Nelson MCP — LibreOffice

Nelson MCP

A LibreOffice extension that turns your documents into an MCP server. External AI clients connect over HTTP and get full access to document tools — reading, editing, navigating, formatting, and more.

Works with any MCP-compatible client: Claude Code, OpenCode, Goose, ollmcp, etc.

How it works

Nelson MCP runs an HTTP server inside LibreOffice and speaks the Model Context Protocol. AI agents connect to it and use tools to interact with your open document — no copy-paste, no file export.

┌─────────────────┐     HTTP/MCP     ┌────────────────┐

│ AI Client │ ───────────────► │ LibreOffice │ │ (Claude Code, │ tools/call │ + Nelson MCP │ │ OpenCode…) │ ◄─────────────── │ │ └─────────────────┘ tool results └────────────────┘

What it looks like

An agent writes a quarterly report into an empty Writer document, one tool call at a time:

The finished report: the content as Markdown, heading colours with style_set, a shaded table, a Page {page} of {pages} footer, and a comment on the hiring line. The Nelson MCP sidebar logs every call. Select a call to see its caller, status, duration and parameters.

Everything is set in Tools > Options > Nelson. For example, the HTTP page sets the port, the bind address and the access token:

Features

• 145 document tools — read content, edit text, manage styles, insert images, handle tables, charts, conditional formatting, hyperlinks, track changes, page headers/footers, navigate headings, search, and more. The list is filtered by the active document type, so a client sees only what applies (97 with a Writer document, 50 with Calc)
• Self-configuring agents — with the config API on, an agent reads and changes Nelson's settings itself (read limits, exchange format, the name on its changes), while the settings that protect you stay reserved to Options. See Let the agent configure Nelson
• Custom MCP endpoints — expose only the tools your agent needs. Built-in presets (minimal, writer-edit, writer-read, calc, gallery) or create your own filtered endpoints
• Writer, Calc, Draw, Impress — tools adapt to the active document type
• Calc `=PROMPT()` — call an LLM directly from a spreadsheet cell
• Built-in launchers — launch Claude Code, Gemini CLI, or OpenCode directly from LibreOffice with one click. Nelson handles MCP config, prompt injection, and working directory setup automatically
• AI image generation — generate and edit images from text prompts using Stable Diffusion (A1111/Forge), OpenAI, or AI Horde. One-click detect/install/launch for Automatic1111
• Beginner-friendly setup — all tools come with install buttons, auto-detection of existing installations, and guided configuration. No manual config files to edit
• Tunnels — expose the MCP server externally via ngrok, Cloudflare, bore, or Tailscale, behind an access token
• SSL — optional HTTPS with auto-generated certificates
• Modular — each feature is a self-contained module with its own config, services, and tools

Install

1. Download the latest .oxt from the releases page
2. In LibreOffice: Tools > Extension Manager > Add
3. Restart LibreOffice
4. The MCP server starts automatically (default: http://localhost:8766/mcp)

Quick start

Once installed, point your MCP client at the server:

{

"mcpServers": { "nelson": { "type": "http", "url": "http://localhost:8766/mcp" } } }

Open a document in LibreOffice, then ask your AI client to read or edit it.

For AI agents: see QUICKSTART.md — a step-by-step guide for LLM agents on how to discover documents, navigate structure, and use tools effectively.

Let the agent configure Nelson

An agent doesn't have to stop and ask you to open Options. With the config API switched on, it reads and changes Nelson's settings itself, over the same HTTP server, and changes apply at once. For example, it can:

• raise the read limit before a long document;
• switch to HTML to rewrite a formatted one;
• sign its tracked changes with its own name.

Switch it on in Tools > Options > Nelson > Http > Enable Config API, then:

# Read one setting, a module, or everything

curl 'http://localhost:8766/api/config?key=writer.max_content_chars' curl 'http://localhost:8766/api/config?module=writer'

Change several at once

curl -X POST http://localhost:8766/api/config \ -H 'Content-Type: application/json' \ -d '{"writer.max_content_chars": 200000, "writer.change_author": "AI agent"}'

Add -H 'Authorization: Bearer ' once an access token is set.

The API tunes how Nelson behaves. It can't undo what protects you. These settings are reserved to Options, and a request that touches one is refused with 403 and changes nothing:

• http.*: the token, the address, allowed origins, SSL, and the config API switch itself;
• tunnel.* and debug.*;
• launcher.*: the commands Nelson runs;
• *.instances: the folders tools can reach, and the AI providers' endpoints and keys;
• core.force_track_changes.

Secrets such as the token and API keys read back as ***.

Security

Whoever can talk to Nelson can read and edit every document open in LibreOffice, and open any file your account can read. What stands between that and other people:

Set the token in Options > Nelson MCP > HTTP > Access Token; it applies without a restart. The config API can't change it or any other protective setting (see above), but it still reads and changes the rest: keep it and the debug API off on any machine that is reachable from elsewhere.

Before exposing Nelson through a tunnel, read the warning at the top of docs/howto/connect-chatgpt-tailscale.md.

Modules

Around 30 modules; the main ones:

Turning off what you do not use

Nelson ships a lot of tools, but a client never sees all of them, and most of the narrowing happens on its own. In order of how much they cut:

So a Calc user is already down to 48 tools without touching anything, and a custom endpoint takes that to whatever you like.

What Nelson deliberately does not have is a switch to stop a module loading. It would be a seventh overlapping control, and it would buy very little: a full startup is ~370 ms for 31 modules, and the optional families are already invisible in the tool list. See #29 for the measurements.

Development

./install.sh              # Set up dev environment

make deploy # Build + install + restart LO + show log make test # Run tests make release # Tag, build and publish a release (see scripts/release.sh)

See DEVEL.md for the complete developer guide and docs/modules.md for the module framework reference.

Developing with an AI agent

This project is developed with AI coding agents, using [wbox-mcp](https://github.com/quazardous/wbox-mcp) to make that practical.

Nelson lives *inside* LibreOffice, so an agent cannot verify a change without actually running LibreOffice — and a headless process cannot show whether a logo really landed in the page header. wbox-mcp is an MCP server that hands the agent a sandboxed LibreOffice in a nested Wayland compositor: it can kill, deploy and relaunch the extension, drive the UI, read the Nelson log, and take screenshots of the result.

That closes the loop. The agent edits the code, redeploys, calls the tool through MCP, and *looks* at the rendered document to confirm the change — instead of assuming it worked.

Documentation

• QUICKSTART.md — agent guide: discovery, workflows, tool patterns
• AGENTS.md — developer cheatsheet: project structure, build pipeline, critical rules
• docs/howto/ — user guides: Tailscale + ChatGPT, Forge, Ollama indexation, image galleries
• docs/roadmap.md — what's planned next
• CHANGELOG.md — version history

Acknowledgments

Nelson MCP is the result of merging and reworking two other projects:

• **[LocalWriter](https://gi