Axion

Axion forecasting

Axion answers a question by running an orchestrator and a team of research agents (web search, market data, SEC filings, Fermi estimates) that cite their evidence and return graded probabilistic forecasts.

When to use

Use Axion when the user wants the odds of a future event, a deal or scenario evaluated, or a research-backed estimate: "how likely is X", "forecast whether Y", "what are the odds of Z", elections, markets, deals, geopolitics, product launches.

Setup

1. Create an account at https://axion.eternis.ai and create an API key on the API Keys page. Any account can create keys.
2. Add credits. API usage is prepaid and billed separately from the monthly plan; new accounts start at zero.
3. Provide the key as the AXION_API_KEY environment variable.

Before calling, confirm the key is present. If AXION_API_KEY is empty, stop and ask the user to set it. Do not proceed.

Procedure

Forecasts run asynchronously: start one, then poll until it reaches a terminal status. Base URL https://api.axion.eternis.ai; the key is a Bearer token on every request.

1. Start a forecast:

curl -s https://api.axion.eternis.ai/forecasts \
  -H "Authorization: Bearer $AXION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "Will the Fed cut rates in June 2026?", "effort": "high"}'

Returns {"id": "<token>", "status": "starting"} — id is an opaque token (e.g. L2d5qNdtxXjyjXGKTmKVMX), no fixed prefix. Use it to poll. Set effort low/medium/high and max_forecasts 1–10 to widen the run.

2. Poll until status is completed or failed, backing off a few seconds between calls:

curl -s https://api.axion.eternis.ai/forecasts/<id> \
  -H "Authorization: Bearer $AXION_API_KEY"

3. Read forecasts[]. Each carries forecast_text, probability (0–1), confidence_lower/confidence_upper, resolution_date, and reasoning. Present the probability and the reasoning behind it.

Full request and response fields, webhooks, and credit rates: see references/axion-api.md, mirrored from https://axion.eternis.ai/docs.md.

Pitfalls

• A positive credit balance is required to start a forecast. A 402 with an insufficient-credits message means add credits at https://axion.eternis.ai.
• A run takes roughly 15 seconds to 3 minutes. Poll with backoff; do not block.
• Maximum 10 concurrent in-progress threads per account (429 with "code": "CONCURRENCY_LIMIT" otherwise). A 429 CAPACITY_QUEUE_FULL is a busy deployment rather than anything you did — wait out its Retry-After and send again.
• A 503 MAINTENANCE_MODE means Axion is read-only for planned maintenance. Reads and already-running forecasts are unaffected; retry the write after Retry-After.
• A 409 THREAD_CANCELED means the thread was stopped while your request was in flight — your own stop racing your own send, since you supply the thread id. Retrying against that thread keeps failing; start a new one.
• The key is read from the environment. Never print or hard-code it. In sandboxed runs, make sure AXION_API_KEY is provided to the sandbox, not only to the host.
• Successful responses are valid JSON, but a transient 5xx (an upstream reset at the ingress) can return a short plain-text body instead of the usual {"error": ...} envelope. Check the HTTP status before parsing rather than parsing the body as JSON. Resending the same request unchanged is the right answer for a 5xx only — the refusals above are deliberate, and each has its own, so retrying one of those on a timer just collects the same refusal.

Verification

A successful run ends with status completed and a non-empty forecasts[] array carrying probabilities between 0 and 1.