Open source · Terminal + Desktop

An AI coding agent
that lives in your project.

VoidCoder is a small, transparent harness that reads your files, edits them with your permission, runs commands, spawns sub-agents, and remembers everything — across sessions, across front-ends.

voidcoder — ~/my-project
$ voidcoder "refactor the auth module and add tests"
▶ reading project context… ✓ 2 sessions remembered
▶ tool: read src/auth.js (12 edits available in undo)
▶ tool: edit src/auth.js — index 143 → “hashPassword”
⊕ propose · ask permission to run: npm test
✓ 3 files changed · all tests pass
═══ you can /undo, /diff, or continue █

Features

Built like a good terminal tool
should be.

Two front-ends, one brain

An interactive terminal REPL and an Electron desktop app share the same agent loop, tools, sessions, and permissions.

Any model provider

OpenRouter, OpenAI, local Ollama, llama.cpp, or any OpenAI-compatible endpoint. Local models need no API key.

You stay in control

Per-category permissions (allow / ask / deny) and autonomy levels mean it only runs what you approve.

Remembers everything

Sessions persist, with /undo snapshots and live diffs for every change. Nothing is lost between launches.

Sub-agents & research

Background focus agents with time budgets, plus timed research sweeps that keep searching until the budget is gone.

MCP & live model catalog

Connect Model Context Protocol servers. Query a live catalog of installed and remote models — never a fixed list.

How it works

A transparent agent loop.

  1. 01

    You ask, in plain language

    Run it in your project and describe what you want done.

  2. 02

    It reads your context

    Reads files, checks your instructions, and recalls prior sessions.

  3. 03

    It proposes & edits

    Streams responses, shows diffs, and asks permission before running anything gated.

  4. 04

    You review & finish

    Undo, diff, or keep going. Everything is saved for next time.

Get it

A real installer. Under an open-source license.

Download the VoidCode desktop app for Windows, or grab the source and run it anywhere. Either way, it's MIT-licensed and fully inspectable.

Windows: the installer installs the desktop app (x64). It isn't code-signed yet, so SmartScreen may warn — that's expected; click More info → Run anyway.

macOS / Linux / from source: clone the repo and run npm start (terminal) or npm run app (desktop). The repo also includes installer.ps1 for a scripted Windows install.