, product_codes)

Build a recent, source-bounded FDA public-record timeline for a device firm: matched recalls, warning letters, and Form 483 citations where exact FEI numbers are available. Product codes are discovered from Constat's AI/ML-device corpus or may be supplied explicitly. Returns attribution and coverage limits with the records; it is not a finding of noncompliance or a prediction of FDA action. Required: firm_name.

postmarket_search(limit, panel, cursor, applicant, product_code, maude_trend_up, has_drift_signal, min_recalls_24mo, signal_specificity)

Find AI/ML devices by postmarket criteria — product code, panel, applicant, whether any drift signal exists, minimum recalls in 24 months, or a rising MAUDE trend. Returns per-device postmarket summaries with drift-signal counts.

predicate_chain(depth, k_number)

Trace the predicate ancestry of a 510(k) device, with each cited predicate's age (how many years old the predicate was when the child cleared). Reveals how AI/ML devices chain to older predicates. Required: k_number.

reimbursement_lookup(cpt_code, k_number)

Trace the clearance-to-payment pathway for an AI/ML device by FDA clearance number (K/DEN, e.g. DEN170073) OR bare CPT code (e.g. 75580). Returns every payment mechanism (NTAP add-on, Category I/III CPT + CMS rate, HCPCS, MAC LCD) with amounts, effective dates, and source links, plus any commercial/MAC payer coverage policies that reference the clearance or its codes. Answers 'who got paid, how much, through which mechanism, on what basis.' CPT codes are bare factual identifiers only — no procedure descriptors; follow the CMS source link for the official descriptor.

reimbursement_search(limit, cursor, applicant, mechanism, ntap_status, cpt_category, has_cms_rate)

Find AI/ML device payment pathways by mechanism — e.g. 'devices that got NTAP', 'devices paid under a Category I CPT code', 'pathways with a known CMS dollar rate'. Filters: mechanism, CPT category, NTAP status, applicant. Returns pathways with amounts, effective dates, and sources. Use reimbursement_stats for the mechanism distribution (never a single pooled reimbursement rate).

reimbursement_stats

Distribution of payment mechanisms across the AI/ML reimbursement corpus — pathway and distinct-device counts per mechanism (NTAP, Cat I, Cat III/APC, …) with the min/median/max dollar amounts for each. Deliberately never a single pooled 'reimbursement rate': NTAP add-on amounts and CMS rates are different measurements and are reported separately with their own spreads.

vehicle_risk_lookup(make, year, model)

Look up NHTSA safety history for a vehicle by make, model, and model year. Returns recall campaigns and complaint statistics (crashes, fires, injuries, top components). Required: make, model, year.

watchlist_diff(limit, since, product_codes)

Return machine-generated FDA public-record changes detected for monitored product codes since a caller-supplied date, plus each code's latest category snapshot and postmarket coverage. Defaults to Constat Radar's five-code watchlist and the last seven days. Analyst verdict text and internal review status are excluded; use next_since as the next polling cursor.

Last successful function declaration observed on 2026-08-13. Source: https://constat.dev/api/mcp. We list what the server declared; we do not call any of these functions.

Endpoint status observed on 2026-08-13. Source: https://constat.dev/api/mcp.

Signals

These are separate measurements of different things. They are deliberately not combined into one score, because a popularity number that mixes website traffic with saves and stars cannot be checked or acted on.

Where to get it

• Repository — https://github.com/healthai-hq/fda-risk-radar
• Website — https://constat.dev/connect
• Remote endpoint (streamable-http) — https://constat.dev/api/mcp

Also from healthai-hq

• Clarity by Health AI — repository gone
  Condition-aware ingredient and product checks for agents, with evidence tiers and citations.