AI Model Trainer API

Call the models you train in StayCharted AI Model Trainer from your own code, a spreadsheet, or an automation tool.

Base URL https://amt.staycharted.com

You train a model in the app by uploading examples you've already labelled. Then send it new text, or links to pictures, and get back a category and a confidence. Three endpoints cover everything:

The API is included on Essentials and Business. See Pricing.

1. Create a key. Sign in to AI Model Trainer, open Settings → API keys, and create one.

• The full key (it starts with amt_live_) is shown once. Store it then.
• Up to ten keys can be active at once. Give each integration its own, so you can revoke one without breaking the others.

2. Find your model's id. It's listed on the same screen. Or call GET /api/v1/models, which also confirms the key works and costs nothing. Only published models can be called.

3. Classify something.

curl -sS https://amt.staycharted.com/api/v1/models/$MODEL_ID/classify \
  -H "Authorization: Bearer $STAYCHARTED_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"input":"My invoice charged me twice this month"}'

{
  "label": "Billing",
  "confidence": 0.94,
  "modelId": "8f14e45f-ceea-467a-9b2c-1d5c2b0e7a11",
  "modelName": "Support tickets"
}

label is the answer. confidence runs from 0 to 1.

Every record is one prediction, drawn from the same monthly allowance your file fills in the app use. A request that is refused, for any reason below, costs nothing.

The rate limit is counted in records per minute, for the whole workspace: 30 a minute on Essentials, 60 on Business.

• Every key draws on that one allowance, so adding keys doesn't add throughput.
• The allowance refills continuously; it doesn't reset on the minute.
• A request that doesn't fit is refused whole with 429 and a Retry-After header, and nothing is charged.

So the API is for a steady trickle, not a flood. It suits a ticket as it arrives, a row as it's created, a form as it's submitted. For thousands of rows, upload the file in the app instead. It runs in one pass on hardware sized for the job, and you download the filled file.

Google Sheets: =CLASSIFY(A2)

In your sheet: Extensions → Apps Script, paste this, save, and go back to the sheet.

const STAYCHARTED_KEY   = 'amt_live_...';  // your API key
const STAYCHARTED_MODEL = '8f14e45f-...';  // your model id

/**
 * Classify one cell.
 * @param {string} text The cell to classify.
 * @return {string} The label.
 * @customfunction
 */
function CLASSIFY(text) {
  if (!text || !String(text).trim()) return '';

  const url = 'https://amt.staycharted.com/api/v1/models/'
    + STAYCHARTED_MODEL + '/classify';

  // Two attempts: a column of formulas recalculates all at once and the
  // per-minute allowance is shared. Retry-After says exactly how long to wait.
  for (let attempt = 0; attempt < 2; attempt++) {
    const res = UrlFetchApp.fetch(url, {
      method: 'post',
      contentType: 'application/json',
      headers: { Authorization: 'Bearer ' + STAYCHARTED_KEY },
      payload: JSON.stringify({ input: String(text) }),
      muteHttpExceptions: true,
    });

    const code = res.getResponseCode();
    if (code === 200) return JSON.parse(res.getContentText()).label || '';

    if (code === 429 && attempt === 0) {
      const wait = Number(res.getAllHeaders()['Retry-After'] || 5);
      Utilities.sleep(Math.min(wait, 25) * 1000);
      continue;
    }
    // Show the reason rather than an empty cell.
    return 'ERR ' + code + ': ' + (JSON.parse(res.getContentText()).code || '');
  }
  return 'ERR rate limited';
}

Then type =CLASSIFY(A2) in a cell and drag it down.

• Start with a few rows at a time. Sheets recalculates the whole column at once. Large selections can exceed the shared rate limit and return ERR rate limited. For more than a few hundred rows, fill the file in the app instead.
• The key travels with the sheet. Anyone you share the sheet with can read the script. Give it its own key, and revoke the key when you're done.

Zapier

Add a Webhooks by Zapier → Custom Request action.

The next step can use label, confidence, modelId and modelName as plain fields. Zapier counts one task per action, so a thousand records is a thousand tasks.

Make

Add HTTP → Make a request.

• URL: https://amt.staycharted.com/api/v1/models/<MODEL_ID>/classify
• Method: POST
• Headers: Authorization = Bearer <YOUR_API_KEY>
• Body type: Raw, content type application/json
• Request content: {"input": "{{yourField}}"}
• Parse response: yes. Without it, the next module sees a string instead of label.

Your own code: /predict for batches

Send up to 100 records in one call. Answers come back in the order sent, each with its input echoed.

curl -sS https://amt.staycharted.com/api/v1/models/$MODEL_ID/predict \
  -H "Authorization: Bearer $STAYCHARTED_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"inputs":["My invoice charged me twice this month","The parcel arrived with a cracked screen"]}'

{
  "model": { "id": "8f14e45f-ceea-467a-9b2c-1d5c2b0e7a11", "name": "Support tickets" },
  "predictions": [
    { "input": "My invoice charged me twice this month", "label": "Billing", "confidence": 0.94 },
    { "input": "The parcel arrived with a cracked screen", "label": "Shipping", "confidence": 0.71 }
  ],
  "usage": { "predictions": 2 }
}

The whole request is checked against your remaining monthly allowance before any of it runs. A request that would go over is refused entirely, never half-answered.

AI assistants

To use your models from Claude or ChatGPT instead of code, connect the assistant instead. See Use AI Model Trainer from Claude and ChatGPT.

Authentication

Send your key in every request:

Authorization: Bearer amt_live_…

Keys are created and revoked only in the app, never through the API. A key that could create keys couldn't really be revoked.

GET /api/v1/models

Lists the models this key can call. A model appears only when it is published, switched on, and allowed by your plan. This call spends no predictions and has its own allowance of 60 requests a minute.

{
  "models": [
    { "id": "8f14e45f-ceea-467a-9b2c-1d5c2b0e7a11", "name": "Support tickets", "custom": true },
    { "id": "3c59dc04-8e88-4950-9f2b-7a1c2a1e0b02", "name": "Product feedback", "custom": false }
  ]
}

custom is true for a Dedicated AI Model. It is false for a Classifier Model or an AI Classifier Model. All of them are called the same way.

POST /api/v1/models/{id}/classify

One record in, one flat answer out.

POST /api/v1/models/{id}/predict

Exactly one of:

predictions is always an array, even for a single input.

Pictures

For a model that reads pictures:

• Each picture counts once against your monthly picture allowance, as well as once as a prediction.
• A picture that can't be used comes back with label and confidence set to null, plus an error code and a message, and isn't counted. The codes are: not_found, forbidden, unreachable, timeout, too_large, web_page, not_an_image, unsupported_format, too_small, corrupt, blocked_address, not_a_link.
• Accepted formats: JPEG, PNG, WebP, GIF, BMP and TIFF.

Rate-limit headers

Errors

Every error is JSON with a stable code. Branch on code, never on the error sentence, which is written for people and may be reworded.

A refusal caused by your plan also carries entitlement, limit, used, resetsAt and upgrade. Those let a client tell "wait until the month resets" from "this plan will never allow it".