AI agents

Your AI agent wakes up only when the market does

Polling prices and computing indicators all day is expensive work for an agent. Subtrigger does the watching and calls your agent with the event, the values and your instructions.

Subtrigger is the market-watching layer for AI agents. It reads live prices from seven crypto exchanges and every US-listed stock and ETF, computes indicators, evaluates your conditions, and calls your agent only when one of them becomes true, with the event, the values and your instructions in the request.

Why agents should not watch markets themselves

An agent that wants to react to "BTC dropped 3 % in ten minutes" or "RSI on NVDA crossed 30" has two choices: keep a loop running that polls exchanges, stores bars and computes indicators, or let something else do it and wake the agent up.

Polling inside an agent is the expensive choice:

Watching inside the agentWith Subtrigger
A loop polling prices every few seconds, around the clockNo loop: the agent is called on events
Tokens spent reading prices that did not change anythingTokens spent only when a condition is true
Candles, warm-up and indicator math to maintain12 indicators computed live and tested against reference values
Rate limits and outages of seven exchanges to handleRetries, back-off and reconnects handled for you
A crash means missed eventsEvents are recorded, with every delivery and the agent's answer

The agent keeps doing what it is good at, deciding, explaining and acting, and stops being a cron job.

How a call looks

You create an alert like any other, with conditions over prices and indicators, then choose an AI agent as where it goes. When the alert fires, Subtrigger sends:

  • A text part: your instructions for this sink, followed by the alert's message (written by you, with live values).
  • A data part: the full event as JSON: alert id and name, the stream that fired, the latest value of every input, each comparison with its result.
{
  "alert": "…",
  "name": "BTC dip inside uptrend",
  "fired_at": 1794750979080,
  "input": "btcusdt_close",
  "inputs": {
    "btcusdt_close": { "stream": "atom:binance.miniticker:BTCUSDT:close", "value": 64812.4 },
    "btcusdt_value": { "stream": "ind:rsi_14_1m@binance.miniticker:BTCUSDT:value", "value": 28.6 }
  },
  "conditions": [ { "left": { "input": "btcusdt_value", "value": 28.6 }, "op": "lt", "right": { "value": 30 }, "result": true } ]
}

With A2A, Subtrigger reads your agent card at /.well-known/agent-card.json, finds the JSON-RPC endpoint and sends SendMessage (A2A 1.0) or message/send (0.3). With HTTP, it posts {instructions, message, format, event, conversation_id} to your URL. The agent's reply, or the state of the task it started, is saved in the delivery history.

Conversations that remember

Choose one conversation per alert and every event of that alert arrives in the same A2A context (contextId), so the agent can see what it concluded last time ("third dip this morning, the previous two recovered"). Choose new every time for independent tasks. Subtrigger stores no conversation itself: memory is the agent's.

Let the agent manage alerts too

Give the agent an API key and it can use the Subtrigger MCP server: search streams, read live values and alert history, create and pause alerts. A common pattern: the agent tightens or loosens its own triggers after each event.

Alert idea

BTCUSDT ROC 10, % · 1m · Binance > 2

Create this alert

What you can trigger on

  • Prices of every spot pair on Binance, Bybit, OKX, Bitget, Gate, KuCoin and MEXC, and of every US-listed stock and ETF.
  • 12 indicators: moving averages, RSI, MACD, Bollinger Bands, ATR, stochastic, rate of change, volatility, z-score, Donchian channels and the spread between exchanges.
  • Arithmetic on the compared stream: price on Binance > price on Bybit + 200.
  • Any number of conditions joined with and / or, fired on crossing, on every match or once, with a cooldown.

Safety by default

  • Requests go only to public https addresses; private networks, cloud metadata addresses and redirects are refused.
  • Authorization headers you add are stored as secrets and never shown again.
  • One attempt and a 60-second timeout by default, so a slow agent is not called twice for the same event.

FAQ

What is the difference between an AI agent sink and a webhook?

Both send an HTTP request. The agent sink speaks agent protocols (A2A), carries your instructions with every event, keeps a conversation per alert and records the agent's answer. A webhook sends the event or the message as-is, signed, for any code.

Which agent frameworks work?

Any agent that serves A2A (Google ADK, LangGraph, CrewAI, Pydantic AI and others have A2A support) or any HTTP endpoint you control, including n8n and serverless functions.

Does Subtrigger run my agent or call an LLM?

No. It only calls your agent. Models, tools and costs stay on your side.

Can the agent create alerts by itself?

Yes, through the MCP server with an API key that has write access.

Is it free?

Calling AI agents is part of Pro, which is free during the pilot: nothing to pay and no card to add.

Your AI agent wakes up only when the market does

Free to start, Pro free during the pilot. No card, no exchange keys.

Create an alert