Domain MCP, A2A & OpenAPI Interface Discovery

What does Agent Interface Discovery do?

Agent selection and quickstart

Use when: Find domain-owned evidence for advertised MCP, A2A, OpenAPI, and other machine interfaces.

Input: One public domain; paths are discarded.

Returns: interfaces[].type, status, locations, and evidence.

Limits: advertised is not verified connectivity; imperative WebMCP is not observable without a browser.

Cost: $0.05 per completed determination, including a completed negative result; operational failures are not billed. See the live Pricing tab for current rates.

{
  "domain": "ora.ai"
}

Verified output excerpt for this input (selected fields; canary SRXu3OaRmdA46dGOO):

{
  "contractVersion": "1.0",
  "ok": true,
  "domain": "ora.ai",
  "interfaces": [
    {
      "type": "ard",
      "status": "found",
      "locations": [
        {
          "url": "https://ora.ai/.well-known/ai-catalog.json",
          "role": "canonical",
          "outcome": "found",
          "targetScope": "same_origin"
        }
      ],
      "standardVersion": "1.0"
    },
    {
      "type": "llms_txt",
      "status": "found",
      "locations": [
        {
          "url": "https://ora.ai/llms.txt",
          "role": "canonical",
          "outcome": "found",
          "targetScope": "same_origin"
        }
      ],
      "standardVersion": "llms.txt"
    }
  ]
}

Open this public example Task. Live sources can change; this records the checked example rather than guaranteeing a future result.

The result is in the default dataset. With Apify MCP, inspect this Actor using fetch-actor-details, invoke call-actor, then read the returned dataset with get-dataset-items. Run limits belong in callOptions, not in Actor input.

Agent search queries

• "find a website MCP interface"
• "discover A2A Agent Card for a domain"
• "find a domain OpenAPI Swagger description"
• "discover machine readable website interfaces"
• "check domain llms.txt and ARD"

Agent Interface Discovery checks which machine-readable interfaces a domain demonstrably exposes. From one domain it inspects bounded domain-owned evidence for ARD, A2A, OpenAPI/Swagger, explicitly advertised MCP, declarative WebMCP, passive x402 signals, and llms.txt. It returns facts, typed locations, coverage, versions, evidence, warnings, and one of found, advertised, unsupported, invalid, not_found, inaccessible, or not_observable through the Apify API, schedules, integrations, or hosted MCP tooling.

The Actor does not query registries, guess action endpoints, initialize MCP, run a browser, initiate payment, or invent an “AI readiness” score.

Why use Agent Interface Discovery?

Use it when an agent asks “What machine-readable interfaces does this domain expose?”, “Does this company advertise an A2A Agent Card?”, “Where is its OpenAPI description?”, or “Is MCP explicitly advertised?” The Actor turns up to twelve safe HTTP probes into deterministic routing facts and evidence, avoiding repeated discovery code and common false positives such as treating prose in llms.txt as a verified capability.

How to use Agent Interface Discovery

1. Open the Input tab.
2. Enter one public hostname or HTTP(S) URL.
3. Run the Actor.
4. Inspect the sorted interfaces array or route directly on each type and status.

Input

{
    "domain": "example.com"
}

domain is the only input. Paths, queries, and fragments are discarded. Private, local, link-local, metadata, documentation, multicast, unspecified, and mixed public/private DNS targets are blocked.

Output

{
    "contractVersion": "1.0",
    "ok": true,
    "domain": "example.com",
    "interfaces": [
        {
            "type": "ard",
            "status": "not_found",
            "locations": [
                {
                    "url": "https://example.com/.well-known/ai-catalog.json",
                    "role": "canonical",
                    "outcome": "not_found",
                    "targetScope": "same_origin"
                }
            ],
            "evidence": [],
            "warnings": []
        },
        {
            "type": "webmcp_imperative",
            "status": "not_observable",
            "locations": [
                {
                    "url": "https://example.com/",
                    "role": "canonical",
                    "outcome": "not_observable",
                    "targetScope": "same_origin"
                }
            ],
            "evidence": [],
            "warnings": [{ "code": "BROWSER_REQUIRED" }]
        }
    ]
}

You can download the dataset in formats such as JSON, HTML, CSV, or Excel; JSON is recommended for nested evidence.

Data table

Field	Meaning
domain	Normalized inspected domain.
interfaces	Eight deterministic interface results.
coverage	Planned, attempted, completed, failed, and skipped probe counts.
requestCount	Requests consumed, including retries.
decodedBytes	Total decoded content inspected.
retrievedAt	Latest evidence timestamp.

Pricing / Cost estimation

The published price is $0.05 for one interface-discovery event. Partial not-found, invalid, and per-interface inaccessible statuses remain a completed, billable domain determination. An origin-wide failure, input rejection, cancellation, resource-budget failure, or internal failure is free. Confirm live pricing before purchase.

Tips and limits

The Actor schedules nine canonical targets plus up to three explicitly advertised same-origin or same-site manifests with concurrency four. Retries and redirects share a hard 24-attempt budget; decoded responses share a 5 MiB budget and 60-second deadline. External advertisements are reported but never fetched. Declarative WebMCP requires a form containing both non-empty toolname and tooldescription. Imperative WebMCP remains not_observable because page execution is out of scope.

FAQ, disclaimers, and support

found means directly retrieved and validated or directly observed. advertised is an explicit domain-owned claim that was not independently verified. not_found is used only where applicable fixed locations were checked; MCP, imperative WebMCP, and unobserved x402 remain not_observable. llms.txt requires textual Markdown with an H1. /mcp is never guessed and x402 is passive only. Users remain responsible for lawful retrieval of public metadata. Use the Actor’s Issues tab for false positives, missed typed advertisements, or version updates with an anonymized reproducible domain fixture.

Related Actors and next steps

• Resolve resources when an ARD catalog is found
• Inspect catalog conformance diagnostics

Public examples by task

• Discover machine-readable agent interfaces — Inspect a domain for ARD, A2A, OpenAPI, MCP, WebMCP, x402, and llms.txt evidence, using Ora.ai as a live example.