curl -fsSL https://chalie.ai/install | bash
chalie # → http://localhost:31025
Beta — on purpose. The bar is software you'd trust with your own life's admin, and it isn't all the way there yet. Hit a sharp edge? Open an issue — we respond fast.
Why Chalie is different
Most AI tools forget you the moment you close the tab. Chalie runs on your own machine as a reasoning engine that keeps working while you step away: it remembers what matters and lets the rest go, and acts only behind an Allow / Ask / Deny policy spanning you, its own background work, and other agents. One SQLite file, credentials encrypted at rest, zero telemetry, encrypted whole-instance backup — no Redis, no Postgres, no queue, just one Python process.
What it can do today
| |
|---|
| 🧠 Self-managing memory | A background pass distills conversation into durable facts and remembered moments — recalled automatically, never something you save by hand. |
| 🎯 Goals & proactive research | Spots goals from casual mentions; researches topics in the background before you ask. |
| 👁 Vision | Reads photos, screenshots, and scans — and indexes them so you can find an image by what's in it. |
| 🌐 Real web browsing | Drives a live browser: clicks, fills forms, scrolls, and inspects its own screenshots. |
| 🔌 MCP, in and out | Connects to remote MCP servers and exposes its own tools to other agents. |
| 📬 Email, calendar, contacts | IMAP, CalDAV, CardDAV — the accounts you already have. |
| 🗓 Scheduler | Natural-language recurring jobs. |
| 🧰 Files, shell & code | Searches files, runs guarded shell commands, and delegates coding tasks to an agent that writes and runs TypeScript in its own persistent, sandboxed workspace. |
| 🎙 Voice, fully local | Moonshine STT + Kokoro TTS, both ONNX. No cloud transcription, ever. |
| 💾 Backup & restore | Snapshot the whole instance to one file, optionally AES-256 encrypted. |
Which model drives it best?
We ran ten models — frontier and open-weight — through the same battery of real tasks and scored how well each drives Chalie. A 31B open model you can run under your desk landed in the front pack, ahead of a 550B one. Size wasn't the story; the chassis was. → Read the benchmark
Install
Fastest start — the wizard picks your provider on first boot:
curl -fsSL https://chalie.ai/install | bash
chalie # → http://localhost:31025 · Anthropic, Gemini, OpenAI, +17 more, or any OpenAI-compatible host
Fully local with Ollama — nothing leaves your network:
ollama pull gemma4:31b # the open model that placed in the benchmark's front pack
From source:
git clone https://github.com/chalie-ai/chalie.git && cd chalie
uv pip install --system -e backend/ # all dependencies, voice (TTS/STT/VAD) included · Python 3.11+
./run.sh # → http://localhost:31025/on-boarding/
Full walkthrough → chalie.ai/guide/installation.
Supported systems
| Platform | Status |
|---|
| macOS — Apple Silicon | Supported. Intel Macs are not: onnxruntime no longer publishes wheels for them. |
| Linux — apt (amd64/arm64) | Supported. Verified on Debian 12 and Ubuntu 24.04. |
| Linux — dnf (amd64/arm64) | Supported. Verified on Fedora 40 and 41. |
| Anything else | The installer refuses outright rather than leaving you a half-installed instance. Docker runs anywhere → installation guide. |
Every supported platform is re-proved on each merge to main. installer/verify runs the real installer on a pristine machine, boots Chalie, and passes only once readiness, the web interface, Chromium, a Kokoro → Moonshine voice round trip, and Deno all check out.
System requirements
What the runtime itself needs — a model's own requirements are separate:
| |
|---|
| Python 3.11+ | Already installed. The installer checks for it and deliberately never installs it, so distributions shipping something older — Ubuntu 22.04 (3.10), AlmaLinux 9 (3.9) — are refused until you supply one. |
Root or sudo | Linux only, for the system build packages and the CLI. macOS needs neither. |
| ~2 GB RAM | Resident set measured at 1.5–1.7 GB once the voice and embedding models have warmed up. Verified on a 4 GB machine; below that is untested. |
| ~3 GB disk | The Python virtualenv and native wheels, the Chromium build Playwright downloads, the local voice models, and Deno. |
| Port 31025 | The default; chalie --port=9000 moves it. |
How it's built
One Python process. Flask + flask-sock WebSocket. SQLite with sqlite-vec and FTS5 for semantic + lexical recall. Vanilla ES6 modules on the frontend — no build step. The whole cognitive runtime — memory, goals, scheduler, voice — runs as daemon threads inside that one process, and different cognitive functions can use different models.
Curious? → Architecture & internals · Schema
CLI
chalie # start
chalie --port=9000 # custom port
chalie stop # stop the daemon
chalie update # update to latest
chalie logs # tail the log
Philosophy
- Judgment over activity. Fewer high-confidence actions beat many low-confidence ones.
- Restraint builds trust. Every token, notification, and action earns its place.
- Continuity is intelligence. The persistent runtime — not any single response — is the product.
- Constraints are features. Forgetting on purpose, hard limits, and token budgets are what make wisdom emerge.
Full product compass and design rationale live at chalie.ai/docs.
Documentation
Full documentation lives at chalie.ai.
| |
|---|
| 📘 Guide | Setup, providers, and day-to-day usage. |
| 🏗 Docs | Technical reference and architecture. |
| 🤝 Contributing | How to get involved and ship a PR. |
Community
chalie.ai · Releases · Blog · X · Reddit
Contributing
PRs welcome. Start with the contributor handbook — vision, principles, and mechanics in one index — or open an issue. Found a security issue? Read SECURITY.md.
License
Apache 2.0 — use it, fork it, ship it.
