MCP Server

Quick Setup with Claude Desktop

{
  "mcpServers": {
    "pressa": {
      "command": "npx",
      "args": ["-y", "@pressa/mcp"]
    }
  }
}

Optional: add an API key

Without a key the server starts in anonymous mode and registers only compile, with 3 successful compilations per day. A free key raises the limit to 50 compilations per month and allows image assets; it also registers the template, asset and render tools, which need a paid plan to save or render. Create one in your dashboard (keys start with pressa_) and pass it as an environment variable:

{
  "mcpServers": {
    "pressa": {
      "command": "npx",
      "args": ["-y", "@pressa/mcp"],
      "env": {
        "PRESSA_API_KEY": "pressa_your_key_here"
      }
    }
  }
}

Setup with Other Tools

Any MCP-compatible AI tool can use the Pressa server. The configuration format is similar across tools.

Cursor

{
  "mcpServers": {
    "pressa": {
      "command": "npx",
      "args": ["-y", "@pressa/mcp"]
    }
  }
}

To use a key, add the same env block as in the Claude Desktop example above.

Windsurf

Use the same JSON structure in Windsurf's MCP configuration file. Refer to Windsurf's documentation for the exact file path.

n8n (AI Agent)

No community node is needed. Attach an MCP Client Tool to your AI Agent node and set:

• Endpoint: https://api.pressa.dev/mcp
• Server Transport: HTTP Streamable
• Authentication: Bearer Auth, with your Pressa API key as the token

The agent writes the LaTeX, calls compile and fixes its own errors from the diagnosis. The tool returns a link to the PDF; to attach it in Gmail or upload it to Google Drive, add an HTTP Request node on that link with Response Format set to File. Use a key: n8n Cloud sends requests from a shared pool of addresses, so keyless calls from every Cloud workflow share one anonymous daily limit. The n8n guide covers the HTTP Request route as well.

Any MCP client

Clients that support remote servers can use https://api.pressa.dev/mcp (Streamable HTTP) directly. Clients that only run local servers can launch npx -y @pressa/mcp (stdio). Neither needs an API key to start. With a key, set PRESSA_API_KEY for the local server, or send an Authorization: Bearer pressa_xxx header to the hosted one.

Available Tools

Without an API key only compile is registered. The other tools appear once a key is set, because a tool that cannot succeed is worse for the agent than an absent one.

compile

Compiles LaTeX source code to PDF and returns the result URL with metadata.

Parameters

Example: extracting a logo from a PDF and including it

Typical AI workflow when the user uploads a PDF that contains images:

1. Agent reads the PDF and extracts the logo as PNG bytes.
2. Agent base64-encodes the PNG.
3. Agent calls compile with assets: { "logo.png": "iVBORw0..." } and generated LaTeX that references \includegraphics{logo.png}.
4. Pressa compiles the PDF with the real logo embedded.

Save once, reuse many

If the user will reuse the same logo across many future compiles (monthly invoices, repeat contracts, recurring reports), save it to the library once instead of re-encoding the bytes every compile:

1. Agent calls save_asset with the base64-encoded logo bytes, name logo.png. Returns .
2. Every future compile passes use_stored_assets: ["logo.png"] instead of re-uploading the bytes. The LaTeX still references \includegraphics{logo.png} normally - the server loads the bytes from storage.
3. When the user wants to see or refresh what's saved, call list_assets. When they change brand, call save_asset again with the same name to upsert.

Returns

render

Render a saved V2 template (one with Liquid placeholders and a schema) by supplying just the data values - no LaTeX in the call. The data is validated against the template schema server-side, the compiler is auto-detected from the source, and stored assets referenced via \includegraphics resolve automatically. Prefer this over compile whenever the same document layout is reused. Paid plans only.

Parameters

Returns the PDF URL with page count, timing, the template name and version that was rendered, and the account's usage block.

save_template

Save or update a LaTeX template. If a template with the same name exists, it gets updated. Paid plans only.

Each template carries the LaTeX layout plus two optional fields with different roles. description is a short human-readable summary surfaced in list_templates for UIs. instructions is a longer prose markdown playbook (up to 50000 chars) the agent reads alongside the LaTeX when filling the template - defaults, workflow rules, edge cases, conditional logic. Use description for humans, use instructions for yourself.

Parameters

Returns

list_templates

List all your saved templates. Takes no parameters. Each row carries a has_instructions boolean so you can see which templates ship with an agent playbook before fetching.

Returns

get_template

Get a single template by ID or name. Returns the full LaTeX content alongside the instructions playbook (or null if the template has none) so the agent receives layout and rules in a single round trip.

Parameters

Returns

delete_template

Delete a template by ID or name.

Parameters

Returns

Confirmation message on success.

save_asset

Upload a binary file (logo, signature, diagram, embedded PDF) to the user's persistent asset library. Upsert by name. Paid plans only. Reuse the saved asset in any future compile via use_stored_assets: ["name"] - no need to re-encode bytes each time.

Parameters

Returns

list_assets

List all assets in the user's library. Takes no parameters. Use to discover names available for use_stored_assets or before uploading a replacement.

Returns

get_asset

Fetch a single asset's metadata plus its base64-encoded content. Accepts either numeric ID or filename.

Parameters

Returns

delete_asset

Delete an asset by ID or name. The database row and the underlying bytes are both removed. Future compiles referencing the deleted name via use_stored_assets will fail with asset_not_found.

Parameters

Returns

Confirmation message on success.

usage

Returns your current plan, compilation usage and plan limits. Takes no parameters. Requires an API key.

Returns

Environment Variables

How it works

When you ask your AI assistant to create a document, here is what happens behind the scenes:

1. The AI understands your request and generates LaTeX source code.
2. It calls the compile tool, passing the LaTeX to the Pressa MCP server.
3. The MCP server sends the LaTeX to Pressa's API, which compiles it on the server.
4. Pressa returns a PDF URL. The AI shares the link with you.

The entire process takes a few seconds. You never need to install LaTeX, write LaTeX yourself, or leave your AI tool.