A shell that holds AI, filled with mini-apps you can switch on and off. Conversation is the entry: say what you need, and the agent decides whether to query your data, produce a card, or run a skill you've saved — not another chat box, but a personal workbench that grows on your own data.
Try without installing · Data stays in your browser · MIT open source
The agent main view — conversation as the entry
Product
Product
Status
v0.1 · Public beta
Updated
July 21, 2026
Most AI tools keep your data on someone else's server; privacy rests on a promise not to read, instead of the absence of a path to read.
Local-first, inverted: data lives on your machine by default (SQLite on desktop), keys go only into the system keychain — never files, never logs; private fields sit in a separate store the AI structurally cannot reach. Even which business data the AI may query passes three gates: the app is on, the app declared it, and you granted it by hand.
Flow
Entry
Say what you need
Conversation is the entry; a slash-command palette and quick skills run on one tap.
Decision
The agent decides what to do
Query your data, produce a structured card, or run a skill you've saved.
Canvas
The canvas appears only when needed
The right-side canvas shows up on demand; switch an app off entirely and the data stays.
01
Agent entry.
Driven by natural language; slash-command palette and quick skills run on one tap.
The slash palette: /match, /resume, /interview… one tap to run (live web demo)
02
Capability center.
Connectors (MCP), memory, knowledge base, Skills, scheduled tasks, projects — one place to see and manage.
Capability center: connectors (MCP), tools, and memory in one place — the AI never sees endpoints or keys
Tauri 2 (Rust core + system WebView), a native HTML/CSS/JS front end with zero frameworks — the installer is about 7MB and it opens instantly. The platform layer and the app layer are physically separated and speak only through contracts; adding an app is roughly a directory plus a manifest.
This project was built by one person pairing with AI: the AI wrote the code, I steered — through a hundred-plus rounds of independent review. Every security red line is pinned by a test; every "undoable" was verified to be real. It is a product, and also an experiment in how seriously a human + AI pair can do engineering.
No red-alert UI, no countdown pressure; every deletion is undoable, and bulk or clear-all actions require a preview and confirmation.
02The AI can only propose destructive operations — execution always waits for your hand, and that confirmation cannot be forged by the model at the code level.
03The AI cannot schedule tasks for itself, cannot edit project instructions, cannot change settings through chat: the paths to self-perpetuation deliberately do not exist.
Version 0.1, collecting its first real feedback — smooth or rough, I want to hear it.
— Feedback
GitHubFeedback.
Feedback is powered by GitHub Discussions. Share suggestions, issues, progress questions, or notes on the work.
Feedback did not load this time. Refresh to retry, or continue on GitHub Discussions.