Add Draconic intelligence to your software.

One conversation. Related updates.

Create a revocable key in Agent & API Access and keep it on your server, in an environment variable. Check current coverage before relying on a symbol or timeframe. Coverage and account-usage checks are free.

First request — requires your own key

curl https://api.draconic.ai/v1/agents/ask \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $DRACONIC_API_KEY" \
  -d '{"question":"Explain the current NIFTY50 structure and options context, including conflicting evidence and conditions that would change the view.","market":"nse","instruments":["NIFTY50"],"timeframe":"multi-timeframe"}'

Save the complete response. For a related follow-up, send the next question to the same endpoint with the returned chat_id. Replace RETURNED_CHAT_ID below. Each successful analysis uses one shared credit; running both examples uses two.

{
  "question": "Since the previous read, what changed? Preserve the evidence that challenges the view as well as the evidence supporting it.",
  "market": "nse",
  "instruments": [
    "NIFTY50"
  ],
  "timeframe": "multi-timeframe",
  "chat_id": "RETURNED_CHAT_ID"
}

A conversation per instrument and trading session is a useful default, not a requirement. The service uses up to twenty recent turns. Start a new conversation for an unrelated task and retain your own full history for longer comparisons.

Want a working starting point? The two-step Python market brief example asks for a NIFTY read, saves the complete answer, and challenges that read in the same conversation when you choose. It needs no extra Python packages and makes no automatic background calls. Both successful steps use two credits in total.

JSON outside. Market analysis inside.

The HTTP response is JSON. Its analysis field is readable Markdown, not a fixed numerical signal schema. Put the market task in question; your application handles code, parsing, storage and any execution. You do not need to ask Draconic to output JSON or explain its implementation.

analysis

Keep the complete text, including the observations, contradictions and “What changes the read”. Do not reduce it to a direction or the first few paragraphs.

chat_id / message_ids

Use chat_id for related follow-ups. Keep message identifiers, when present, for support and reconciliation. Identifiers belong to the same authenticated account.

coverage

Retain the market, instruments, timeframe and source_updated_at. Different sources can have different times; keep any additional timing stated in the analysis too. After-close data can be useful when clearly dated.

credits / chart

Handle these optional fields without crashing when absent. Chart availability varies; the analysis should remain usable without it.

If another AI summarizes the response, ask it to preserve source times, counter-evidence, missing data and the conditions that would change the view. Keep Draconic's original alongside that summary. If you require fixed fields, validate your downstream extraction and retain an unknown state; a missing number is not zero.

Keep the integration predictable.

• Send one intelligence request at a time per account. On 429, wait for the running request or follow the returned limit information. Do not create a rapid retry loop.
• On 401, check the key or sign-in. On 402, check the shared credit balance. On 403, check whether your account has the required access. On 404, check the instrument, market or conversation ownership. On 400 or 422, correct the request rather than retrying it unchanged.
• On a timeout or 5xx error, the outcome may be uncertain: a saved or charged request can outlive the client connection. Check account history and usage, or contact support with available identifiers before replaying it. Requests are not guaranteed to be idempotent, so a new request can incur another charge.
• Request observations and conditional scenarios, not a guaranteed daily range, unconditional reversal or buy/sell instruction. Missing source data cannot be supplied by stronger wording.
• API access does not read broker positions, saved profiles or presets, create alerts, place orders, or provide a raw market-data feed. Include relevant context in your question. Use your own application for scheduling and execution.
• Never place your key in a public page, shared collection or log. Treat all model output as data, not executable code. Review it before using it in a consequential workflow.