Skip to content
    Build on Ansvar · evidence infrastructure

    One endpoint that cites every row.

    Point your agent or your product at the gateway. Results carry source URL, publisher and licence; reports carry a receipt you can verify later. The contracts are published and versioned, and your data leaves with you.

    # any MCP client
    url      https://gateway.ansvar.eu/mcp
    auth     OAuth 2.1, dynamic client registration
    ask      "Using Ansvar, which DORA articles apply to ICT third-party risk?"
    returns  rows · source_url · publisher · license · receipt
    The contracts

    What a row promises, and what a report promises.

    Two published contracts, both versioned. Break either and the gateway fails closed rather than serving a thinner answer.

    Serving contract

    Every served row carries a citation triple: the original publisher's URL, the publisher, and a licence code from the published catalogue. A lookup hint on a search row re-fetches the same row with the same triple. Rows that cannot carry the triple are withheld with a visible reason.

    Read the contract in the docs
    "_citation": {
      "source_url": "https://data.europa.eu/eli/reg/2016/679/oj",
      "publisher":  "Publications Office of the European Union",
      "license":    "EUR-Lex-Decision-2011-833",
      "lookup":     { "tool": "get_provision",
                      "args": { "source_id": "eu-regulations",
                                "canonical_ref": "GDPR:art_17" } }
    }

    Delivery receipt

    Every report response carries a server-authored receipt: the executive summary verbatim, integrity metrics computed or marked unsupported, every attention item uncapped, artifact links. Your agent relays it unchanged, so what the reader sees is what the server produced.

    See it on an assessment
    summaryVerbatim executive summaryserver-authored
    metricsComputed, or marked unsupported. Never zero for a metric the report does not carry.typed
    attentionEvery item, uncapped, verbatimcomplete
    artifactsReport links with content hasheshashed

    Verification contract

    Send back the references a draft relies on and verify_citations re-retrieves each one under your own scope, checks the producer-declared identity of the row it got, and compares your quotation with the body. One verdict per subject — quote_checked and reference_resolved for the two ways a subject can hold, then quote_mismatch, no_quotation_submitted, identity_mismatch, identity_absent, reference_unparseable, not_served, withheld and unavailable — and coverage is advertised rather than assumed, so a subject the identity seam does not cover refuses instead of passing. Available on every plan; signed attestations from Premium.

    See a verified set
    verify_citations(subjects=[
      { "subject_id": "s1",
        "source_id":  "eu-regulations",
        "reference":  "GDPR:art_17",
        "quotation":  "The data subject shall have
                       the right to obtain ..." }
    ])
    -> { "subject_id": "s1", "verdict": "quote_checked" }
    Query receipt
    What was served. On Company: a signed, tamper-evident record of what was asked, what came back and when — read it with get_receipt, list with list_receipts, check it with verify_receipt.
    Delivery receipt
    What the reader was handed. The block above: server-authored, relayed unchanged with a report, so the summary, the metrics and the attention items cannot be trimmed on the way to a person.
    Audit ledger
    What an auditor can verify without us. On Company: per-tenant receipts in a cryptographic ledger, exported as an offline-verifiable package and decrypted client-side with your own key.
    Running headless

    Agents on seats, credentials you rotate.

    1. Service credentials

      From Premium, create org-owned credentials for n8n, CI jobs and scheduled monitors. Each takes a seat with the same tool surface as a signed-in person. Rotate and revoke any time; revoking works on every plan.

    2. Published limits

      Calls per day, concurrency and requests per minute are published per plan. A refusal returns your usage, the reset time and the upgrade path in a shape your agent can act on.

    3. Skills and connectors

      Compiled, checksummed workflow skills for Claude, Copilot and Cursor; the Microsoft 365 agent; the Gemini CLI extension. Every example question leads with Using Ansvar.

    4. Export guarantee

      Audit packages, OpenVEX, git-native ISMS, a local architecture workspace. What you built on Ansvar leaves with you.

    Which route applies to you

    Automate internally, deliver to clients, or embed in what you sell.

    Three different permissions, and the difference matters before you plan a dependency. The terms are the authority; this is the short read.

    Internal automation
    Included in your plan. Your own agents, CI jobs and monitors call the gateway for your own business. Seats and credentials are the only thing to size.
    Client work
    Included, within section 8a. You may produce deliverables for your clients on your own responsibility. What you may not do is hand a client the content itself — a feed, an extract series or a recurring compilation.
    Embedding and reselling
    By written agreement. Putting Ansvar inside a product you sell, or giving third parties outputs on demand, needs a partner agreement first. Ask before you build the dependency, not after.
    The binding text is sections 8a, 12 and 13 of the terms. For an embedding or reseller route, start with a partner enquiry.
    One integration, measured

    What it looks like under someone else’s agent.

    A European legal-expenses insurer is evaluating Ansvar underneath its own copilot. The shape below is measured from the gateway’s own telemetry over thirty days, not from a questionnaire.

    They kept their stack. We are the layer underneath.

    Two channels, both on service credentials: an agent built in Microsoft Copilot Studio on the insurer’s own Power Platform identity, and the team’s own Python harness. Neither runs on our interface, and neither sends us a model.

    The workload is national law and case law — court decisions, disciplinary rulings, statutes and the preparatory works behind them. About half the calls are exact-reference lookups of a named article or decision; the other half are searches that fan out across the jurisdiction. Legislative history arrives inside those searches rather than through a lookup of its own.

    In thirty days neither channel hit a quota, a rate limit or a licence refusal, including one unattended run of roughly a thousand calls inside an hour. Where a decision corpus was withdrawn pending a privacy assessment, the questions that needed it came back empty and said so, rather than being filled in from somewhere else — visible in their logs as it happened.

    The limits that run underneath it
    clientsCopilot Studio agent · the team’s own Python harnesstheir stack
    authTwo service credentials, each on its own seatheadless
    callsRoughly half exact lookups, half fan-out searchesmeasured
    refusalsNo quota, rate-limit or licence refusal in the windownone
    gapsA withdrawn corpus answered empty, not approximatelyvisible
    Why not build it yourself

    The part that never stops: keeping the sources true.

    Retrieval is a weekend. What follows is the standing cost, and it is the work Ansvar takes off your roadmap.

    01

    Source integrations

    Every publisher has its own site, its own bulk feed, its own quirks — and changes them without telling you. We hold the fetchers and the parsers.

    02

    Licensing and provenance

    Each source is cleared for commercial reuse before a single fetch, and every served row carries its publisher, URL and licence code.

    03

    Coverage and freshness

    What is in force, what changed, when each corpus was last refreshed — tracked per source, with the age published rather than assumed.

    04

    Verification machinery

    Citation round-trips, content hashes, withheld-row reasons, refusals that say why. The plumbing that makes an answer checkable.

    What a production month costs

    Seats, credentials, runs. Nothing metered in the dark.

    An agent is a seat like a person is a seat, and every limit that shapes the bill is published.

    Premium, per seat
    Each holder is a seat, person or agent, with the same tools and concurrency. Runs past the monthly allowance are never billed — the ceiling refuses instead.
    Team, per workspace
    One price covers three contributors, their pooled quota and five service credentials. Extra credentials take extra seats.
    The runs
    Team's pooled assessment allowance is the only metered line: past it a run is priced, and the hard stop refuses rather than billing on.

    Worked totals for a busy workspace, and every quota the gateway enforces, are on pricing and limits.

    Connect in two minutes.

    Free for businesses. OAuth in the client you already run.