Intelligence you can put your name on.

One engine. Your name on the front.

You are selling intelligence to your users, so the engine should answer as your product. Send a persona with any request and the engine presents itself by that name, in your tone, on behalf of your organization. What you rename is the voice. What you cannot rename is the conscience.

curl -X POST https://api.askasha.org/a2a/v1/message:send \
  -H "Authorization: Bearer $DNAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "asha",
    "message": {"role": "user", "parts": [{"text": "Can I take ibuprofen with lisinopril?"}]},
    "metadata": {
      "persona": {"name": "Nova", "organization": "Northwind Clinics", "tone": "warm"}
    }
  }'

Persona fields are validated, not trusted. A name that reads like an instruction ("ignore the rules") or a credential ("Dr. Nova", "Nova RN") is rejected and the request runs unbranded; the task record echoes the persona that was accepted. Credentials are yours to state in your product, never the engine's to imply. Deeper customization, a dedicated corpus boost, or a named clinical persona for your practice is an Enterprise conversation: hello@dnai.systems.

Ten domains, one protocol

Every domain runs the same engine: retrieval, evidence gating, falsification, then the answer. What differs by domain is the corpus, the refusal posture, and the disclaimers. The agent_id is a routing key: use the domain key (health, legal, ...); the legacy agent ids in parentheses stay valid forever. Your users never see either.

Depth. Production routes each question by clinical stakes to one of three architectures, Base (one retrieval-grounded pass), Mid (three-model merge), or Apex (multi-expert panel with a domain composer). Set metadata.depth to "base" to pin the fast path or "auto" to allow routing; every task reports depth_served. Free, Developer, and Pro keys run Base; auto is honored on Enterprise keys. The three depths are measured side by side on dnai.systems/benchmarks.

The live agent card lists every routing key, its capabilities, and its knowledge collections. It also lists a platform-service key used by DNAi's own tooling.

What is A2A?

A2A (Agent-to-Agent) is an open interoperability protocol for AI agents to discover and call each other. Your stack speaks A2A once and gets every domain engine an organization has decided to expose. DNAi runs A2A v1.0 in production at api.askasha.org: bearer auth, JSON in, JSON out, a task-based lifecycle you can re-fetch, stream, or cancel.

If you know the OpenAI Chat Completions or Anthropic Messages API, A2A will look familiar. The difference is on the other end: the engine carries its own corpus, its own falsification loop, and its own refusal posture, and it tells you when it does not know.

Quickstart

Three steps to a production call.

1. Provision an API key

curl -X POST https://api.askasha.org/api/a2a/signup \
  -H "Content-Type: application/json" \
  -d '{"email": "you@yourdomain.com", "tier": "free"}'

Returns a one-time api_key. Store it in a secret manager. Free Sandbox starts at 250 queries/month and 50 queries/day (20 RPM), no credit card required. Top up $10 Pay-As-You-Go via POST /api/a2a/topup to unlock 1,000 evaluation queries/month through January 2027, or subscribe to a monthly plan any time.

2. Call a domain, as yourself

curl -X POST https://api.askasha.org/a2a/v1/message:send \
  -H "Authorization: Bearer $DNAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "legal",
    "message": {"role": "user", "parts": [{"text": "My landlord kept my deposit in Texas. What are my options?"}]},
    "metadata": {
      "session_id": "user-8812",
      "depth": "base",
      "persona": {"name": "Casey", "organization": "Renters Union", "tone": "concise"}
    }
  }'

Swap legal for any routing key above. The response includes the answer, the evidence trail (source list and citations), token usage, and a task ID you can re-fetch through GET /a2a/v1/tasks/{id}. Reuse session_id to keep a conversation going.

3. Stream, retry, cancel, or falsify

POST /a2a/v1/message:stream returns server-sent events: retrieval and reasoning phases, text chunks, then the final task. List your tasks with GET /a2a/v1/tasks. Cancel one in flight with POST /a2a/v1/tasks/{id}:cancel. Stress-test any factual claim with POST /api/feng/falsify, the same falsification engine the domains run on themselves. It takes a signed-in DNAi account session; API keys are not accepted on that route.

Endpoints

Errors

Every error returns one JSON object. Branch on error.type: it names the cause and does not change. error.code follows JSON-RPC 2.0 and A2A v1.0 where a standard number exists, so related causes can share a code. error.param names the field or header at fault, error.doc_url links to its row below, and error.request_id matches the X-Request-ID header. Limits send Retry-After in seconds, also in error.retry_after. detail is the older field, kept as it was for existing clients. Send Accept: application/problem+json to get the same facts as RFC 9457 Problem Details.

{
  "error": {
    "code": -32061,
    "type": "daily_quota_exceeded",
    "message": "Daily query limit reached (50/50). Resets at midnight UTC. Top up $10 or upgrade tier for more.",
    "retry_after": 21600,
    "doc_url": "https://dnai.systems/a2a#err-daily_quota_exceeded",
    "request_id": "9b2e6c1a-4f3d-4a8e-b1c7-2d5e8f0a3c61"
  },
  "detail": "Daily query limit reached (50/50). Resets at midnight UTC. Top up $10 or upgrade tier for more."
}

A task that runs and fails still returns HTTP 200 with task.status.state: "FAILED" and the reason in task.status.message, per A2A. The OpenAI-compatible /v1 routes use the OpenAI error shape, with the same type value in error.code.

Retries and request IDs

An answer can take 10 to 60 seconds, so connections drop. Send an Idempotency-Key with message:send and retry with the same key: the first call runs, and any retry within 24 hours gets the saved response with Idempotent-Replayed: true, without running again or using your allowance again.

KEY=$(uuidgen)  # reuse $KEY when you retry this request
curl -X POST https://api.askasha.org/a2a/v1/message:send \
  -H "Authorization: Bearer $DNAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $KEY" \
  -d '{"agent_id":"asha","message":{"role":"user","parts":[{"text":"What is sepsis?"}]}}'

• Keys are 1 to 255 characters. Use a UUID v4 per question; never put personal data in a key.
• Same key with a different body: 400 idempotency_key_reused. Same key while the first call runs: 409 idempotency_in_progress.
• Nothing is saved when a call fails auth, limits, or validation, so you can fix it and retry with the same key.
• Idempotency-Status: unavailable means the retry store was unreachable and the call ran without protection.
• Streams take no key. After a dropped stream, read the result with GET /a2a/v1/tasks/{id} using the id from task_created.
• Every response carries X-Request-ID, repeated as error.request_id on errors. Send your own to match your logs. Quote it when you write to support.

Changelog

Newest first. Every /a2a/v1 response carries A2A-Version: 1.0. Additive changes ship without a version change, so clients must ignore fields they do not know. Behavior changes say what to check.

Pricing

Start free with 250 queries/month, top up $10 Pay-As-You-Go to unlock 1,000 evaluation queries/month through January 2027, or pick a monthly plan. One key covers all ten domains. Enterprise adds a HIPAA Business Associate Agreement and dedicated capacity.

Token & Depth Metering Schedule

Every standard query includes up to 2,000 input and 2,000 output tokens (roughly 8,000 characters). Long-context documents or extended syntheses above 2,000 tokens are metered at Corti-parity token rates ($4.00 / 1M input, $16.00 / 1M output) while including 144M-passage clinical grounding and PMID verification.

Clinical Performance & Competitor Comparison

Measured on peer-reviewed clinical rubrics and pre-registered rare-disease red-team evaluations (dnai.systems/benchmarks · GET /api/benchmarks).

Spending caps are hard limits. We will not silently exceed them. If you hit your cap mid-month, the API returns HTTP 429 with an upgrade or top-up hint until the next cycle or until you raise the cap. Want one key per domain with usage split out on your invoice? Ask us; volume pricing and BAAs go through the same address.

Fiduciary Contract

What a DNAi domain engine gives you that a base-model API does not.

What we can show you

• The method is patented. US Patent 12,555,008 B1, granted 17 February 2026: evidence-gated cognition with a cryptographic audit trail. Details at dnai.systems/patents.
• The corpus is committed. A cryptographic commitment to the knowledge corpus is available to customers, so you can verify that the corpus your answers came from has not changed underneath you.
• The health engine is benchmarked in the open. DNAi Health in Apex mode scored 65.5% on HealthBench Professional (official length-adjusted metric, September 2026) on our harness with one grader we chose, applied identically to every generally available frontier model we compared. Research previews and vendor self-reports on other harnesses are a different table. The register with confidence intervals is available on request; we publish method and interval together or not at all.
• The safety floor is measured, not asserted. Crisis protocols, refusal under uncertainty, and the AI disclosure are part of every domain and cannot be switched off by a persona.

Live status

Your browser calls each public endpoint directly; nothing below passes through us on the way to you. Re-check

• waitingAPI livenessapi.askasha.org/health/live
• waitingA2A fleet health (11 agents)api.askasha.org/a2a/v1/health
• waitingAgent cardapi.askasha.org/.well-known/agent-card.json
• waitingPublic catalogapi.askasha.org/api/agents/
• waitingCorpus cryptographic commitmentapi.askasha.org/api/audit/corpus-state/public
• waitingBenchmark registerapi.askasha.org/api/benchmarks
• waitingAPI alias on dnai.systemsapi.dnai.systems/health/live

Checks start when the page loads.

Reference

• Live agent card · api.askasha.org/.well-known/agent-card.json
• Health probe · api.askasha.org/a2a/v1/health
• Public catalog · api.askasha.org/api/agents/
• API reference · dnai.systems/api-reference
• OpenAPI 3.1 · api.askasha.org/openapi.yaml (also .json)
• A2A protocol spec · a2a-protocol.org
• Engineering whitepaper · How DNAi Thinks
• Talk to a human · hello@dnai.systems