A web application for writing novels and short stories that you run on your own hardware or network. Named after the Celtic goddess of poetry.
Brigid keeps your manuscript as a structured outline rather than one long file, so you can move a chapter without cutting and pasting twenty pages. It handles the things a novel needs and a word processor doesn't: parts and chapters and scenes that know what they are, formatting decided once and applied everywhere, a canvas that lays the whole shape of the book out in front of you, compiling to a submission manuscript, and — if you want it — a local AI that reads the book and tells you what shape it has, maps your characters to archetypes, and lets you chat about your work.
Nothing leaves your machine or server until you're ready to compile into a Word document or a PDF. There is no account with anyone, no subscription, and no company that can change the terms later. If you use the AI features, the model runs on hardware you control too.
This is a vibe-coded app. Claude did most of the heavy lifting.
A computer to run it on. A spare Linux box, a Mac, or a Windows machine. It can be the same computer you write on, or another one on your network that you reach from a browser. It needn't be powerful — unless you want the AI features, which want a decent graphics card.
About twenty minutes, most of it waiting for downloads.
Enough comfort with a terminal to open one, paste a command, and read what comes back. You don't need to know what any of it means. The installers do the work and tell you what they did.
Pick your platform. Each installer does the same job: installs what Brigid needs, sets up its database, builds the application, and arranges for it to start automatically when the computer boots.
All three are safe to run twice. If one stops partway with an error, fix what it complained about and run it again — it checks what is already done and skips it.
git clone https://github.com/jpmoo/brigid.git
cd brigid
./install.shIt will ask for your password once or twice. Installing software and setting up a background service both need administrator rights.
You need Homebrew first — that page has a one-line install if you don't have it.
git clone https://github.com/jpmoo/brigid.git
cd brigid
./install-macos.shOpen PowerShell as Administrator (right-click it, choose "Run as administrator"), then:
git clone https://github.com/jpmoo/brigid.git
cd brigid
Set-ExecutionPolicy -Scope Process Bypass -Force
.\install-windows.ps1It asks for the PostgreSQL password — the one set when PostgreSQL was installed. If the script installs it for you, that's the password you were prompted for during that step.
The Windows installer is newer than the other two and has had less real-world use. If it goes wrong, the Linux and macOS paths are better trodden.
When the installer finishes it prints an address, something like:
http://192.168.1.40:8090
Open it in a browser. The first visit asks you to create an account — one username and password, yours, stored on your own machine. Nothing to confirm by email, nothing to verify.
Then make a manuscript and start writing.
From another computer on your network, use the address the installer printed
rather than localhost. If it doesn't load, see When something's wrong.
The outline down the left is always the same manuscript. What changes is how the middle shows it, and you switch with the three buttons above it.
Book sets your writing the way a novel is set: a comfortable measure, generous leading, nothing on the page but the words. This is the one to write in.
Manuscript sets it exactly as your submission templates specify — double spaced, the right indents, the right font — so what you see is what will come out of the compiler. Useful when you are close to sending it somewhere.
Clicking into a section in either one starts writing there. The style bar carries bold, italic, and the rest; spelling is checked as you go, and the search box walks every match in the whole manuscript rather than just the section you are in. Zen (the expand button) drops the header and floats the outline, so there is nothing left but the page.
Canvas shows the shape of the book rather than its text: every part, chapter, and scene as a card on an endless surface, nested inside the thing that contains it, with arrows for the order they are read in.
Anything with something inside it is drawn as a container rather than a card, and its own writing — a chapter's opening, before its first scene — gets a card of its own at the top of it. So a chapter is a rectangle holding its opening and then its scenes, with the sequence running through all of them.
The arrows are worked out from the outline every time it is drawn, so reordering a chapter redraws them immediately — there is no stored connection that could quietly disagree with the book. Dragging never re-parents: a scene dropped inside another chapter's rectangle has been moved on screen, not moved in the book. Structure stays the outline's business, so an accidental drag can't rewrite it.
- Getting around. Drag the background to pan, or scroll with two fingers. Pinch or Ctrl-scroll to zoom, or type a percentage. The dot grid can be turned off.
- Arranging. Drag a card to move it, drag a corner to resize it. Chapters grow and shrink around their scenes as you go. Everything is remembered, and Reset (hold to confirm) throws the arrangement away and lays it out fresh.
- Notes. Drag from a tab on any side of a card to hang a note off it, and drop it wherever you like — a dotted arrow keeps it pointed at the section it belongs to. Notes are bookmarks — the same list at the top of the outline — so one made here appears there, and one dropped on a line while writing appears here beside the section holding it.
- Color. Cards shade against their level's word goal, the same reading the outline gives — short of it, or past it. The goal belongs to the chapter, so it is the chapter that shades and not the card holding its opening.
- Searching. There is no reading order to scroll along, so search works differently here: every card holding the term lights up where it sits and the count is of sections rather than occurrences. Open one and you see all of its matches at once.
- Writing. Double-click a card to open that section in a room of its own — zen without the outline, the same editor with all the same features.
Brigid runs in the background and starts itself when the computer does. You shouldn't need to think about it. When you do:
| Linux | macOS | Windows | |
|---|---|---|---|
| Update | ./restart.sh |
./restart.sh |
see below |
| Restart | sudo systemctl restart brigid |
sudo launchctl kickstart -k system/com.brigid.app |
Restart-ScheduledTask -TaskName Brigid |
| See logs | journalctl -u brigid -f |
tail -f data/brigid.log |
Task Scheduler → Brigid |
| Stop it | sudo systemctl stop brigid |
sudo launchctl bootout system /Library/LaunchDaemons/com.brigid.app.plist |
Stop-ScheduledTask -TaskName Brigid |
Updating on Windows:
git pull; pnpm install; pnpm build:web; pnpm db:migrate; Restart-ScheduledTask -TaskName BrigidBrigid backs itself up nightly — by default at 1am, keeping the last ten. You can change that, take one immediately, or download one, under Settings → Backup. Restoring can bring back a single manuscript or everything.
Keep a copy somewhere else as well. A backup living on the same disk as the thing it is backing up is not really a backup. Downloading one now and then and putting it wherever your other important files go is enough.
Optional, and off until you set them up. Brigid can read your manuscript and tell you about its shape, its characters and the way you write. It talks to a language model running on your own hardware, so the manuscript is never sent anywhere.
Brigid talks to a language model you run yourself. It does not care which server you run it with — give it an address and it works out what is answering.
- Install a model server on the machine with the best graphics card — the same one as Brigid, or another on your network. Ollama is the easiest starting point; llama.cpp, LM Studio and vLLM all work too, as does anything else serving the OpenAI shape.
- Give it a model to serve. With Ollama that is
ollama pull qwen2.5:14b, or any model you prefer; with the others it is usually an argument you start the server with. - In Brigid, open Settings → AI Model, enter the address, and press Connect.
What happens next depends on what answered. With Ollama you get a list of the models you have installed and pick one, and Brigid reads how large a window that model can hold and uses all of it — rather than the much smaller default Ollama would otherwise serve, which truncates every chapter silently. With anything else, whatever the server is already serving is what will answer, because that was settled when you started it.
If your server wants an API key, there is a field for one. llama.cpp needs
none; vLLM started with --api-key does. It is stored in your own database in
the clear, next to the manuscript — which is fine for a box on your network and
is not a secret store.
Whatever address you give is where your manuscript goes when you use the AI features. On a machine you run, it does not leave your network.
Once connected, Brigid reads the manuscript in the background, section by section. On a novel that takes roughly an hour. You can carry on writing while it happens, and it keeps up with your edits by itself.
When it finishes, three tools appear under Project Settings → AI.
Story shape rates the manuscript against seven narrative frameworks — Hero's Journey, Freytag's Pyramid, Three-Act, Save the Cat, the Story Circle, Kishōtenketsu, and Seven-Point — judged on where your turns actually fall rather than on whether they exist at all. A weak fit to all seven is a legitimate finding, not a failure.
Characters profiles each character on ten role axes drawn from Vogler and Propp — Hero, Mentor, Shadow, Trickster, Ally and so on — as a radar chart whose shape describes their function in the story. Before anything is profiled you review what the reading gathered: every recorded action can be moved to whoever actually did it, reworded, or thrown out. Nothing is scored on a record you haven't approved.
Chat discusses the manuscript with you, drawing on everything above plus the actual passages your question bears on — so it can talk about your sentences as well as your structure.
The model is a reader, not an authority. It will sometimes be confidently wrong about your book, and it has no idea what you are trying to do.
Every score above the lowest is backed by the specific events it rested on, and those are shown so you can check them. That is deliberate: a profile whose evidence you cannot inspect is a profile you should not trust. When something looks wrong, look at what it rested on — often the answer is that the reading misattributed an action, which you can correct.
Treat all of it as one attentive reader's opinion, which is exactly the weight it deserves.
It doesn't load from another computer. Brigid is probably running fine and the machine's firewall is in the way. The Linux and Windows installers try to open the port; macOS may prompt you to allow incoming connections the first time, and you have to say yes.
It worked yesterday and doesn't now. Look at the logs — the table above says how. The usual cause is PostgreSQL not having started; restarting the machine generally sorts it, and Brigid waits for the database rather than giving up.
The AI says the model timed out. Usually the model is bigger than the graphics card can comfortably hold, so it spills into ordinary memory and runs many times slower. A smaller model is the fix.
The reading is taking forever. Roughly a minute a section, so a long novel is an hour or more. You can stop it and pick it up later — Characters → Stop reading — and nothing already read is lost.
Anything else. The logs almost always say. They are in plain English, and the last twenty lines are usually enough.
In PostgreSQL, on your own machine, in a database called brigid. The
application's settings live in .env.local and data/brigid.config.json inside
the folder you cloned.
Nothing is in the cloud, because there isn't one. That is the point of Brigid, and it is also why the backups are your responsibility rather than someone else's.
To do it by hand, or to run Brigid behind a reverse proxy — including under a
subpath like example.com/brigid — see docs/deploy.md, with
the systemd unit in deploy/brigid.service.
Brigid is a pnpm workspace: a Fastify server and a React client, TypeScript throughout, PostgreSQL underneath with hand-written numbered migrations. The design notes are in docs/brigid-spec.md.
pnpm dev:server # and, in another shell, pnpm dev:webVite proxies /api to port 8090, so the session cookie behaves exactly as it
does in production.
MIT — see LICENSE. Use it, change it, run it, share it. It comes with no warranty, which for something holding a novel is worth reading as: keep your own backups.