For agents
An agent reads Attify three ways: the read-only API at /api/v1, the MCP server at /mcp, and llms.txt. Every answer is typed JSON with a versioned schema, and every published score carries a signed attestation anyone can check.
The read-only API
- GET /api/v1/vendors
- GET /api/v1/vendors/{id}
- GET /api/v1/attestations/{id}
- POST /api/v1/verify
- GET /api/v1/keys
- GET /api/v1/sample/{verdict|findings}
- GET /api/v1/jobs[/{id}]
- GET /api/v1/standards[/{id}]
- POST /api/v1/ask
- GET /api/v1/index
- GET /api/v1/index/editions[/{YYYY-MM}]
- GET /api/v1/definitions
The OpenAPI document of exactly these paths: /api/v1/openapi.json. Errors share one shape: {"error": {"code", "message", "status", "details"}}.
The MCP server
Streamable HTTP at POST https://attify.app/mcp, one server for Claude and ChatGPT. Every tool is read-only.
The tools that read your own organization need sign-in: OAuth 2.1 with PKCE (S256), found from /.well-known/oauth-protected-resource/mcp. An app registers by its Client ID Metadata Document or dynamically; its token is for https://attify.app/mcp only, scope attify.read, and only reads. Called without one, such a tool answers 401 with WWW-Authenticate.
- get_verdict
- The answer first: one verdict sentence, then claimed, audited and the gap with 95% intervals, the evidence tier, provenance (connected, witnessed or uploaded; for witnessed data whether the customer reconciled it), judge validation, the drivers of the gap, the period, sample sizes and the data's scope. For a published vendor (by slug) or the public sample audit.
- list_findings
- The public sample audit's findings: each a template sentence over its numbers, with the evidence tier, sample size and the evidence ids it rests on. Filter by kind: gap_driver, contact_reason, evidence_tier, judge_validation, provenance, window_sensitivity, no_request, handoff_outcome, claim_rule, rule_choice. A vendor's published score lists its drivers only (it quotes no conversation).
- get_evidence
- One evidence record of the public sample by its id (ev_…): the outcome, the failed check, the evidence tier, the redacted excerpt, quoted as customer text, and why it failed.
- compare_subjects
- Two to five published vendors of one job type side by side: audited against claimed with 95% intervals, the gap, evidence tier, provenance, period and, for support resolution, always each one's contact-reason mix. Vendors of different job types or standard versions are refused, and differences inside the intervals are not distinguishable.
- explain_check
- What one check of a job type's standard asks, which question it answers (did it happen, did it stick, did a human have to redo it) and which evidence tier decides it. Checks by job type: support-resolution: "repeat_contact", "handoff", "judge_unresolved"; sales-meetings: "not_ai", "not_new", "stepped_in", "cancelled", "no_show", "not_sales"; bookings: "not_ai", "not_new", "staff_fixed", "cancelled", "no_show", "wrong_booking".
- get_standard
- A job type's published standard: number, title, version, status, how the claim is reconstructed, the checks, the evidence tiers, the changelog and the full text's address.
- ask_attify
- A plain-language question about the public sample audit or a published vendor's score, answered from that data only: numbers copied with their intervals and tier, every claim cited (verdict fields, finding ids, evidence ids), "not in the data" instead of a guess. Requests for advice on raising a score, and for private data (brand names, source ids), are refused.
- get_index
- What AI agents achieve, by job type: the pooled audited share of the outcomes they claimed, with its 95% interval (each company a cluster), evidence tier, provenance mix, failure modes and monthly trend, from companies that contribute their anonymous data. Published for a job type only with enough contributing companies and none above half the claims. Names no company, vendor or agent. Give job_type for one job.
- define_term
- What a term of AI agent verification means (resolution rate, handoff, containment, meetings booked, show rate, evidence tier, provenance...): Attify's glossary definition, with how vendors' own definitions differ where they do, and each standard's own definition in its own words, with the anchor to cite. Nothing else is guessed: an unknown term says so.
- lookup_vendor
- A vendor's published Attify verification: status (Claim confirmed · within 5.0 points when Attify is 95% sure the vendor's count overstates audited outcomes by at most the claim tolerance, the gap's upper 95% limit being within it; Claim not confirmed; Lapsed; or Revoked), audited rate against the vendor's own claimed rate with 95% intervals, the gap, evidence tier, provenance (connected, or witnessed with how many counted customers' data the vendor streamed and they reconciled), the contact-reason mix (support resolution), rank within its job type, period, standard version and the signed attestation. Give the profile slug, or a name to search. Rates compare only within one job type, and for support resolution next to the contact-reason mix.
- verify_attestation
- Checks an attestation's Ed25519 signature against Attify's published keys, then whether it is still current: a vendor verification (not superseded, expired or revoked) or a traction report (not expired or withdrawn). Give the attestation JSON as published, or a vendor attestation's id.
- explain_provenance
- What each provenance level means and what it can support: connected (Attify pulled every record from the customer's own system), witnessed (the vendor's agent streamed each interaction to Attify as it happened, signed and timestamped; it counts toward a verification only once the customer consented and reconciled it each month) and uploaded (a file, self-reported, never a verification). Also how an agent on its own platform sends witnessed data: the Ingest API, the Python and TypeScript SDKs, or OpenTelemetry, with the endpoints, schema and docs. Give a level to explain one.
- get_traction_report
- A company's audited traction report, from a share link it gave you: the summary and auditor-style opinion, up to three key numbers, each claim tested (supported, not supported, not testable), then usage, retention, pilot to production, concentration and outcome rates per customer-month, every number with its sample size, period, evidence tier and provenance (connected, or witnessed data the customer reconciled; never uploads), and the signed attestation. Opening it is logged in the company's view log. Links limited to named people need a browser sign-in.
- get_my_account
- Who is signed in, in which organization, and what kind it is: a vendor of an AI agent, a company that buys AI agents, or an investor. Says which signed-in tools answer for it. Needs sign-in to Attify; reads only.
- get_monthly_verdict
- Your company's verdict for a month on each AI agent it pays for, as emailed on the 1st: the verdict sentence, the audited share of the vendor's claimed outcomes with its 95% interval, the cost per audited outcome, the change since the month before, the evidence tier, provenance and standard version. Last month by default. For companies that buy AI agents. Needs sign-in to Attify; reads only.
- get_agent_outcomes
- One month of your company's AI agents, answer first: for each agent the verdict sentence, claimed against audited outcomes with the 95% interval, what the month cost, the cost per audited outcome against the price per claimed one, and where it fails: each failure mode's share of the claims, with suggestions for your own deployment and redacted examples, quoted as customer text. For companies that buy AI agents; what a plan does not include says so. Needs sign-in to Attify; reads only.
- compare_my_agents
- Your company's AI agents of one job type side by side for a month: each one's audited share of claims with its 95% interval, cost per audited outcome and main failure mode, and the pooled rate of other contributing companies where it can be shown (anonymous, each company a cluster). Differences inside the intervals are not distinguishable. Side by side comes with the Business plan. Needs sign-in to Attify; reads only.
- get_renewal_packs
- The renewal packs your company signed for vendor negotiations: one agent and month each, the audited numbers it carries (claimed, audited, the audited share with its interval, cost per audited outcome), when it expires, and its signed attestation (check it with verify_attestation). Reading only: a new pack is signed in the app. Needs sign-in to Attify; reads only.
- get_program_status
- Your verification programs, as a vendor: phase, public status (Claim confirmed · within 5.0 points, Claim not confirmed, Lapsed or Revoked, or private), the one next step, each publishing minimum met or missing in your own numbers, what the latest score still waits for before it publishes, the claimed unit as Attify recorded it, and your customers' connection and consent. It states where the program stands; it never advises how to raise a score. Needs sign-in to Attify; reads only.
- get_badge_embed
- The code for your live Attify badge, once a program's score is public: one line for a website (a plain link to your profile, then the script that draws the current badge and its hover card), the image for email signatures and READMEs, and the badge's addresses. Before that, what publishing still needs Needs sign-in to Attify; reads only.
- list_traction_reports
- Your company's audited traction reports: status, summary, the months and customers covered; for one report, its share links (who each is for, state, expiry, email allowlist, views) and who opened it, from the view log. A link's address is shown once when it is made and never stored, so it cannot be read back. Needs sign-in to Attify; reads only.
- get_ingest_health
- Health of your deployments that send witnessed data (the Ingest API, an SDK, OpenTelemetry, or read from your own agent platform account): healthy, degraded or no data and why, events and interactions in the scoring period, late interactions (never scored as witnessed), sequence gaps and breaks, the signed share, the last event, volume by month, and each customer's consent and reconciliation. Needs sign-in to Attify; reads only.
Verifying an attestation
POST it to /api/v1/verify, or check it offline against the keys at /.well-known/proven-keys.json: the signed bytes are the payload as canonical JSON (sorted keys, no whitespace, UTF-8), signed with Ed25519. A person can paste one at /verify.