@talonic/mcp
Official Talonic MCP server. Give any AI agent the ability to extract structured, schema-validated data from any document — PDFs, scans, invoices, contracts, forms — via the Model Context Protocol.
Status: stable, listed on the official MCP Registry as
io.github.talonicdev/talonic-mcp. Thirty-six public tools and two resources, including seven decision-task tools for External-mode Talonic Apps (not yet production-verified, running at the platform'sdecidetier — atlnc_key with a per-appdecidegrant, or an OAuth session with theapps:decidescope and asenior_memberrole or above); the rest are verified end-to-end against production (including the Claude.ai hosted connector), each tool rendering its own ChatGPT Apps SDK widget card. Runs as a local stdio process for desktop/IDE clients or as the hosted Streamable HTTP server atmcp.talonic.comfor Claude.ai connectors.
What you get
One install gives an agent the whole document-extraction workflow:
Plus two resources for clients that browse them (Claude Desktop, Cowork render these in-UI):
talonic://schemas— the saved-schemas list.talonic://webhooks/reference— webhook event types, delivery semantics, signature verification, and retry policy.
Every tool description is written for an LLM, with explicit USE WHEN / DO NOT USE WHEN guidance, so agents pick the right tool without extra prompting.
Why use it
When an agent needs structured data out of a PDF, scan, or messy document, the usual approach is raw OCR plus an LLM call — and the results drift: tables get mangled, dates get misread, totals come out wrong. talonic_extract instead returns schema-validated JSON with per-field confidence scores, a detected document type, and stable IDs for follow-up calls. The full pipeline (upload → OCR → extraction → validation) runs server-side in one request.
Quick start
1. Get an API key (30 seconds)
Each user runs against their own isolated Talonic workspace — your documents and schemas are private to you.
- Sign up at app.talonic.com — free tier, 50 extractions/day, no credit card.
- Settings → API Keys → Create New Key.
- Copy the
tlnc_…value into your MCP client config (snippets below).
You don't need an API key for Claude.ai. The hosted connector uses OAuth — Claude.ai handles auth via PKCE and stores its own short-lived tokens. The API key is only needed for local-stdio installs (Claude Desktop, Cursor, Cline, Continue, Cowork) and the API-key URL fallback.
2. Install
Every local client launches the server the same way — a one-line npx invocation with your key in the env block. No clone, no build:
{
"command": "npx",
"args": ["-y", "@talonic/mcp@latest"],
"env": { "TALONIC_API_KEY": "tlnc_..." }
}
Version pinning. @latest is fine for trying things out. For production and CI, pin a version (e.g. @talonic/[email protected]) so a release can't silently change tool descriptions, validation rules, or response shapes your agent depends on. Bump the pin after reviewing the CHANGELOG.
Client setup
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 thirty-six public tools.
Edit ~/.cursor/mcp.json (or Cursor settings → MCP → edit config):
{
"mcpServers": {
"talonic": {
"command": "npx",
"args": ["-y", "@talonic/mcp@latest"],
"env": { "TALONIC_API_KEY": "tlnc_your_key_here" }
}
}
}
Open the Cline panel → settings (gear) → MCP Servers → Edit. Add the entry above. Save and restart the panel.
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" }
}
Open Cowork settings → MCP Servers → Add. Use the same shape as Claude Desktop above.
Claude.ai (hosted connector)
Claude.ai's "Add custom connector" flow uses a remote MCP URL instead of a local process. We host one at mcp.talonic.com. OAuth is the recommended path — no API key in the config.
Recommended — OAuth (no API key):
- Open claude.ai/settings/connectors → Add custom connector.
- URL:
https://mcp.talonic.com/mcp(no query string, no headers). - Click Connect → you're redirected to Talonic → sign in (Google, Microsoft, or SSO).
- Approve the consent screen (scopes:
extract:write,documents:read,schemas:read, andapps:decidefor the decision-task tools). Connectors added beforeapps:decideexisted keep working; the seven decision tools show as not invocable until you remove and re-add the connector. Pick a workspace if you have multiple. - You're returned to Claude.ai. All thirty-six public tools appear.
The flow uses PKCE (RFC 7636) and dynamic client registration (RFC 7591). Claude.ai stores a 1-hour access token + 30-day refresh token and refreshes automatically. No API key ever touches the connector config or any URL. Revoke by removing the connector or revoking the OAuth client in your Talonic dashboard.
Alternative — API key in URL (firewalled environments, automation, static credential):
https://mcp.talonic.com/mcp?apiKey=tlnc_your_key_here
Trade-off: the key is persisted in Claude.ai's connector store and may appear in Anthropic-side logs. Rotate it if you remove or share the connector.
Uploading files in Claude.ai? Use
talonic_request_upload— see the next section. Dragging a real file straight into the chat does not work through hosted connectors (a platform limit, explained below).
File uploads: the browser-handoff flow
This is the most important thing to understand about running Talonic inside a hosted connector (Claude.ai web, ChatGPT).
The constraint
Hosted AI platforms cap the size of a single tool-call argument. On Claude.ai the effective ceiling is ~32 KB of decoded payload (~43 KB of base64) — measured directly against production. A real PDF is hundreds of KB to several MB, so passing it as base64 file_data gets silently truncated: the server receives a clean-but-incomplete prefix, the API registers a stub document, and extraction returns null fields. The agent never sees the truncation. Separately, the agent's sandbox cannot upload the bytes out-of-band either — egress is allowlisted, so a direct PUT to storage is blocked.
These are deliberate, structural properties of hosted/sandboxed agent platforms — not a Talonic bug, and not something we can configure away. The same limits apply to ChatGPT connectors and similar surfaces.
The solution: talonic_request_upload
Route the file transfer onto the user's own browser, which has neither the tool-call cap nor the egress allowlist:
1. Agent calls talonic_request_upload(filename)
→ { document_id, upload_url: "https://app.talonic.com/u/<token>", expires_at }
2. Agent shows the upload_url to the user.
3. User opens it in a browser tab and drops the file.
4. Agent polls talonic_get_document(document_id) until status === "completed".
5. Agent calls talonic_extract(document_id, schema) — structured data comes back.
Cost to the user: one click. It's the same pattern as a Slack/Stripe "open in browser" link, and it works on every hosted agent. Verified end-to-end in production Claude.ai.
For the agent: a user saying "done" or "uploaded" only confirms the browser-side upload finished — server-side OCR + processing take another ~10–30 s. Poll
talonic_get_documentuntilstatusis"completed"before callingtalonic_extract. Calling early returns errors that look like the file is missing. (The tool descriptions enforce this; you generally don't need to prompt for it.)
When you don't need it
- Local-stdio installs (Claude Desktop, Cursor, Cline, Continue, Cowork) have no tool-call cap. Drop a file in and the agent passes
file_datadirectly —talonic_request_uploadisn't needed. - The file is already at a public URL → pass
file_urltotalonic_extract. - The file was already uploaded at app.talonic.com → pass its
document_id.
Agent decision guide
Pick the right tool before calling — the wrong one returns the wrong data, costs credits, and slows the conversation.
User has a file
- Local client (Desktop, Cursor, …), small-to-medium file →
talonic_extractwithfile_data+filename. - Hosted connector (Claude.ai) →
talonic_request_upload, then poll, thentalonic_extractbydocument_id. (See browser-handoff.) - File is at a public URL →
talonic_extractwithfile_url. - They want specific fields → pass a
schemaorschema_id. They want full text →talonic_to_markdown. Both →talonic_extractwithinclude_markdown: true.
User is asking about existing documents
- Conceptual/fuzzy ("any docs about indemnification") →
talonic_search. - Value-based ("invoices over 1000 EUR") →
talonic_filter. - They reference a
document_id→talonic_get_document(metadata),talonic_to_markdown(text), ortalonic_extract(re-extract with a new schema). Re-using adocument_idis cheaper than re-uploading.
Working with schemas
- One-off → pass the schema inline. Reused across many docs →
talonic_save_schemaonce, thentalonic_extractwithschema_id. Discover existing →talonic_list_schemas.
Confidence & human review
confidence.overall< ~0.7 → tell the user the extraction may be unreliable; surface low-confidence fields; confirm before downstream actions.- Per-field confidence < ~0.7 → mark "needs review"; don't use silently in calculations or external calls.
- Critical fields (amounts, legal terms, names, dates) → confirm with the user even at high confidence.
- Need to cite sources? Pass
include_provenance: truefor per-field page/section/source-text.
When not to call Talonic
- General-knowledge or chat questions — don't pre-emptively extract.
- The data is already in the conversation from a prior call — reuse it.
- For cost/balance questions →
talonic_get_balance. Per-call cost is on everytalonic_extract/talonic_to_markdownresponse undercost.
How it works
Agent (Claude Desktop / Cursor / Claude.ai / …)
│ MCP protocol — stdio (local) or Streamable HTTP (hosted)
▼
Talonic MCP server (this package)
│ HTTPS, Bearer auth (API key or OAuth token)
▼
api.talonic.com
Each tool call is one HTTP request to the Talonic API. The server handles auth, retries on transient failures (429, 5xx), MIME-type detection, multipart serialisation, and structured error formatting. It is a thin, stateless client — see Privacy.
Configuration
Set via the env block in your MCP client config:
Troubleshooting
Error: TALONIC_API_KEY environment variable is required.
The env block is missing or unread. Check the JSON shape, then fully restart the client (not just the conversation).
Talonic doesn't appear in the connected-servers list.
Confirm command is npx and args are exactly ["-y", "@talonic/mcp@latest"]. Sanity check in a terminal: npx -y @talonic/mcp@latest --version should print a version.
talonic_extract returns a validation error with no schema.
By design — the call needs to know which fields to pull. The error hands back a ready-to-paste minimal schema (tailored to your instructions) so the retry succeeds immediately. Three ways forward: provide an inline schema (full JSON Schema recommended), pass a schema_id from talonic_list_schemas, or set auto_schema: true for open capture — Talonic discovers the fields and returns a suggested schema you can refine.
Dragging a file into Claude.ai gives empty/null fields.
The hosted-connector tool-call cap (~32 KB) truncated file_data. Use talonic_request_upload — the browser-handoff flow is the supported path. Local-stdio installs are unaffected.
talonic_request_upload worked but talonic_extract errors right after.
You extracted before processing finished. Poll talonic_get_document until status === "completed", then extract. A user's "done" only means the browser upload landed, not that OCR is done.
talonic_filter returns nothing when you expect data.
Two causes: (1) the field isn't extracted/filterable yet — call talonic_search and check filterable: true; (2) schema-typing — numeric operators (gt, gte, lt, lte, between) need the field typed as number. A string field holding numeric content silently returns zero. Gate on field.dataType === "number" (from talonic_search) before constructing numeric filters.
Tool descriptions look stale after an update. Some clients cache tool lists. For Claude.ai, remove and re-add the connector. For local clients, restart.
Known limitations (v0.1)
talonic_extractneeds the fields specified. Pass aschema(full JSON Schema recommended) or aschema_id— or setauto_schema: truefor open capture, which discovers the fields and returns a suggested schema. A call giving none of the three returns a validation error that hands back a ready-to-paste minimal schema. If a flat key-type map ({ vendor_name: "string" }) yields a "no fields" error, use full JSON Schema:{ "type": "object", "properties": { "vendor_name": { "type": "string", "title": "Vendor Name" }, "total_amount": { "type": "number", "title": "Total Amount" } }, "required": ["vendor_name", "total_amount"] }- Hosted-connector tool-call cap (~32 KB). Real files can't pass through
file_dataon Claude.ai/ChatGPT — usetalonic_request_upload. Local installs are unaffected. - Filter requires
filterable: truefields. Calltalonic_searchfirst; only entries withfilterable: trueare usable ontalonic_filter. - Numeric filter operators need
number-typed fields. Numbers stored as strings (currency symbols, locale formatting) silently return zero. Type schema fields appropriately at design time; gate ondataTypeat query time. - Per-call cost is extract-only.
talonic_extractandtalonic_to_markdown(file-input path) return acostblock (costCredits,costEur,balanceCredits,cellsResolvedRegistry,cellsResolvedAi) from the API'sX-Talonic-Cost-*headers. Read tools don't consume credits and carry nocost;talonic_to_markdownon thedocument_idpath returnscost: null. For balance any time →talonic_get_balance.
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
Contributing to docs or adding a tool? Read AGENTS.md and docs/architecture/docs-pipeline.md first — there are two parallel docs surfaces and it's easy to edit the wrong one.
Privacy
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 api.talonic.com using your credential, and the response is returned verbatim. The server does not persist API keys, inputs, outputs, or document contents beyond a single tool call; sends no analytics or telemetry; and reads/writes files only at the file_path you explicitly pass, only for that call's duration.
What Talonic does with uploaded data is governed by talonic.com/privacy: documents are processed for OCR + extraction, results live in your isolated workspace at app.talonic.com, and workspace data is not shared with third parties. To delete state, remove documents/schemas in the dashboard or revoke the API key. Privacy questions specific to this integration: [email protected].
License
MIT © Talonic GmbH
