Status: Core surface stable. Live extract is verified end-to-end against production. documents.filter() now passes canonical field names directly to the API for server-side resolution; remaining edge cases are tracked in Known issues.

Looking for the AI agent path? @talonic/mcp wraps this SDK as a Model Context Protocol server. Easiest install is the hosted endpoint at https://mcp.talonic.com/mcp from Claude.ai's "Add custom connector" flow, which authenticates via OAuth 2.1 against app.talonic.com (no API key in the connector config). For IDE-style clients (Claude Desktop, Cursor, Cline, Continue, Cowork), use the local stdio install: npx -y @talonic/mcp@latest with TALONIC_API_KEY in env. Either way, any MCP-aware agent can extract documents directly.

Looking for the Python path? talonic on PyPI. pip install talonic. Same API surface as this Node SDK, both sync and async.

Install

npm install @talonic/node

Requires Node.js 18 or newer. Zero runtime dependencies.

Get an API key (30 seconds)

Every user runs against their own Talonic workspace, so each user needs their own key. Workspaces are isolated; your documents and schemas are private to you.

Sign up at https://app.talonic.com. Free tier: 50 extractions per day, no credit card.
Settings → API Keys → Create New Key.
Copy the tlnc_ value.
Set it as the TALONIC_API_KEY environment variable, or pass it directly to the client constructor.

Quickstart

import { Talonic } from "@talonic/node"

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

const result = await talonic.extract({
  file_path: "./invoice.pdf",
  schema: {
    type: "object",
    properties: {
      vendor_name: { type: "string", title: "Vendor Name" },
      invoice_number: { type: "string", title: "Invoice Number" },
      total_amount: { type: "number", title: "Total Amount" },
      due_date: { type: "string", title: "Due Date", format: "date" },
    },
    required: ["vendor_name", "total_amount"],
  },
})

console.log(result.data)
// { vendor_name: "Acme Corp", invoice_number: "INV-2024-0847", total_amount: 14250, due_date: "2024-03-15" }

console.log(result.confidence.overall)
// 0.97   — overall extraction confidence (0..1)

console.log(result.rateLimit)
// { limit, remaining, resetAt }   — parsed from X-RateLimit-* headers

Pass full JSON Schema with type: "object" and properties for reliable results. The SDK auto-populates required from properties when omitted, so you cannot accidentally end up with the silent-empty-data footgun where the API returns null for fields you intended to extract. Pass include_provenance: true to also receive per-field source evidence (page, section, source text).

API surface

// Top-level
await talonic.extract({ ... })

// Documents
await talonic.documents.list({ per_page: 50 })
await talonic.documents.get(id)
await talonic.documents.getMarkdown(id)
await talonic.documents.reExtract(id)
await talonic.documents.delete(id)

// Extractions
await talonic.extractions.list({ document_id })
await talonic.extractions.get(id)
await talonic.extractions.getData(id)             // returns object
await talonic.extractions.getData(id, { format: "csv" })  // returns string
await talonic.extractions.patch(id, { corrections: [...] })

// Schemas
await talonic.schemas.list()
await talonic.schemas.get(id)
await talonic.schemas.create({ name, definition })
await talonic.schemas.update(id, { ... })
await talonic.schemas.delete(id)

// Asynchronous batch jobs
await talonic.jobs.create({ schema_id, document_ids })
await talonic.jobs.list()
await talonic.jobs.get(id)
await talonic.jobs.getResults(id)
await talonic.jobs.cancel(id)

// Credits
await talonic.credits.getBalance()
// returns { balance_credits, balance_eur, burn_rate_30d_credits,
//           projected_runway_days, tier, tier_resets_at }

Per-call cost is also surfaced on the extract response under result.cost (costCredits, costEur, balanceCredits, cellsResolvedRegistry, cellsResolvedAi), parsed from the X-Talonic-Cost-* and X-Talonic-Balance-* response headers. null on calls that do not run through the extract path.

CLI

The package ships with a talonic binary:

talonic schemas list
talonic documents list --per-page=20
talonic extract ./invoice.pdf \
  --schema='{"type":"object","properties":{"vendor_name":{"type":"string"},"total_amount":{"type":"number"}},"required":["vendor_name","total_amount"]}'
talonic --help

Configuration

const talonic = new Talonic({
  apiKey: process.env.TALONIC_API_KEY!,
  baseUrl: "https://api.talonic.com", // default
  timeout: 60_000, // ms; default 60s
  maxRetries: 3, // 429, 500, 502, 503, 504, network, timeout
  fetch: customFetch, // optional override (e.g. for testing)
})

The SDK retries automatically on 429 (respecting X-RateLimit-Reset), 500, 502, 503, 504, network errors, and timeouts. Backoff is exponential with jitter, capped at 16s. The API may set retryable: false on a specific error; the SDK respects that and does not retry.

Rate limits

Every successful response is wrapped in WithRateLimit<T> and includes a rateLimit block parsed from the X-RateLimit-* response headers:

const result = await talonic.schemas.list()

result.data // SchemaList
result.rateLimit // { limit, remaining, resetAt } | null

Since 0.1.8, rateLimit is nullable. When the response carries no X-RateLimit-* headers (e.g. enterprise / unlimited tier, or any path that does not run through the rate-limit interceptor) the SDK returns rateLimit: null instead of a sentinel zeros object. Handle the null case explicitly when reading result.rateLimit.limit and friends. TalonicRateLimitError.rateLimit stays non-null on 429 since the response always carries the headers in practice.

Errors

Every failure is a TalonicError subclass:

import {
  TalonicError,           // base
  TalonicAuthError,       // 401, 403
  TalonicNotFoundError,   // 404
  TalonicValidationError, // 400, 409, 413, 422
  TalonicRateLimitError,  // 429 (after retries exhausted)
  TalonicServerError,     // 500, 502, 503, 504 (after retries exhausted)
  TalonicNetworkError,    // DNS / TCP failures
  TalonicTimeoutError,    // request exceeded timeout
} from "@talonic/node"

try {
  await talonic.extract({ ... })
} catch (err) {
  if (err instanceof TalonicRateLimitError) {
    console.log(`Reset at ${err.rateLimit.resetAt}`)
  } else if (err instanceof TalonicError) {
    console.error(`${err.code} (status ${err.status}, request ${err.requestId}): ${err.message}`)
  }
}

Known issues

Schema is required on extract. Schema-less extraction is unreliable in v0.1. Always provide a schema (full JSON Schema recommended) or a schema_id. The MCP wrapper at @talonic/mcp rejects schema-less calls at the layer above; the SDK passes through but you will hit unreliable behaviour.
Schema definitions: prefer full JSON Schema. The flat key-type map ({ vendor_name: "string", ... }) is documented as supported and the API's own error message lists it as accepted, but as of writing the server-side normaliser does not always translate it. If a flat-map save returns a "no fields" error, fall back to:
```
schema: {
  type: "object",
  properties: {
    vendor_name: { type: "string", title: "Vendor Name" },
    total_amount: { type: "number", title: "Total Amount" },
  },
  required: ["vendor_name", "total_amount"],
}
```
Filter requires filterable: true fields. Call talonic.search(...) first; only entries in the response where filterable: true can be used as field (or fieldId) on talonic.documents.filter(...). Entries with filterable: false exist in the schema but have no extracted data yet.
Schema field type affects filter operators. Numeric operators (gt, gte, lt, lte, between) only work on fields typed as number in the schema. Numeric values stored as strings (with currency symbols, locale formatting, etc.) silently return zero results. Type your schema fields appropriately at design time.
is_not_empty filter checks materialized data. The operator works correctly for documents that have completed extraction and materialization (typically within seconds via the DOCUMENT_RESOLVED event). For batch-mode documents, materialization occurs after the batch poll cycle applies results. If a value is known to be populated but is_not_empty does not return it, give the materialized-values index a moment after extraction and retry.

Development

npm install
npm test                      # 100+ unit tests
npm run typecheck
npm run check:spec            # verify SDK paths against the official OpenAPI spec
npm run build
npm run debug                 # diagnostic against the live API (needs TALONIC_API_KEY)
npm run test:live             # full live integration tests

Source of truth: `@talonic/docs/openapi.json`

The canonical Talonic API contract is the OpenAPI spec shipped via the @talonic/docs npm package. The SDK is built against it and CI verifies every call site stays in sync.

If you ever see CI fail with API spec drift detected, one of two things has happened: the spec changed and the SDK needs an update, or someone introduced a path that does not exist in the spec. Either way, fix the code or bump @talonic/docs. Do not bypass the check.

We previously kept a talonic-api-reference.md in the repo root; that document was incomplete and incorrect on several paths and has been removed. Use the OpenAPI spec instead.