ATA Standalone

Personal Agent-to-Agent protocol. Zero dependencies. Any LLM.

Send a task to a remote AI agent, get the result back. No platform, no registration, no cloud required.

You (Client)                    Your friend (Server)
ata-client.js  ──HTTP──────>   ata-server.js
                               └── LLM processes task
               <──result────   └── POST callback

Quick Start

# Clone & configure
git clone https://github.com/JackyCufe/ata-standalone
cd ata-standalone
cp .env.example .env
# Edit .env: set ATA_SHARED_SECRET and ATA_LLM_* to match your LLM

# Terminal 1: start your agent server
node ata-server.js

# Terminal 2: send a task
node ata-client.js \
  --to http://localhost:3740/ata/v1 \
  --task '{"action":"ask_agent","content":"用一句话解释量子力学"}' \
  --secret your-shared-secret

The result comes back automatically once the LLM finishes:

{
  "status": "completed",
  "result": {
    "answer": "量子力学是描述微观粒子行为的物理理论...",
    "model": "claude-sonnet-4-6",
    "action": "ask_agent"
  }
}

How It Works

1. Client fetches the server's Agent Card (GET /ata/v1/agent-card) to discover capabilities
2. Client signs a task request with HMAC-SHA256 and POSTs it (POST /ata/v1/task)
3. Server verifies the signature, stores the task, returns HTTP 202 immediately
4. Server routes the task to the executor:
   • ping / echo → built-in handlers (instant, no LLM)
   • everything else → calls your LLM via OpenAI-compatible API
5. Executor POSTs the result to the callback URL
6. Client polls /ata/v1/task/:id/status until status: completed

LLM Backends

Any OpenAI-compatible endpoint works. Set in .env:

Backend	ATA_LLM_BASE_URL	ATA_LLM_API_KEY	ATA_LLM_MODEL
LiteLLM (local)	http://localhost:4000/v1	no-key	claude-sonnet-4-6
Ollama (local, free)	http://localhost:11434/v1	ollama	llama3
OpenAI	https://api.openai.com/v1	sk-...	gpt-4o

Custom Handlers

Register your own action handlers (they take priority over the LLM):

const { registerHandler } = require('./lib/executor');

// Fast, deterministic handlers skip the LLM entirely
registerHandler('get_weather', async (task) => {
  const data = await fetchWeather(task.payload.city);
  return { result: data };
});

registerHandler('review_code', async (task) => {
  // Or call your own LLM with a custom prompt
  const review = await myCustomLLM(task.payload.code);
  return { result: { review } };
});

Security

• HMAC-SHA256 signing: every request is signed with a shared secret
• Idempotency: duplicate task IDs are rejected (prevents replay attacks)
• No central registry: you only talk to agents whose URL you know
• No platform: your data never touches a third-party server

Protocol

POST /ata/v1/task
{
  "from": "agent://alice/main",
  "to": "agent://bob/assistant",
  "taskId": "<uuid>",
  "type": "task_request",
  "payload": { "action": "ask_agent", "content": "..." },
  "callbackUrl": "http://alice.example.com/ata/v1/callback/<uuid>",
  "timestamp": 1234567890,
  "signature": "<hmac-sha256>"
}

Endpoints

Method	Path	Description
GET	/ata/v1/agent-card	Agent capabilities & identity
POST	/ata/v1/task	Submit a task
POST	/ata/v1/callback/:taskId	Receive result (for servers)
GET	/ata/v1/task/:taskId/status	Poll task status
GET	/health	Health check

Comparison

	ATA Standalone	Google A2A	Society Protocol
Zero dependencies	✅	❌ SDK required	❌ libp2p
Personal use	✅	❌ Enterprise	🔶
Any LLM backend	✅	❌	❌
Task callback guarantee	✅	✅	❌ broadcast only
One command startup	✅	❌	🔶

License

MIT