A real AI editor — plan, edit, run, verify — on models you run yourself. Nothing leaves the box.
subtract adds instead of subtracts; divide needs a zero guard. Editing, then running pytest.pytest -qThe agent plans, edits, and verifies — entirely on your GPU.
Local models stall as coding agents for one reason — reliability: the raw model emits a tool call that's subtly malformed, the agent can't parse it, and the loop breaks. Everyone else bolts an editor onto a model server they don't control, or a model onto an editor they can't tune. We build all three. So we fix this below the model — well-formed tool calls, turn after turn — where a cloud IDE or a standalone server can't reach. A real agent on top: plan mode, diff-and-approve, checkpoints, MCP tools. A genuine pair-programmer, not a toy autocomplete.
Because we build the server, the agent and the editor, we force every turn into a well-formed tool call below the model — the failure that stalls every other local agent, fixed where a plugin can't reach.
Watch it reason about the change before it touches a file — and approve every diff.
Locked to your runtime at 127.0.0.1. Cloud providers are removed, not just hidden.
Terminals, extensions, a real file system. Installs like any Windows app.
Your model, your GPU, your code — never leaves the box.