# Pretick — complete agent guide

Forecasting AI: Give it numbers. Get the next set.
Base URL: https://pretick.ai
Public docs: https://pretick.ai/docs
Interactive reference: https://pretick.ai/docs/reference
Contract: https://pretick.ai/openapi.json

1. Request a free trial with POST /v1/trial and JSON {"email":"you@example.com"}. Five free successful forecasts. No card required. Signup limits apply.
2. A human retrieves the pretick_live_ key from email and gives it to their application's server or trusted agent. Do not expect the signup response to contain a key. Keep keys out of browser bundles, repositories, URLs and logs.
3. Call POST /v1/timeseries/forecast with Authorization: Bearer YOUR_KEY and Content-Type: application/json.
4. Call GET /v1/usage with the same authorization to see credit and request accounting.
5. To buy credit, use https://pretick.ai/pricing or POST /api/stripe/checkout with {"bundle":"start"}. Have a human complete the returned Stripe URL using the SAME email as the trial. $10 adds 500 forecasts at $0.02 each. No subscription. The key stays the same.

Input: history.columns (1–8 unique names), history.values (3–2048 equally spaced chronological rows, oldest first; one finite number per column; magnitude <=1e12); frequency (positive integer followed by s, min, h, d, w); targets (unique subset of columns); horizon (integer 1–1024); quantiles (strictly increasing unique levels between 0.1 and 0.9, default [0.1,0.5,0.9]); model optional, pretick-default.
Every target is jointly forecast. Other history columns are past-only covariates. No irregular timestamps, nulls, future covariates, categorical inputs or training endpoints in this release. Resample/align before calling. Frequency describes the interval, not a calendar input.

Example (16 observations):
{"history":{"columns":["sales"],"values":[[102],[108],[111],[104],[99],[106],[115],[120],[118],[114],[109],[112],[121],[125],[119],[116]]},"frequency":"1d","targets":["sales"],"horizon":7,"quantiles":[0.1,0.5,0.9]}

Response: forecasts[] entries each have target, median[horizon], quantiles {"0.1":[...],"0.5":[...],"0.9":[...]}. Also model_version (pinned Chronos-2 checkpoint), model, frequency, horizon, request_id, usage dimensions/inference time, cost_cents, balance_cents. No language generation or chat messages.
Pretrained Amazon Chronos-2 processes the input. Requests do not permanently train the model. Input/output series are not saved to the application's database; account, payment and request accounting metadata are retained. Do not submit personal information.

400 invalid input: fix before retrying. 401 invalid key. 402 insufficient balance. 413 payload too large (max 400,000 bytes). 429 trial limit. 503 busy/unavailable: failed inference returns credit. If refund_pending or accounting_pending, retain request_id and contact support. Requests are synchronous, allow 45 seconds. Client retries are separate requests; do not automatically retry a network timeout without checking usage. Start with one concurrent call.

Forecasts are estimates, not guarantees. Quantile bands are not guaranteed calibrated confidence intervals. Evaluate against chronologically held-out data before relying on predictions. Model fine-tuning, tenant workspaces, scheduled jobs and large batch inference are not current public features.
Support: es@textmei.com
Privacy: https://pretick.ai/privacy
Terms: https://pretick.ai/terms

Simple input: POST /v1/timeseries/forecast with {"data":"102, 108, 111"} or a numeric array. 3-1,024 observations. Optional horizon defaults to input count. Common delimiters accepted. Two columns or two rows mean time/key then value. Numeric keys, month/day[/year] dates, ISO dates and timestamps are continued at a regular interval; no-year dates start in the current year. Optional headers retained by the playground. Invalid/irregular times fail before billing.
Read predictions first: flat median array for one target, row matrix for multiple targets in prediction_columns order. times is present for supplied time/key input. forecasts remains available for quantiles and compatibility. Playground copies the best guess in the input layout and separators; default display rounds to two decimals, adjustable without another request.

Alternating date/value lines are also accepted and retained in copied output. Weekday-only daily dates with weekend gaps and occasional skipped dates use an inferred weekday schedule; values are kept in observation order without filling gaps. Future weekends are skipped; future holidays are NOT excluded. This assumption appears in the playground and time_schedule in the API response. Other irregular schedules remain invalid.