# WilmAI
**Wilma, Finland's school system, with AI.** Two ways to use it, both free and open source:
| Assistant | How | Guide |
|---|---|---|
| **Claude Desktop** (Mac, Windows) | Download [`wilmai.mcpb`](https://wilm.ai/get/claude) and open it; Claude asks whether to install it | [wilm.ai#claude](https://wilm.ai/en#claude) |
| **ChatGPT desktop app** (Mac, Windows) | In Work mode (not Chat), send ChatGPT one message and it installs the plugin itself; or add the MCP server `npx -y @wilm-ai/wilma-cli@2 mcp` yourself | [wilm.ai#chatgpt](https://wilm.ai/en#chatgpt) |
| **Grok Bot** | Send one message; the bot installs the CLI and skill itself | [wilm.ai#grok](https://wilm.ai/en#grok) |
| **Meta Muse** (not in Finland yet) | Send one message; Muse installs the CLI and skill itself, and asks you to save your Wilma login in its vault | [wilm.ai#muse](https://wilm.ai/en#muse) |
| **OpenClaw** | Send one message and OpenClaw installs the CLI and the [`wilma`](https://clawhub.ai/aikarjal/skills/wilma) skill; a second one starts a daily briefing with [`wilma-triage`](https://clawhub.ai/aikarjal/skills/wilma-triage) | [wilm.ai#openclaw](https://wilm.ai/en#openclaw) |
| **Claude Code** | `claude mcp add --scope user wilma -- npx -y @wilm-ai/wilma-cli@2 mcp` | [wilm.ai#claude-code](https://wilm.ai/en#claude-code) |
| **Codex** | `codex plugin marketplace add aikarjal/wilmai`, then `codex plugin add wilma@wilmai` | [wilm.ai#codex](https://wilm.ai/en#codex) |
| **Terminal** | `npm i -g @wilm-ai/wilma-cli`, then `wilma login` | [wilm.ai#cli](https://wilm.ai/en#cli) |
Not supported: Claude and ChatGPT on the web and in the phone apps. Your login stays on your own devices and WilmAI runs no server in between; for updates on your phone, use an agent with its own computer, such as OpenClaw or Grok Bot. Not yet available in Finland: Meta Muse (US and Canada so far), Instinct, and OpenAI dots on consumer plans.
## What you can ask
In your own words; every question covers all your children.
- "What do my kids have going on at school this week?"
- "Is there any homework for tomorrow?"
- "Are there any exams coming up? What should she study?"
- "Has any teacher left a note about Eino this month?"
- "Any new messages from school? Read the attached letter too."
- "How did the last exams go? What was on the spring report card?"
Or give an agent a routine: "Every weekday at 7, post a school briefing for both kids to the family channel, and put new exams in the family calendar." The `wilma-triage` skill does exactly that.
## What's inside
- `packages/wilma-client` – TypeScript Wilma client (auth + parsing + tenant list)
- `packages/wilma-cli` – CLI (`wilma`), browser login (`wilma login`), and MCP server (`wilma mcp`)
- `skills/` – Agent skills (`wilma`, `wilma-triage`); also the plugin root for Claude Code and Codex
- `apps/site` – wilm.ai (Next.js, built as static files and served by Cloudflare Workers)
- `apps/home` – WilmAI Home, the desktop app (Tauri), and its eval suite
## Logging in
```bash
wilma login
```
Opens a one-time page on `127.0.0.1` in your browser. Pick your school's Wilma, log in, and the verified login is saved to `~/.config/wilmai/config.json` (or `$XDG_CONFIG_HOME/wilmai/config.json`; override with `WILMAI_CONFIG_PATH`). The password is entered on that page — never in an agent's chat.
Children on different Wilmas (say, a city school and a private school or lukio)? Press **Add another Wilma** on the same page, or run `wilma login` again. Every saved login is used together: every command covers all your children. `wilma accounts` lists the saved logins and `wilma accounts remove
- **Reads and writes.** Reply to teachers, write new messages, sign up for parents' evenings and report absences, without opening Wilma. Pin the messages you keep coming back to.
- **For a student too.** Sign in with a student's own Wilma account and choose "A student": the app then reads things from the student's side.
- **AI is optional.** No AI (the default), OpenAI with your own API key (usually tens of cents a month; the app stops at a monthly limit you set), or local AI on an Apple Silicon Mac with 16 GB of memory, where nothing goes to an AI service.
- **Stays on your computer.** No WilmAI server and no tracking. Your Wilma password goes only to your school's Wilma; the app keeps it (scrambled) only if you turn on staying signed in.
- **Updates itself.** New versions arrive in the app, signed with the project's update key; you choose when to install them.
Mac (Apple Silicon, macOS 13.3 or later) and Windows 10 or 11. The installers aren't signed by Apple or Microsoft yet, so the first start needs one extra step: on a Mac, System Settings → Privacy & Security → Open Anyway (the install window has a link); on Windows, More info → Run anyway. On Windows, notifications and local AI aren't there yet.
The source is in [`apps/home`](apps/home) (Tauri, React and Rust), with its own README for building and the eval suite that measures how well the AI reads school messages.
## Troubleshooting
### `fetch failed` on a managed or corporate laptop
If any command fails with a TLS certificate error, your network is very likely
intercepting TLS ("break-and-inspect") and re-signing certificates with a private
root CA — standard on devices behind a corporate secure web gateway. Node ships its
own CA bundle and ignores the operating system trust store, so a root your browser
already trusts is invisible to Node.
The CLI reports the underlying cause code (for example
`UNABLE_TO_GET_ISSUER_CERT_LOCALLY`) and the fix. In order of preference:
```bash
# 1. Node >=22.19 / >=24.6
NODE_USE_SYSTEM_CA=1 wilma summary
# 2. Node >=22.15 (the flag has wider version support than the env var)
node --use-system-ca "$(command -v wilma)" summary
# 3. Point Node at the proxy root explicitly
NODE_EXTRA_CA_CERTS=/path/to/root-ca.pem wilma summary
```
Never set `NODE_TLS_REJECT_UNAUTHORIZED=0`. It disables certificate verification
entirely, on a network that is already inspecting your traffic.
Two things that make this hard to recognise:
- **A working `npm install` proves nothing.** Inspection policies commonly exempt
`registry.npmjs.org`, so installing the CLI succeeds while every Wilma request
fails.
- **`curl` and your browser work fine**, because they use the OS trust store. Only
Node is affected, which makes the network look healthy.
## Credentials & Privacy
Your Wilma login is stored locally in `~/.config/wilmai/config.json` (or `$XDG_CONFIG_HOME/wilmai/config.json`), readable only by your user account. The password is obfuscated (not encrypted) for convenience — this is a personal productivity tool, not a vault. The current Wilma session is kept next to it (`wilmai-sessions.json`, same protection, at most 6 hours) so commands don't log in each time; set `WILMAI_NO_SESSION_CACHE=1` to turn that off. `wilma logout` ends the session in Wilma and removes it, keeping the login; `wilma config clear` removes both.
The CLI, the MCP server and the Claude Desktop extension have no server behind them: they log in only to your school's Wilma. Besides Wilma, they open links from school bulletins when you ask for an attachment (only public websites, never with your Wilma login), and once a day they ask npm whether a newer version is out (not when started through `npx`, which fetches the newest version itself; `WILMAI_NO_UPDATE_CHECK=1` turns it off). When an AI assistant uses WilmAI, the school information it reads goes to that assistant's provider, the same as pasting it into a chat; your password does not.
**Do not share your config file.** It accesses the same data as the official Wilma app or website; it is your responsibility to handle that data appropriately.
## License
MIT. Questions, bugs and ideas: [GitHub issues](https://github.com/aikarjal/wilmai/issues). If WilmAI is useful to you, a star helps other parents find it.