Carbone MCP

io.carbonePublished by carbone.ioOfficial publisherv2.2.0Updated Nov 7, 2026

Automate enterprise-grade documents from templates, merge and 100+ conversions: PDF, Office, HTML.

Overview

AI-generated overview

Official Carbone server that lets an assistant fill document templates with JSON data, convert files, and manage a template library.

What you can do
  • Fill Word, Excel, PowerPoint, HTML, Markdown or PDF-form templates with JSON data and render them to PDF or other formats
  • Convert documents between PDF, Office, HTML, SVG, Markdown, CSV and image formats
  • Inspect a template's tags and sample data, and list templates, categories and tags
  • Upload, version, categorize, deploy, download and delete templates
  • Check API health and list supported formats and features
When to use it
Use it when you want an assistant to turn data from a conversation or another MCP server into finished, branded documents such as quotes, invoices, certificates or reports, or to convert and finalize existing files.
Requirements
Either the hosted endpoint with OAuth sign-in to a Carbone account, or the local npm package run with npx, which needs Node.js 20.3 or later and a Carbone API key in CARBONE_API_KEY. An on-premise instance can be targeted with CARBONE_BASE_URL, which needs no key. A paid plan removes watermarks and unlocks all formats.
Before you install
Uploaded templates and samples are stored by Carbone, and delivered documents are kept up to 1 hour. delete_template deletes immediately and update_template_metadata can schedule deletion, so keep tool approval prompts on. Documents can carry prompt-injection text. Keep the API key out of version control and rotate it if exposed. A self-hosted HTTP server with a shared key lets anyone who reaches the port spend that account.

Installation

In SourceWeft

  1. Open Carbone MCP in the dashboard and add it to a workspace.
  2. Connect the required credentials. They are stored privately and never shown on public pages.
  3. Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Use with an agent

Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.

Read https://sourceweft.com/mcp/io.carbone%2Fcarbone-mcp/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.

Read the install guide the agent follows

Other MCP clients

Add this to your client's mcpServers config. The client completes OAuth sign-in on first connect.

{
  "mcpServers": {
    "carbone": {
      "type": "http",
      "url": "https://mcp.carbone.io/"
    }
  }
}

README

Carbone MCP Server

[npm version] [MCP Registry] [License: Apache-2.0]

Automate production-ready PDF, Word, Excel and PowerPoint documents from templates and JSON data, and convert files with 100+ conversions, from Claude, ChatGPT, Cursor, VS Code or any MCP client. The official server for the Carbone document generation API.

[Add to Claude] [Add to ChatGPT] [Add to Cursor] [Install in VS Code]

  • Convert files: DOCX, XLSX, PPTX, HTML, SVG, Markdown or PDF to PDF, images, CSV, ODT and the other output formats.
  • Fill templates with JSON: design them in Word, Excel, PowerPoint, LibreOffice, HTML or Markdown with {d.field} tags, or fill an existing PDF form.
  • Finish your PDFs: passwords, watermarks, PDF/A, page ranges, and a choice of converter (LibreOffice, OnlyOffice, Chromium or Carbone ICE).
  • Batch and localize: up to 100 documents per request, collected by the assistant on the hosted server or delivered to your own webhook; translations, time zones, and currency conversion at the rates you give.
  • Keep a template library: upload, version, categorize, deploy and download templates, and ask which data a template expects.

What you can ask

Carbone turns data into finished documents, including data another MCP server fetches (Salesforce, HubSpot, Stripe, Notion, Google Sheets…) in the same conversation:

  • "Get the Acme Corp opportunity and its products from Salesforce, fill my 'Sales quote' template, and give me a PDF with a DRAFT watermark."
  • "Which data does my 'Service agreement' template expect? Map its fields to the HubSpot deal 'Acme renewal' and show me the JSON before you render it."
  • "Create one certificate per attendee in the 'Workshop' Google Sheet with my 'Certificate' template, as a ZIP of PDFs named certificate-{d.lastName}."
  • "Generate this invoice in French, German and Spanish, with prices converted from EUR to USD at 1.08 and dates in New York time."
  • "Build a monthly report from last month's paid Stripe invoices with the 'Statement' template. Render it as a PNG first so you can check the layout, then give me the PDF."
  • "Turn this Word quote into a Carbone template: add tags for the customer, a loop on the line items and the total, then upload it as 'Quote v2' in the 'sales' category."
  • "Make a PDF/A copy of pages 1 to 5 of this contract for archiving, and a password-protected PDF with a CONFIDENTIAL watermark to send to the client."
  • "Deploy the newest version of my 'Invoice' template, and schedule the old one for deletion in 30 days."

Account and pricing

You sign in with a Carbone account, free to create. Without a plan, you get unlimited watermarked PDFs. The Free plan adds every output format without a watermark, with 100 documents per month (pricing).

Quick start

SetupRunsClientsFiles
Hosted server (recommended)https://mcp.carbone.io, nothing to installClaude, ChatGPT, Claude Code, VS Code, Cursor, …: sign in with your Carbone accountURLs, base64, ChatGPT attachments, upload URLs
Local packagenpx carbone-mcp on your machineClaude Desktop, Claude Code, VS Code, Cursor, and any client that starts local serversAlso reads local paths and saves documents to disk
DockerYour own serverClients that can send an Authorization headerURLs and base64

Hosted server (mcp.carbone.io)

The Carbone agent plugin installs this server and the Carbone Skill together, in Claude, ChatGPT, Codex, Cursor, GitHub Copilot, Gemini CLI and more. To add the server alone:

Claude (claude.ai and Claude Desktop)
  1. Open Carbone in the Claude directory and click Connect.
  2. Sign in with your Carbone account and approve.

If your organization does not offer directory connectors, add a custom connector with the URL https://mcp.carbone.io instead, leaving the OAuth client ID and secret empty. In an organization, each member connects with their own Carbone account, and their documents count against it.

ChatGPT

Open the Carbone plugin in ChatGPT, click Add, then sign in with your Carbone account. The plugin also includes the Carbone Skill.

Without the plugin: open Plugins, click Add, then Add custom MCP server. Enter the name Carbone and the URL https://mcp.carbone.io, choose OAuth, tick I understand and want to continue, click Create as a plugin, and sign in.

Claude Code, VS Code, Cursor

Claude Code: add the server, then run /mcp, select carbone and choose Authenticate.

bash
claude mcp add --transport http carbone https://mcp.carbone.io

VS Code (.vscode/mcp.json), which asks you to sign in when it starts the server:

json
{  "servers": {    "carbone": { "type": "http", "url": "https://mcp.carbone.io" }  }}

Cursor (~/.cursor/mcp.json or .cursor/mcp.json), then sign in from Cursor's MCP settings:

json
{  "mcpServers": {    "carbone": { "url": "https://mcp.carbone.io" }  }}

Local package (stdio)

Use it to read and save files on your computer. It needs Node.js 20.3 or later (node --version) and a Carbone API key from account.carbone.io: the test key makes unlimited watermarked PDFs (templates stored with it are kept 30 days), the production key every format (details).

Most clients use this config:

json
{  "mcpServers": {    "carbone": {      "command": "npx",      "args": ["-y", "carbone-mcp"],      "env": {        "CARBONE_API_KEY": "your_api_key_here"      }    }  }}
ClientWhere to put it
Claude DesktopSettings → Developer → Edit Config opens claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\). Quit and restart Claude Desktop afterwards.
Cursor~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project)

Claude Code:

bash
claude mcp add --env CARBONE_API_KEY=your_api_key_here --transport stdio carbone -- npx -y carbone-mcp

VS Code: .vscode/mcp.json uses servers, not mcpServers. With an input variable, VS Code asks for the key once and stores it securely:

json
{  "inputs": [    { "type": "promptString", "id": "carbone-api-key", "description": "Carbone API key", "password": true }  ],  "servers": {    "carbone": {      "type": "stdio",      "command": "npx",      "args": ["-y", "carbone-mcp"],      "env": { "CARBONE_API_KEY": "${input:carbone-api-key}" }    }  }}

Then try: "What can Carbone do?"

Client docs: Claude Desktop · Claude Code · VS Code · Cursor

Files in and out

  • Hosted server: give files as HTTPS URLs or base64, or attach them in ChatGPT. A large template goes through an upload URL, for assistants that can run commands, such as Claude Code.
  • Local package: absolute paths too (/Users/me/invoice.docx, ~/invoice.docx). A bare name such as invoice.docx is read as base64 and fails.
  • Results: text and images come back in the conversation. PDF and Office files come back as a one-time download link (hosted, valid 1 hour) or a saved file (local).

Details: file inputs and output delivery.

Tools

ToolWhat it does
convert_documentConverts a file to another format; Carbone tags stay as written
render_documentFills a template with JSON data, and can convert the result (to PDF, for example)
get_render_resultReturns the download link of a render started with async: true (only when async renders are on)
inspect_templateLists a template's tags and returns a matching sample dataset. Runs one render on your account (counts as one document generation)
list_templatesLists stored templates, filtered by category, name or ID
list_categoriesLists the template categories in your account
list_tagsLists the tags used across your templates
upload_templateStores a reusable template, with versions, category and tags; can return an upload URL for a large file
update_template_metadataRenames, categorizes, tags, deploys or schedules the deletion of a template version
delete_templateDeletes a template or one of its versions, immediately
download_templateDownloads a template's original file, or the sample data stored with it
get_api_statusChecks the Carbone API's health and version
get_capabilitiesSummarizes the formats, features and usage examples

Every parameter, with examples: docs/API.md.

Writing templates

Design a template in Word, Excel, PowerPoint, LibreOffice, HTML or Markdown, and put tags where the data goes:

Dear {d.customer.name},Your invoice total is {d.total:formatC()}.Items:{d.items[i].description}  {d.items[i].quantity}x  {d.items[i].price:formatC()}{d.items[i+1]}

{d.items[i]} … {d.items[i+1]} repeats a line for each item. Start with the basics, or give your AI assistant the Carbone Skill, a template syntax reference for AI tools (download .skill · GitHub).

Troubleshooting

SymptomCause and fix
Carbone is connected, but a Claude chat does not use itTurn it on for the chat: the + button, then Connectors, then Carbone.
mcp.carbone.io asks you to sign in again, or sign-in failsReconnect: in Claude, Disconnect then Connect the connector; in Claude Code, /mcp → Clear authentication, then Authenticate.
Watermark on every PDF, or only PDF worksYour account has no plan (test key): it produces watermarked PDFs only. On mcp.carbone.io, start a plan at account.carbone.io (the Free plan is enough): the server switches to your production key by itself. Locally, use the production key.
Tags such as {d.name} still visible in the outputconvert_document is a pure conversion and keeps tags. Use render_document with your data.
A file given by name (invoice.docx) failsA bare name is read as base64. Pass an absolute path (local package), a URL, or base64.
Request body too large, or the local connection closesA file passed as base64 is too large (60 MB by default). Pass it by URL or local path; for a template, ask for an upload URL.
Rendering by Template ID still uses the old version after an uploadRenders by Template ID use the deployed version (most recent deployedAt). Ask to deploy the new version, or render by its Version ID.
Password, watermark or PDF/A missing from a PDFPDF/A, page ranges and accessibility need the default converter L (LibreOffice); O, C and I apply passwords, permissions and watermarks only.
Rate limit exceeded. Please wait N second(s)The Carbone API answered 429. Wait the time given and try again.
Local package: Authentication failed, or the server stops with CARBONE_API_KEY environment variable is requiredSet CARBONE_API_KEY in the client config's env and restart the client. Check the key at account.carbone.io.
Claude Desktop shows no Carbone tools, or the local server disconnectsRead mcp-server-carbone.log in ~/Library/Logs/Claude/ (macOS) or %APPDATA%\Claude\logs\ (Windows). A good start logs Carbone MCP Server v<version> started (stdio).
You want to test the server without an AI clientRun the MCP Inspector (Node.js 22.19+): npx @modelcontextprotocol/inspector -e CARBONE_API_KEY=your_api_key_here npx carbone-mcp.

Security

  • Tool approval. Keep your client's approval prompts on for actions that change your account: delete_template deletes immediately, and update_template_metadata can schedule a deletion.
  • Prompt injection. A document or template can contain instructions that trick the AI into unintended actions, such as sending data elsewhere or deleting templates. Review what your client is about to do before you confirm a tool call.
  • Disconnecting. Disconnect the connector (in Claude Code, /mcp → Clear authentication) to stop the assistant from using your account.
  • URL inputs and local files. URLs that resolve to a private or internal address (localhost, RFC 1918, cloud metadata) are refused unless CARBONE_ALLOW_PRIVATE_NETWORK=true (details). Local paths are read by the local package only, where the server runs as you.
  • Shared server key (self-hosted HTTP). With CARBONE_API_KEY set on an HTTP server and CARBONE_REQUIRE_CLIENT_AUTH_HEADER=false (the default), every request without a Bearer key spends that key: anyone who can reach the port uses that Carbone account. The server prints a warning at startup. Never publish such a port on a network you don't fully trust, or set CARBONE_REQUIRE_CLIENT_AUTH_HEADER=true so each client must bring its own key.
  • Browser origins (self-hosted HTTP). A request whose Origin header is present but not allowed gets 403, so a malicious web page cannot reach a locally running server (DNS rebinding) and spend its key. List the browser-based clients you trust in MCP_ALLOWED_ORIGINS.
  • API keys. Keep them out of version control, pass them through your client's secret input or environment, and rotate them at account.carbone.io.
  • Your data. Carbone stores the templates and samples you upload, and keeps a document delivered as a download link or to a webhook until it is downloaded, for 1 hour at most. For an async render, the server keeps an encrypted reference to the document for 1 hour. See the privacy policy, or run Carbone on-premise for full control.

Self-hosting

Docker

The image serves Streamable HTTP on port 3000. This quick start sets no server key: each client sends its own key in Authorization: Bearer <key> (signing in with a Carbone account is for mcp.carbone.io only), and the port is reachable from your machine only.

bash
docker run -d --name carbone-mcp -p 127.0.0.1:3000:3000 carbone/carbone-mcpcurl http://localhost:3000/health

Connect your client with your key in the Authorization header. Claude Code:

bash
claude mcp add --transport http carbone http://localhost:3000 --header "Authorization: Bearer your_api_key_here"

VS Code (.vscode/mcp.json, which asks for the key once and stores it):

json
{  "inputs": [    { "type": "promptString", "id": "carbone-api-key", "description": "Carbone API key", "password": true }  ],  "servers": {    "carbone": {      "type": "http",      "url": "http://localhost:3000",      "headers": { "Authorization": "Bearer ${input:carbone-api-key}" }    }  }}

Cursor (~/.cursor/mcp.json), reading the key from your environment:

json
{  "mcpServers": {    "carbone": {      "url": "http://localhost:3000",      "headers": { "Authorization": "Bearer ${env:CARBONE_API_KEY}" }    }  }}

To serve other machines, publish the port behind a TLS reverse proxy, keep the server without CARBONE_API_KEY (or set CARBONE_REQUIRE_CLIENT_AUTH_HEADER=true), and list the hostnames of browser-based clients in MCP_ALLOWED_ORIGINS.

Shared-key variant, for one user on one machine. A server-level key lets clients connect without sending one. Read the shared-key warning first.

bash
docker run -d --name carbone-mcp -p 127.0.0.1:3000:3000 -e CARBONE_API_KEY=your_api_key_here carbone/carbone-mcp

Docker Compose: compose.yml builds the image from this repository, binds 127.0.0.1 only, sets no server key, and passes through every variable of .env.example. Start it with docker compose up -d. Compose also reads CARBONE_API_KEY from your shell or a .env file: if it is set there, you get the shared-key variant.

To run the image as a stdio server for Claude Desktop, see the Docker Hub page.

On-premise Carbone

If you run Carbone on-premise, point the server at your instance with CARBONE_BASE_URL; no API key is needed. For example CARBONE_BASE_URL=https://your-carbone-server.com npx carbone-mcp, or add -e CARBONE_BASE_URL=https://your-carbone-server.com to docker run.

Configuration

VariableDefaultPurpose
CARBONE_API_KEYnoneYour Carbone API key: required in stdio for Carbone Cloud, optional over HTTP (Security)
CARBONE_BASE_URLhttps://api.carbone.ioCarbone API URL; set it to an on-premise instance (no key needed)
CARBONE_TIMEOUT60000Carbone API request timeout in ms; Carbone Cloud's own limit is 60 s
CARBONE_MAX_FILE_BYTES104857600 (100 MB)Largest input file read from a path, URL or base64
MCP_TRANSPORTstdio (http in Docker)stdio or http
MCP_PORT3000HTTP port
MCP_PATH/HTTP endpoint path; cannot be /health
MCP_MAX_BODY_BYTES62914560 (60 MB)Largest MCP message: an HTTP body (413 above it) or a stdio message
MCP_ALLOWED_ORIGINSnone (localhost is always allowed)HTTP: browser Origin hostnames to allow, comma-separated
CARBONE_REQUIRE_CLIENT_AUTH_HEADERfalseHTTP: true refuses requests that bring no Bearer key of their own
CARBONE_ALLOW_PRIVATE_NETWORKfalsetrue lets URL inputs reach private and internal addresses
MCP_TICKET_SECRETnoneHTTP: turns on upload URLs for large templates (details); generate with openssl rand -base64 32
MCP_PUBLIC_URLnoneHTTP, with MCP_TICKET_SECRET: the public https origin of this server, where upload URLs and Carbone's webhook point
MCP_S3_*noneHTTP, with MCP_TICKET_SECRET: an S3-compatible bucket for async renders and batches (details)
MCP_JOBS_DIRnoneInstead of MCP_S3_*: a directory every instance shares, such as a Docker or Kubernetes volume (details)

Full descriptions: Configuration in docs/API.md.

Self-hosting troubleshooting

SymptomCause and fix
Tool error Authentication failed: …Send Authorization: Bearer <key> from the client; a server with CARBONE_REQUIRE_CLIENT_AUTH_HEADER=true requires it on every request.
HTTP 403 Invalid Origin: <host>A browser-based client served from a host other than localhost. Add its hostname to MCP_ALLOWED_ORIGINS.
HTTP 405 Method not allowed when you open the endpoint in a browserExpected: the endpoint only answers MCP POST requests. Check the server with GET /health.
HTTP 413 Request body too largeOne MCP message is over MCP_MAX_BODY_BYTES. Pass files by URL, or raise the limit.
Async renders: the job store cannot be read at startupThe S3 access key needs to read, write and list the bucket, or MCP_JOBS_DIR must be writable.

Upgrading from 1.x

HTTP deployments should upgrade now (security fixes): read the upgrade notes.

Contributing

Bug reports, fixes and documentation improvements are welcome. CONTRIBUTING.md covers the development setup, tests, debugging and the pull request process.

Support

License

Apache 2.0, see LICENSE.

Source: README.md at commit c9cd2ec

Tools

0
This server asks you to sign in, so SourceWeft could not read its tools without an account and has not indexed them. What a connected account sees can depend on that account and its permissions.

Version history

8
  1. v2.2.0LatestNov 7, 2026
  2. v2.1.4Oct 26, 2026
  3. v2.1.3Oct 20, 2026
  4. v2.1.2Oct 20, 2026
  5. v2.1.1Oct 15, 2026
  6. v2.1.0Oct 15, 2026
  7. v2.0.0Oct 14, 2026
  8. v1.6.0Sep 16, 2026

Score

85% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 daysRequired25 pts
  • Auth declaredRequired10 pts
  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts