DarkHorseCode is a desktop app that turns a plain-English idea into a working project. Coding agents read, write and run code while you watch every step and a live preview update.
Open Settings, choose a provider, and paste a key or sign in. No project needs to be open.
On New App, type your idea and choose a build type: Full-stack, Mobile, Landing page or Website.
You land in the Builder. Follow each tool call and the live preview, and send follow-ups.
One file, no setup scripts. Pick your system and run it.
On Linux, make it runnable once: chmod +x DarkHorseCode.AppImage, then double-click it. · All versions
Five tabs cover the whole loop: start a project, build it, reason about its structure, reopen it later, and ship it.
Describe what you want in plain English. This always starts a brand-new project and never reuses the one that is open.
The main workspace. Chat with the coding agent, see each read, edit and command as a card, and watch the live preview change.
A system-tree view of the project's logic and structure. It has its own chat, separate from the Builder, so you can talk design without touching the live build.
Every app you have built, newest first.
Ship the open project through a deployment connector you have added, such as Vercel.
Every past conversation with your agents for this project, newest first, separate from the main Builder thread.
Each agent has a persona, a model and a set of allowed tools. Create custom ones in Agent Studio and switch them on or off without deleting them.
Edits files and runs commands. This is the agent that actually changes your project.
Maps the project as a tree and plans structural changes without editing anything.
Proposes an approach for a change before any code is written.
Talk it through with Plan first, then switch to Build to make the change. Long sessions burn tokens fast, so start a new session for each new feature.
Every agent can use these while it works. Custom agents get only the tools you tick in Agent Studio, plus any MCP connectors you add.
Model Context Protocol servers give agents extra tools: Supabase for Postgres, auth and storage; Vercel for deploys and env vars; Playwright to drive a real browser.
Tokens and cost across every agent run: total tokens, input and output split, cost and agent steps, with a per-day breakdown. It updates live while an agent runs.
Choose your model provider, sign in, and set agent behaviour. It works without a project open, since it is also where you log in.
Create an agent with a name, persona and system prompt, then choose its model and allowed tools.
One provider switcher; the model list changes with it.
Point it at any OpenAI-compatible endpoint, or use OpenCode Zen, opencode.ai's gateway with several free coding models. Zen needs a free API key with no billing and no card.
Runs Claude models: Opus 5, Sonnet 5, Fable 5.1 and Haiku 4.5. A Claude Pro or Max subscription cannot be used, because Anthropic rejects that login token from third-party apps (401 "API key is invalid"). Get a key from console.anthropic.com.
No API key. Sign in with your ChatGPT account and run on your existing plan. ChatGPT Go covers three models: gpt-5.5, gpt-5.6-luna and gpt-5.6-terra. The rest of the GPT-5/6 line needs Plus or Pro.
Each model is tagged (Go & Plus/Pro) or (Plus/Pro only). A Plus-only model fails on a Go plan.
No key, no cloud, no cost. Models run on your own GPU or CPU. Settings detects your hardware and suggests models that fit your VRAM and RAM, from qwen2.5-coder:7b up to gpt-oss:120b.
ollama pull qwen2.5-coder:7blocalhost:11434Press ? in the app to see this list anytime.
Dark comes from China's dark factories: production lines that run with the lights off and no one on the floor. That is the speed the app chases, agents that build end to end with nobody standing over them. A dark horse is the unexpected one that comes from nowhere and wins. Code is what it does.