Your AI chat. Your machine. Your keys.
Local-first. No account. No telemetry. Your keys stay in your OS keychain. Your chats never leave your machine.
A serious AI chat app for people who'd rather their conversations weren't on someone else's server. You bring the model — a local one through Ollama or LM Studio, or a cloud one through your own API key — and the app gives it real tools: your files, a terminal, the web, a browser it can drive, knowledge, memory and a voice you can talk to.
Windows and Linux, with a browser build that needs no install at all. Ships with no API keys of its own.
Built with Tauri v2, React 19 and TypeScript, with a Rust backend.
Conversations — streaming chat with markdown, branching, editing, snapshots, folders, and automatic compaction so long chats don't run out of context. Projects group chats that share a brief, documents and memory.
Tools — files, terminal, web, browser control, media generation, and your own scripts in JavaScript or Python. Every tool is off until you enable it, and file access is confined to a working directory you choose.
Knowledge — index documents and the model searches them semantically, so it answers from your material rather than from what it half-remembers.
Memory — facts persist across conversations in three scopes (chat, project, global), capped at a fifth of the model's context window so recall never crowds out the conversation.
Voice — talk to it. Whisper for speech recognition and Kokoro for synthesis, both running locally; cloud engines (OpenAI, ElevenLabs, Cartesia, Groq) when quality matters most. Optional 3D VRM or MMD avatar.
Browser control — a real browser embedded in the conversation that the model can read and click, or an extension that drives the Chrome you already use, with the sessions you're already signed into.
Automation — save a role as an agent, chain steps into a workflow, run either on a schedule, or put several agents on one job at once.
MCP — connect Model Context Protocol servers over stdio or HTTP. Tools are disclosed progressively, so connecting ten servers costs the same startup context as connecting one.
Device mesh — pair the machines you own so they can share models, tools and knowledge. Discovery over the LAN, pairing by a code that never crosses the wire.
Models — any OpenAI-compatible endpoint plus Anthropic. Connect a subscription you already pay for (Claude Pro/Max, GitHub Copilot) as a backend, chain providers into combos that fail over on their own, and see which ones are rate-limiting you before you wonder why replies got slow. Compare models side by side, score them against your own eval set, and check public benchmarks before paying for one.
Open API — the app is also a server: an OpenAI- and Anthropic-compatible
endpoint on loopback, so Claude Code, opencode, Aider or any SDK can drive
your models, agents and combos — local ones included. hs endpoint prints
the configs; Settings → Devices shows them with copy buttons.
Value — live prices for ~6,600 AI models, plus VPS and GPU compute, read straight from the providers' own public price lists. No key and no account: these are published numbers, and every row links to the page it came from. Ranks on what a workload would actually cost — cache reads, cache writes, batch discounts, long-context tiers and rate-limit feasibility, not the sticker price.
Full documentation lives in docs-site/ — 40 pages including
end-to-end walkthroughs for real jobs: working with a codebase, research,
automating a recurring report, document processing, web scraping.
npm run docs:dev # http://localhost:5174
npm run docs:build # static site → docs-site/distStart with what this is, your first chat, or the use cases.
Prerequisites: Node.js 18+, Rust, and
a C++ toolchain — VS 2022 Build Tools with the C++ workload on Windows,
build-essential on Linux.
npm install
npm run tauri dev # run the app
npm run tauri build # produce an installerThe first Rust build compiles several hundred crates and takes a few minutes. Later builds are incremental.
npm test # 699 tests
npx tsc --noEmit # type-check app and docs site
# Price feeds change upstream; this checks the adapters against the real APIs.
PRICING_LIVE=1 npx vitest run tests/pricing.live.test.tsOne folder, plain JSON, no database:
~/.harnessx/
├── settings.json # providers (no keys), instructions, preferences
├── conversations/ # one JSON file per chat, plus an index
├── presets/ # saved prompt + parameter combinations
├── snapshots/ # chat snapshots
├── agent-memory/ # per-agent memory
├── avatars/ # imported VRM / MMD characters
├── models/ # models downloaded through the app
└── engines/ # Whisper, Piper
API keys are not in there. They go in your OS credential store — Windows Credential Manager, or Secret Service on Linux.
Worth stating plainly, because "AI app" usually implies otherwise.
Your prompts, files and documents go to the model provider you chose, and nowhere else. With a local model they don't leave at all. There is no account, no telemetry, and no server holding your conversations.
The one exception is public benchmark data, fetched through the gateway in
server/ so the Benchmarks panel isn't empty. It carries no key of
yours and no data about you — it's public information about models.
Provider keys, media keys and MCP credentials are deliberately not routed through that gateway. They're yours, they stay on your machine, and they go only to the service they belong to.
src/ |
React frontend — views, state, provider clients, tool implementations |
src-tauri/ |
Rust backend — audio capture, MCP, browser bridge, mesh, keychain, speech |
extension/ |
Chrome MV3 extension for driving the user's own browser |
server/ |
Gateway for shared benchmark data (optional; not needed to run the app) |
docs-site/ |
Public documentation — markdown plus a small React renderer |
docs/ |
Internal planning notes and release process |
tests/ |
Vitest suite |
Early software, used daily, with rough edges — the documentation says where they are rather than glossing over them. Two worth knowing before deploying it widely:
- Not code-signed. Windows shows an unsigned-application warning on first launch.
- The device mesh is not encrypted yet. The handshake protects credentials and blocks replay, but request bodies are plaintext. Fine on your own network; across the internet, put it inside a VPN or tunnel.
See docs/roadmap.md for what's planned.