Veritap
OfficialUses Wikidata as a data source to verify claims about businesses, listings, objects, and places, providing evidence bundles with sources and confidence scores.
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., "@VeritapCan you verify this restaurant listing is accurate?"
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.
Veritap
An MCP endpoint where agents check whether a real-world fact can be verified — before acting on it.
Live: https://veritap.dev/mcp · npm: veritap-mcp · MCP registry: dev.veritap/veritap
{ "mcpServers": { "veritap": { "command": "npx", "args": ["-y", "veritap-mcp"] } } }Free, no auth, read-only, and it never returns a bare failure. Covers businesses, listings, objects and places. Claims about individuals are refused as a matter of policy.
What works today
Business-exists claims are answered on the spot -- an automated desk check across GLEIF LEI records, SEC EDGAR filings, Wikidata, and website liveness returns a dated evidence bundle with sources and an honest confidence. Free while verification is new. Never a bare verdict.
Triage (
triage_unknowns): batch every uncertainty in a task; each is classified as answerable-yourself (with a suggested source), verifiable here, or not determinable (with planning advice). Items you can settle yourself are routed away, not sold to you.Plan mapping (
plan_verification): which steps of a multi-step plan rest on checkable real-world facts.Feasibility checks (
check_feasibilityand friends): call speculatively before relying on, buying from, or citing anything seen online.
Answers repeat instantly from a shared attestation cache. Nothing is charged; paid verification is not open.
Related MCP server: Lenz MCP Server
What this actually is
A demand sensor. The product is the ledger, not the verification.
Agents constantly hit steps that depend on facts they cannot confirm, and today they either guess or silently drop the step — so the demand is invisible. Every call here writes a structured row: what was asked, what it was worth to the caller, and whether anything could answer it. Requests we cannot fulfil are the most valuable rows in the database.
That dataset cannot be backfilled, which is why the endpoint went live before fulfillment existed.
Design decisions worth knowing
No human in the loop, ever. No dispatch, no verifier network. If the copy ever implies otherwise,
scripts/a3-audit.mjsfails the build.No outbound referrals. A miss keeps its signal in-house, so it must stay decision-grade on its own: closest alternatives, honest self-help, and a real promise to report back (
src/notify.ts).Quarantine, don't drop. Suspected flooding tags rows rather than refusing them. A wrongly-throttled agent prunes us from its tool rotation permanently; a wrongly-tagged row costs nothing and can be un-flagged later.
Claims about people are refused, not deferred. FCRA exposure and the obvious harassment vector. Refusals are logged without their text.
scripts/policy-check.tsis a two-sided control battery — a gate that refuses nothing and a gate that is switched off look identical from the outside.The fingerprint is not an identity. MCP has no installation id, so components are stored separately rather than baking a guess into a hash.
Layout
src/index.ts Worker entry, MCP protocol layer, cron
src/router.ts the core: normalize -> hash -> cache -> catalog -> ledger
src/policy.ts people-claim exclusion (hard refusal)
src/triage.ts unknown classification, routes web-answerable items away
src/notify.ts the return path: report back when a gap closes
src/ledger.ts append-only demand ledger, batched writes
src/admin.ts operator dashboard, Cloudflare Access protected
public/ agent-legible surface: llms.txt, openapi, well-known, pages
docs/decisions.md every deviation from spec, with reasoningDevelop
npm install
npx wrangler d1 migrations apply groundtruth --local
npx wrangler devGates that must pass before any deploy:
npx tsc --noEmit
node scripts/a3-audit.mjs # honesty: no dispatch, no referral, no purchasability
node scripts/policy-check.ts # people-claim gate, both directions
./scripts/history-scan.sh # secrets across git HISTORY, before any public pushhistory-scan.sh exists because a working-tree audit cannot catch what a
public push actually exposes. Grepping git ls-files only ever sees the
current checkout — it passes happily while earlier commits, and commit
messages, still carry the thing you removed.
Secrets are never committed. ALERT_EMAIL and NTFY_TOPIC are set with
wrangler secret put — the ntfy topic in particular is a capability, since
anyone holding it can both read the digest and publish alerts to the operator.
Stack
Cloudflare Workers (stateless createMcpHandler — no Durable Objects), D1,
@modelcontextprotocol/server v2. A frozen project costs $0 to leave running,
which is the point: the sensor keeps accruing whether or not anyone is watching.
MIT
This server cannot be deployed
Maintenance
Related MCP Connectors
Check whether a real-world fact can be verified before an agent acts on it. Free, no auth.
Check claims against a fact-store: consistent, contradicts, or unverifiable — with a receipt.
Tamper-evident proof creation and verification for AI agents via MCP, A2A, and REST.
Experimental MCP server for current empirical verification of explicit public HTTPS endpoint claims.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables agents to verify claims with evidence-based truth scores and confidence levels by running a deterministic pipeline of evidence lanes and adversarial checks.26MIT

Lenz MCP Serverofficial
AlicenseNot gradedqualityAmaintenanceFact-check claims against independent sources via a multi-model pipeline, providing verdicts and confidence levels from any MCP client.Apache 2.0- AlicenseAqualityCmaintenanceRead-only MCP server exposing a W3C PROV knowledge graph of verified facts with provenance, enabling AI agents to list, search, and check facts while enforcing that writes remain CLI-only.5MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to cross-verify candidate claims against caller-supplied source texts, flagging hallucinations, numerical drift, entity mismatches, contradictions, and unverified assertions. It returns sentence-level verdicts with matched evidence snippets and machine-readable factual grounding confidence scores, exposed over MCP stdio, HTTP REST, and A2A discovery routes.MIT