LibreOffice and Ollama: LocalWriter, AI in Writer

LibreOffice has no built-in AI assistant, but a third-party extension can connect Writer to a model served by Ollama on your machine. This guide starts from LocalWriter, an open source project published on GitHub: where to find it, how to point it explicitly at http://localhost:11434, what permissions it gets, and what can break. The LibreOffice + Ollama combination doesn't guarantee privacy on its own: it all depends on the server address you enter in the extension.
#Why connect LibreOffice and Ollama
Writer is the writing tool of choice for many individuals, nonprofits, and government agencies. Commercial office suites now offer writing assistants, but they send your text to an online service. With Ollama installed on the same computer, an open-weight model can rephrase, extend, or correct a passage without the document leaving your machine, provided the whole chain is configured correctly end to end.
The principle is simple. Ollama runs in the background and responds to HTTP requests on port 11434. A LibreOffice extension grabs the selected text, sends it to that address along with a prompt, then inserts the response into the document. The extension contains no model: it acts as an intermediary between Writer and a server. That's why the server address determines everything, including privacy.
This setup is suited to short, verifiable tasks: rewording a paragraph, suggesting a continuation for a draft, changing the tone, summarizing a section. It doesn't replace proofreading. A local model makes mistakes just like an online model, and often more if it's small.
#LocalWriter: what the extension does
Benefit from our experience and a guided path to build your local assistant and use your AI every day.
- Lifetime online access
- PDF + files
- Refunded within 30 days
LocalWriter is a third-party LibreOffice extension (from the GitHub account balisujohn) and isn't affiliated with The Document Foundation or the Ollama team. When we checked on October 30, 2026, the README described two main uses in Writer: continuing the selected text, and editing the selection based on an instruction you type. The extension connects to a server that's compatible with the OpenAI API. Ollama provides one, as do other local servers.
- Nature
- LibreOffice extension file (.oxt), distributed on the Releases page of the GitHub repository.
- Described features
- Extend a selection; edit a selection based on an instruction. Check the README for the exact menu labels, any keyboard shortcuts, and features added since (for example, for Calc).
- Target server
- Any server that exposes an OpenAI-style API: Ollama, but also other local engines or an online service. Nothing requires a local server: you enter the address yourself.
- Status
- Community project. No one guarantees the update cadence, maintenance, or compatibility with new LibreOffice versions. Check the date of the last commit and the open issues.
Other extensions or macros have the same goal. The principles in this guide (explicit address, permission review, compatibility checks) apply to all of them: don't treat LocalWriter as a recommendation, but as a documented example.
#Prerequisites
- LibreOffice
- A recent version installed from libreoffice.org or your distribution. Note the exact version number (Help > About LibreOffice): you'll need it to check the compatibility the extension claims.
- Python in LibreOffice
- Extensions written in Python require LibreOffice's Python support. It's included in the official Windows and macOS installers; on some Linux distributions, it's a separate package (see step 2).
- Ollama
- Installed and running, listening on http://localhost:11434. If it isn't, follow the installation guide for Ollama first (link at the bottom of the page).
- A model that fits your memory
- Rough figures at Q4_K_M: a 3B uses about 2 GB of VRAM, a 7B about 5 GB, a 14B about 9 GB. A 12 GB card like the RTX 3060 comfortably handles a 7B to 9B.
- A test document
- A file with no sensitive data for your first tests, and a backup of your important documents before installing any extension.
#Step 1: Set up Ollama and check the API
Before touching LibreOffice, make sure Ollama responds on its OpenAI-compatible API. If this step fails, the extension will fail too, and you'll know the problem isn't the extension. First, download a reasonably sized model. The two tags below were listed in the Ollama library on October 28, 2026; replace them with a model you've already installed if you prefer.
Next, query the API the same way the extension will. The first command lists the models visible to the API; the second sends a real chat request. A JSON response containing a choices field means the server is ready.
Also check which address Ollama is listening on. By default, it only accepts connections from the machine itself (127.0.0.1), which is the right setting for use in Writer on the same computer. If someone has set OLLAMA_HOST=0.0.0.0, the server is reachable from the local network: that isn't needed here.
#Step 2: install the extension in LibreOffice
- 01Read the repository before downloadingOpen https://github.com/balisujohn/localwriter. Read the entire README, then the Releases page: latest version number, date, release notes, and the LibreOffice versions mentioned. Also check the open issues: a known incompatibility with your version often shows up there before it gets documented.
- 02Download the .oxt file from ReleasesUse the .oxt file attached to a published release, not a copy found on some other site. If the project publishes a SHA-256 checksum, compare it against the downloaded file; otherwise, just note the version number and download date.
- 03Inspect the contents (optional but useful)An .oxt file is a zip archive. Listing its contents shows you the Python scripts LibreOffice will run, and therefore what the extension can do (see the Permissions section).
- 04Install with the Extension ManagerIn LibreOffice: Tools > Extension Manager > Add, then select the .oxt file. LibreOffice may show a warning about the extension's origin or ask whether to install it for you only or for all users: on a shared computer, choose "only for me."
- 05Fully restart LibreOfficeClose all LibreOffice windows, including the Quickstarter in the system tray if it's running, then relaunch Writer. A new menu or new menu items should appear; the README tells you exactly where.
On Windows, unopkg is in LibreOffice's program folder (for example, C:\Program Files\LibreOffice\program). On Fedora, Python support comes from the libreoffice-pyuno package. The Flatpak and Snap versions of LibreOffice bundle their own environment. If you run into problems, compare with your distribution's packaged version or the one from libreoffice.org.
#Step 3: explicitly configure it to use your local Ollama
The extension opens a settings window (the README describes where to find it). This is the most important step in the guide: each field determines where your text goes and how the model handles it. Don't leave any default value you haven't read.
- Server address
- Enter http://localhost:11434 or http://127.0.0.1:11434. Depending on how the extension builds its requests, it expects either this root or the address with the /v1 prefix: follow the example in the README and, if you get a 404 error, try the other form.
- Model name
- Exactly the name shown by ollama list, including the tag (qwen3.5:9b, not qwen3.5). An incorrect name produces a "model not found" error on the Ollama side.
- API key
- A local Ollama doesn't require one. If the field is mandatory, any value usually works; never paste a real online service key there for local use.
- System prompt
- If the extension lets you edit the prompts, write them in French and keep them short: “You only edit the passage provided. Reply in French, with no commentary or quotation marks.” This limits wordy responses being inserted into the document.
- Maximum length
- If the token limit is too low, the response gets cut off mid-sentence. If it's too high, the model may ramble on. Start with a modest value and adjust.
For your first try, open the test document, select a sentence, and run the edit function with a simple instruction, such as "Make this paragraph more formal." While it's generating, running ollama ps in a terminal shows the loaded model: that's direct proof the request actually reached your local Ollama and not somewhere else. The first call may be slow while the model loads into memory.
#Permissions and privacy: what the extension can do
A LibreOffice extension isn't sandboxed like a browser extension. Its Python scripts run with your user account's privileges: they can read and write your files, open network connections, and read the contents of the open document. No permissions dialog spells out these rights during installation. Installing an .oxt therefore means trusting its author and the version you downloaded.
- The code is readable
- LocalWriter is open source, so you can review the scripts listed in step 2. Look for hard-coded network addresses and HTTP request functions. A trustworthy extension only calls the address you entered.
- Updates matter too
- A new version can change behavior. Read the release notes before updating, and avoid unreleased development builds for real-world use.
- The address determines privacy
- With http://localhost:11434, the selected text stays on your machine. With another PC's address, it travels across your network. With the address of an OpenAI-compatible online service, it goes to that provider, under its terms.
- Ollama-hosted models
- Recent versions of Ollama also offer models that run on ollama.com servers, accessible with an account. If you pick such a model in the "model" field, your text is sent online even though the address is still localhost. Check the model's page to confirm it's a local model, and that ollama ps shows it loaded on your machine.
#Compatibility limits to be aware of
Several layers must line up: the LibreOffice version, its Python interpreter, the extension, and the Ollama version. None of them is controlled by the other projects, and this guide hasn't tested any combination. Here are the most likely friction points, to be checked against the README and the repository's issues.
- LibreOffice version
- An extension tested on one branch may break on the next if the internal API changes. Compare your version with the ones mentioned in the repository. If none are listed, treat compatibility as unknown.
- Python bundled or missing
- LibreOffice bundles its own Python on Windows and macOS, and often uses the system Python on Linux. An extension that imports a missing library will fail to load.
- API format
- The extension speaks the OpenAI dialect. Ollama implements it under /v1, but not necessarily every option. An unsupported option may be silently ignored.
- Complex documents
- Selections that span tables, footnotes, fields, or tracked changes may lose their formatting: the model receives plain text and returns plain text.
- Length
- The selection and the prompt must fit within the context window configured on the Ollama side. Beyond that, the beginning of the text may be truncated without warning. Work one paragraph at a time.
- Calc, Impress
- Don't assume a feature designed for Writer works elsewhere: only the applications listed in the README are supported.
- Frozen interface
- Depending on the implementation, Writer may stop responding while the model is generating. With a model too large for your VRAM, the wait gets significantly longer: choose a model that fits entirely in GPU memory.
#Troubleshooting LibreOffice with Ollama
- 01No menu appears after installationFully restart LibreOffice, including the Quickstarter. Check in Tools > Extension Manager that the extension is listed and enabled. On Linux, install Python support (libreoffice-script-provider-python or libreoffice-pyuno), then try again.
- 02"Connection refused" or timeoutOllama isn't running or is listening on a different address. Rerun the curl commands from step 1: if they fail, the problem is on the Ollama side. If they succeed, compare the address entered in the extension character by character (http, not https; port 11434).
- 03404 ErrorThe path doesn't match: the extension may already append /v1, or conversely expect it in the address. Try the other form given in step 3.
- 04“model not found”Copy the exact name shown by ollama list, including its tag. If the model isn't installed, run ollama pull.
- 05Response in English, wrapped in commentary or quotation marksMake the system prompt stricter (reply in French, text only), or switch models. Small models are worse at following this kind of instruction.
- 06Writer very slow or frozenRun ollama ps: if the PROCESSOR column shows a CPU share, the model is spilling out of VRAM. Switch to a smaller model or a lighter quantization, and close any other loaded models.
#Go further
This guide focuses on the bridge between Writer and Ollama. Installing the server, sharing it over the network, and the legal framework each have their own guide.
- Install Ollama
- Step-by-step installation on Windows, macOS, and Linux, prerequisite for step 1: https://quelllm.fr/guide/installer-ollama
- Share Ollama on your local network
- If Writer runs on a different machine than the server, see how to set OLLAMA_HOST and secure access: https://quelllm.fr/guide/ollama-serveur-reseau-local-partage
- Local LLM and GDPR
- What local processing changes, and doesn't change, for personal data in a business: https://quelllm.fr/guide/llm-local-donnees-privees-rgpd
- Sources
- LocalWriter repository: https://github.com/balisujohn/localwriter (accessed 30/10/2026, not rechecked on 17/11/2026); Ollama OpenAI-compatible API documentation: https://docs.ollama.com (not rechecked on 17/11/2026). Check them before you install: versions and features change.
Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.