@talonic/mcp

Official Talonic MCP server. Lets AI agents extract structured, schema-validated data from any document via the Model Context Protocol.

Status: Listed on the official MCP Registry as io.github.talonicdev/talonic-mcp. Eight tools and two resources live: talonic_extract, talonic_search, talonic_filter, talonic_get_document, talonic_to_markdown, talonic_list_schemas, talonic_save_schema, talonic_get_balance, plus the talonic://schemas and talonic://webhooks/reference resources. Verified end-to-end against production.

Why an agent should use this

When an agent needs to pull structured data out of a PDF, scan, image, or messy document, the usual approach is raw OCR plus an LLM call. Results are unreliable; tables get mangled, dates get misread, totals drift.

With this MCP server installed, the agent has a talonic_extract tool that returns schema-validated JSON with per-field confidence scores, a detected document type, and stable IDs for follow-up calls. Seven other tools cover the rest of the workflow: searching the workspace, filtering by extracted field values, fetching a document's metadata, getting OCR markdown, listing saved schemas, saving new ones, and reading the workspace credit balance for budget-aware behaviour.

Get an API key (30 seconds)

Each user runs against their own isolated Talonic workspace. Your documents and schemas are private to you.

1. Sign up at https://app.talonic.com. Free tier: 50 extractions per day, no credit card.

2. Settings → API Keys → Create New Key.

3. Copy the tlnc_ value into your MCP client config (snippets below).

Install

The package is on npm. Every MCP client launches it the same way: a one-line npx invocation with your API key in the env block. No clone, no build.

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

The -y flag skips the npm install prompt.

Version pinning. @latest is fine for trying things out and for personal use. For production deployments and CI, pin to a specific version (e.g. @talonic/mcp@0.1.16) so a future release cannot silently change tool descriptions, validation rules, or the response shape your agent depends on. Bump the pin manually after reviewing the CHANGELOG.

Talonic uses a single API key per workspace. The same key authorises all tools. There is no scoping mechanism in v0.1; treat the key like any other secret and store it in your client's secret store rather than in version control.

Per-client snippets are below.

MCP client setup

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

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

Fully restart Claude Desktop (Cmd+Q on macOS, not just close the window). Talonic appears in the connected servers list with all eight tools.

Cursor

Edit ~/.cursor/mcp.json (or open Cursor settings → MCP → edit config):

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

Cline (VS Code extension)

Open the Cline panel → settings (gear icon) → MCP Servers → Edit. Add the entry above. Save and restart the panel.

Continue (VS Code / JetBrains)

Edit ~/.continue/config.json. Add to the mcpServers array:

{
  "name": "talonic",
  "command": "npx",
  "args": ["-y", "@talonic/mcp@latest"],
  "env": {
    "TALONIC_API_KEY": "tlnc_your_key_here"
  }
}

Cowork

Open Cowork settings → MCP Servers → Add. Use the same shape as Claude Desktop above.

Claude.ai (hosted MCP)

Claude.ai's "Add custom connector" flow uses a remote MCP URL instead of a local stdio process. We host one at mcp.talonic.com so Claude.ai users can install Talonic without running anything locally:

1. Open https://claude.ai/settings/connectors.

2. Click "Add custom connector".

3. URL: https://mcp.talonic.com/mcp?apiKey=tlnc_your_key_here

4. Click Add. The 8 tools appear.

Claude.ai's UI does not currently accept a custom Authorization header on connectors, so the API key is passed as a ?apiKey=... query parameter. Less secure than the Bearer header pattern (the key is persisted in the connector store and may appear in Anthropic-side logs), so rotate the key in your Talonic dashboard if you remove the connector. IDE-style clients (Cursor, Cline, Continue) that accept custom headers should use the Bearer header instead.

Caveat: drag-and-drop file uploads through talonic_extract currently stall on the hosted endpoint. Use file_url (a publicly reachable URL) or document_id (an already-uploaded document) for now. The local stdio install is unaffected.

Tool reference

Each tool's description is written for an LLM, with explicit USE WHEN / DO NOT USE WHEN sections. Agents pick the right tool reliably without further prompting. Briefly:

• talonic_extract, status: stable. Extract structured, schema-validated data from a document. Inputs: one of file_data + filename (recommended for chat clients, see below), file_path, file_url, or document_id, plus a schema or schema_id. Returns JSON with data, per-field confidence, document metadata, and a cost block (per-call credits / EUR / post-call balance) parsed from the API's X-Talonic-Cost-* response headers. Schema is required; the MCP layer rejects schema-less calls.

• talonic_search, status: stable. Omnisearch across documents, fields, sources, and schemas in the workspace. Use for conceptual or fuzzy queries.

• talonic_filter, status: stable. Filter documents by extracted field values using composable conditions (eq, gt, between, contains, is_empty, etc.). Accepts canonical field names (e.g. vendor.name) which the Talonic API resolves to ids server-side, or UUIDs directly. is_not_empty is intentionally not exposed in v0.1; see Known limitations.

• talonic_get_document, status: stable. Fetch full metadata for a single document by id, including processing log and link URLs.

• talonic_to_markdown, status: stable. Get OCR-converted markdown for a document. Accepts document_id (cheapest), file_data + filename, file_path, or file_url. Returns the same cost block as talonic_extract when an extract step ran (i.e., on the file inputs); null on the document_id path.

• talonic_list_schemas, status: stable. List all saved schemas with their definitions. Returns both UUID and SCH-XXXXXXXX short id; either is accepted by talonic_extract.

• talonic_save_schema, status: stable. Save a schema definition to the workspace for reuse.

• talonic_get_balance, status: stable. Read the workspace credit balance, EUR value, 30-day burn rate, projected runway, tier, and next-tier-reset timestamp. Use it for budget-aware decisions before kicking off large batches.

Two resources are also exposed for clients that browse them separately (Claude Desktop and Cowork render these in the UI):

• talonic://schemas: saved-schemas list.

• talonic://webhooks/reference: webhook event types, delivery behavior, signature verification algorithms, and retry policies. Use this when an agent needs to set up or troubleshoot a webhook integration without leaving the MCP context.

Agent decision guide

Pick the right tool before you call. The wrong tool returns the wrong data, costs unnecessary credits, and slows the conversation.

User has a file (or just dropped one in)

• They want specific fields (vendor, total, dates, parties): talonic_extract with schema or schema_id.

• They want full text content for summarisation, translation, or analysis: talonic_to_markdown.

• They want both: talonic_extract with include_markdown: true, one upload.

User is asking about existing documents

• Conceptual or fuzzy ("any docs about indemnification"): talonic_search.

• Value-based on extracted fields ("invoices over 1000 EUR"): talonic_filter.

• They reference a document_id: talonic_get_document for metadata, talonic_to_markdown for text, talonic_extract to re-extract with a new schema. Re-using a document_id is cheaper than re-uploading.

User is working with schemas

• One-off extraction: pass schema inline.

• Same schema across many documents: talonic_save_schema once, then talonic_extract with schema_id.

• Discover existing schemas first: talonic_list_schemas.

• Iterate inline before saving. Avoid clutter.

Confidence and human review

• confidence.overall below ~0.7: tell the user the extraction may be unreliable, surface low-confidence fields, confirm before any downstream action.

• Per-field confidence below ~0.7: mark as "needs review", do not use silently in calculations or external API calls.

• Critical fields (amounts, legal terms, names, dates): confirm with user before acting, even at high confidence.

• Per-field source provenance (page, section, source text snippet) is available by passing include_provenance: true on talonic_extract. Use it when the user wants to cite the source of a value or verify against the original document.

When not to call Talonic

• General-knowledge or chat questions. Do not pre-emptively extract.

• The data is already in conversation history from a previous tool call. Re-use it.

• The user wants to discuss or revise an extraction you already produced. Reason over the previous result instead of re-extracting.

• Cost, EUR price, and remaining balance are not surfaced in v0.1 tool responses. If the user asks about cost or credit balance, point them to https://app.talonic.com.

Drag-and-drop in chat clients

When the user drag-drops a PDF (or any supported file) into a chat-style MCP host such as Claude Desktop, Cowork, or Cursor, the file lands in a host-owned sandbox directory the MCP server cannot read. The path the host then hands the agent (something like /mnt/user-data/uploads/abc.pdf) is meaningless to a separately-running npx MCP process, so file_path calls fail with a filesystem error.

@talonic/mcp@0.1.4 and later solve this by accepting file_data (base64-encoded file bytes) and filename on talonic_extract and talonic_to_markdown. The agent reads the file bytes from the conversation, base64-encodes them, and passes them through the MCP tool call. The MCP server decodes, infers MIME type from the filename, and uploads to the Talonic API as a normal multipart request. The file never has to live on the MCP server's disk.

This works reliably in local-stdio installs (Claude Desktop, Cursor, Cline, Continue, Cowork) where the MCP server runs as an npx process spawned by the host and tool-call arguments are passed directly between agent and server.

Caveat: Claude.ai connectors and hosted MCP

When the hosted MCP at mcp.talonic.com/mcp is added as a custom connector inside Claude.ai's web UI, Claude.ai imposes a small hard limit on tool-call argument size (effectively under ~1KB per parameter). A real PDF base64-encoded is hundreds of KB, so the file_data string is truncated before reaching the MCP server. The Talonic API receives a few hundred bytes, registers an empty document, and returns a successful-looking response with null extracted fields. The agent often misdiagnoses this as a server bug and goes into recovery loops.

This is a Claude.ai platform limit on connector tool-call payloads, not a Talonic MCP server bug. There is no workaround at the MCP layer.

Recommended paths for Claude.ai users with real files:

1. file_url — host the file at a publicly reachable URL (S3, Drive share with public read, GitHub raw, etc.) and pass the URL.

2. document_id — upload the file at https://app.talonic.com first, take the resulting document id, and pass it.

3. Use a local-stdio install instead of the hosted connector. Claude Desktop, Cursor, Cline, Continue, and Cowork all spawn the MCP server locally and have no parameter cap.

A future architectural fix is pre-signed upload URLs (a tool that returns a one-time HTTPS PUT target so agents can upload outside the MCP tool-call channel). Tracked as an engineering follow-up.

How it works

Agent (Claude Desktop / Cursor / Cline / etc.)
  ↓ MCP protocol over stdio
Talonic MCP server (this package)
  ↓ HTTPS, Bearer auth
api.talonic.com

Each tool call is one HTTP request to the Talonic API, using your API key. The server handles auth, retries on transient failures (429, 5xx), MIME-type detection on file uploads, multipart serialisation, and structured error formatting.

Configuration

Set via the env block in your MCP client config:

Troubleshooting

Tool calls return Error: TALONIC_API_KEY environment variable is required. The env block in your MCP client config is missing or not being read. Double-check the JSON shape. After editing the config, fully restart the client (not just the conversation).

Talonic does not appear in the connected servers list. Make sure the command is npx and the args are exactly ["-y", "@talonic/mcp@latest"]. As a sanity check, in any terminal run npx -y @talonic/mcp@latest --version; it should print a version number. If you are on an older 0.1.x and see no output at all, you are hitting the silent-bin bug fixed in 0.1.3; upgrade by setting the args to ["-y", "@talonic/mcp@latest"] and restarting the client.

talonic_extract returns a validation error when no schema is given. By design in v0.1. Schema-less extraction is unreliable, so the MCP layer rejects calls that omit both schema and schema_id before they reach the API. Provide either an inline schema (full JSON Schema recommended) or a schema_id from talonic_list_schemas.

talonic_extract rejects with unsupported_file_type. The MIME type was inferred as application/octet-stream. The SDK infers from common file extensions; if your filename has an unusual extension, pass content_type explicitly to the SDK call (the MCP layer does not yet expose this; a future tool version will).

talonic_filter returns no results when you expect data. Two common causes. First, the field has not been extracted yet: call talonic_search first, then look at fields[] and fieldMatches[] in the response. Only entries with filterable: true are usable with talonic_filter. Second, schema-typing mismatch: numeric operators (gt, gte, lt, lte, between) require the schema field to be typed as number. A field typed as string that holds numeric content silently returns zero. Re-design the schema with the right type and re-extract.

talonic_extract returns empty fields when I drag a file into Claude.ai. Claude.ai's connector pipeline imposes a small hard limit on tool-call argument size (effectively under ~1KB). The base64-encoded PDF is truncated before reaching the MCP server, so the API receives a stub document and the extracted fields come back null. This is a Claude.ai platform limit on connector tool-call payloads, not a Talonic MCP server or API bug. Workaround: use file_url (publicly reachable URL), document_id (file already uploaded at app.talonic.com), or use a local-stdio install (Claude Desktop, Cursor, Cline, Continue, Cowork) which has no parameter cap. See "Drag-and-drop in chat clients" above.

Tool descriptions look wrong in my client. Some MCP clients cache tool descriptions. Restart the client after a server update.

Known limitations (v0.1)

• Schema is required on talonic_extract. Schema-less extraction is unreliable in v0.1 and is rejected at the MCP layer with a validation error. Always pass a schema (full JSON Schema recommended) or a schema_id.

• Schema definition: prefer full JSON Schema. The flat key-type map ({ vendor_name: "string", ... }) is documented as accepted, but if you get a "no fields" error from the API, fall back to:

  {
    "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 field_id) on talonic_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 is not exposed in v0.1. It underreports against fields known to be populated. Workaround: filter with eq/gt/contains against a known value, or use is_empty and invert the result client-side.

• Drag-and-drop file uploads in Claude.ai are capped by Claude.ai's tool-call argument size limit. A base64-encoded real PDF (typically hundreds of KB) cannot fit through Claude.ai's connector tool-call pipe (which truncates parameters under ~1KB). The Talonic API receives a few hundred bytes, registers an empty document, and returns a response with null extracted fields. This is a Claude.ai platform limit on connectors, not a Talonic MCP server bug. Workaround for Claude.ai users: use file_url (publicly reachable URL), document_id (file uploaded at app.talonic.com), or use a local-stdio install (Claude Desktop, Cursor, Cline, Continue, Cowork). The architectural fix is pre-signed upload URLs (engineering follow-up).

• Cost, EUR price, and remaining balance are not surfaced. The API does not return them in tool responses yet. Credit balance must be checked in the Talonic dashboard.

Upgrading from 0.1.0 / 0.1.1 / 0.1.2

Versions before 0.1.3 had a bug where the bundled MCP server bin would exit silently when launched via the npm bin symlink, which is exactly how every MCP client invokes it via npx. If your client config is on an older version and you see no talonic_* tools surface despite the config looking correct, you are hitting that bug.

The fix is to point your args at @latest (or @0.1.3 explicitly) and fully restart the client:

"args": ["-y", "@talonic/mcp@latest"]

Develop

git clone https://github.com/talonicdev/talonic-mcp.git
cd talonic-mcp
npm install
npm run build
npm test
node dist/server.js --version

Privacy Policy

This MCP server is a thin client. It does not collect, store, log, or transmit any data on its own. Every tool call is forwarded directly to the Talonic API at api.talonic.com using your TALONIC_API_KEY, and the API response is returned verbatim to the MCP client.

The server does not:

• Persist API keys, tool inputs, tool outputs, document contents, extracted data, or any other information beyond the lifetime of a single tool call.

• Send analytics, telemetry, crash reports, or usage metrics to Talonic, Anthropic, or any third party.

• Read or write files on your machine other than the file paths you explicitly pass to talonic_extract / talonic_to_markdown (file_path), and only for the duration of that call.

What Talonic itself does with the data you upload via this server is governed by Talonic's privacy policy at https://talonic.com/privacy. In summary, Talonic processes documents you submit so it can perform OCR and structured extraction, stores results in your isolated workspace at app.talonic.com, and does not share workspace data with third parties. For data retention, deletion requests, and contact information, refer to that policy.

If you uninstall the MCP server (e.g. remove @talonic/mcp from your client config), all server-side state is in your Talonic workspace, not on this server. To delete that state, sign in to app.talonic.com and remove the documents and schemas via the dashboard, or revoke the API key the server was using.

For privacy questions specific to this MCP integration, contact info@talonic.ai.

License

MIT (c) Talonic GmbH