Ingest once, query forever.

One endpoint in. Two endpoints out.

curl -X POST https://api.talonic.com/v1/extract \
  -H "Authorization: Bearer $TALONIC_API_KEY" \
  -F "file=@invoice.pdf"

const result = await client.extract({
  file_path: "invoice.pdf",
  schema: {
    type: "object",
    properties: {
      vendor: { type: "string" },
      total_eur: { type: "number" },
      due_date: { type: "string", format: "date" },
      line_items: {
        type: "array",
        items: {
          type: "object",
          properties: {
            description: { type: "string" },
            amount_eur: { type: "number" }
          }
        }
      }
    },
    required: ["vendor", "total_eur"]
  }
});
// result.data == { vendor: "Acme Corp", total_eur: 1500.00, ... }
// Every field includes confidence and provenance.

{"type": "object", "properties": { ... }}

{"fields": [
  {"name": "vendor", "type": "string", "description": "..."}
]}

{"vendor": "string", "total": "number"}

# Typed rows over materialised field values. No model call, no credits.
POST /v1/documents/filter
{
  "conditions": [
    { "fieldId": "vendor_name",   "operator": "eq", "value": "Meridian Energy AG" },
    { "fieldId": "contract_year", "operator": "eq", "value": "2026" }
  ],
  "limit": 50
}

# A cited answer over the same corpus.
POST /v1/ask
{
  "question": "Which contracts renew in the next 30 days, and with what notice period?",
  "scope": { "pipeline_id": "pipe_contracts_2026" }
}
# -> rows + the SQL that produced them + an exact source span behind every claim

Same auth, same response envelope, same provenance. Pick the mode that fits the call you are writing.

Authentication

Authorization: Bearer $TALONIC_API_KEY

Manage keys in the dashboard

The agent billing story.

Read the cost on every call

Every synchronous /v1/extract response includes cost headers. Agents track spending per call without a separate request, and can see how many cells came free from the database.

HTTP/1.1 200 OK
X-Talonic-Cost-Credits: 70
X-Talonic-Cost-EUR: 0.07
X-Talonic-Balance-Credits: 64930
X-Talonic-Cells-Resolved-Registry: 0
X-Talonic-Cells-Resolved-AI: 1

Headers appear on synchronous 200 responses. Async 202 poll responses do not include them.

Check balance and runway

A single endpoint returns current balance, 30-day burn rate and projected runway in days.

GET /v1/credits/balance

{
  "balance_credits": 64930,
  "balance_eur": 64.93,
  "burn_rate_30d_credits": 12400,
  "projected_runway_days": 157,
  "tier": "pro",
  "tier_resets_at": "2026-05-01T00:00:00Z"
}

Auto top-up (human-gated)

Errors and retries

See the error reference for all error codes, the response envelope, retry guidance and backoff recommendations.

Webhooks

SDKs and MCP

npm install @talonic/node

import { Talonic } from "@talonic/node";

const client = new Talonic({
  apiKey: process.env.TALONIC_API_KEY!
});

{
  "mcpServers": {
    "talonic": {
      "url": "https://mcp.talonic.com/mcp",
      "headers": {
        "Authorization": "Bearer tlnc_live_..."
      }
    }
  }
}

{
  "mcpServers": {
    "talonic": {
      "command": "npx",
      "args": ["-y", "@talonic/mcp@latest"],
      "env": {
        "TALONIC_API_KEY": "tlnc_live_..."
      }
    }
  }
}

The Node SDK and MCP server are thin wrappers around the same REST API. For other languages, the API works directly. See the OpenAPI spec or jump to Mode 1 above.

Limits and metering

Talonic meters usage in credits, not request quotas. 1,000 credits are one euro, at every pack size.

Free

5,000 credits/month, no credit card. A hard cap: 429 until the monthly reset or a top-up.

Pay as you go

Prepaid credit packs from €10, flat rate at every size, bought inside the app. Credits keep for twelve months.

Enterprise

Volume contracts, invoicing, custom quotas.

Per operation: 100 credits per page ingested, 20 per AI-resolved cell, 0 per cell resolved from the database, 100 per matching or case operation. Batch mode runs at 0.5x.

Supported formats

25+ formats across documents, images, text and archives. German, English, French and Spanish at production quality.

Why not just RAG?

Because the second question should be free, and the same question tomorrow should return the same answer. Three published benchmarks, one method, with the rows we lose left in.

On one-question-per-document numeric reading, structure-first and retrieval tie in our own runs. The benchmark page says so. The advantage is the second question, and every one after it.