Houston Beta

Houston: open-source AI coding agent Powerful enough to ship, safe enough to trust

Bring your own model: Claude, GPT, Gemini, any OpenAI-compatible API, or a local one. Houston reads, edits, searches, and runs code in your project, behind an approval flow you control.

Download Houston Install the CLI View on GitHub
  • Runs locally
  • No telemetry
  • Approval-gated actions
  • Sandboxed shell
  • macOS, Windows, Linux
  • Open source
Houston's model picker open, listing a local Ollama model and the Anthropic Claude, OpenAI GPT, and Google Gemini models, each row showing tool-calling, vision, and reasoning badges and its context-window size.
The model picker: every provider in one list, with each model's context window and capabilities.

Built to be trusted with real code

An agent that edits files and runs commands needs hard limits, not just good intentions.

  • A real OS sandbox

    Shell commands run under Seatbelt on macOS and bubblewrap on Linux (when installed). Writes stay inside the project, and network stays off until you allow it, then only to domains on an allowlist. Where no sandbox exists, Houston says so and never auto-approves.

  • You approve what matters

    Plan mode, ask every time, auto-approve edits, or full auto. Edits show an inline diff before you approve, and your choices can become lasting rules. Reaching the network prompts the first time, even in full auto.

  • Your checkout stays clean

    Each new chat in a git repo gets its own worktree and branch, so the agent never edits the files you are working in. Every turn's file changes revert in one click.

  • Guardrails for teams

    A repo can ship rules that only tighten, like always asking before touching infra/. Project config that would grant more is ignored until you trust the folder, and admins can lock a machine-wide policy.

  • Web pages are data, not instructions

    Fetched pages are fenced off as untrusted content. A page that looks like a prompt injection is read by a separate model call with no tools and no history, and only its summary reaches the agent.

  • Releases you can verify

    macOS builds are signed and notarized. Every release ships cosign-signed checksums, an SBOM, and SLSA build provenance, so you can check exactly what you installed and how it was built.

What Houston can do

More than chat. It works in your repository with real tools, and you stay in the loop on every step that matters.

  • Any model, your keys

    Claude (direct, Bedrock, Vertex AI, or Foundry), GPT, Gemini, any OpenAI-compatible endpoint, or a local model through Ollama or LM Studio. Switch per conversation.

  • Local and private

    Runs on your machine with no analytics or telemetry. Your code goes only to the model providers you configure.

  • Real tools

    Read, write, and patch files, run shell commands, inspect git, fetch and search the web, and work with pull requests. Reads run in parallel; writes stay in order.

  • Structural code search

    Bundled ripgrep and ast-grep. Find calls, declarations, and JSX by syntax-tree shape instead of fragile regex. With a language server installed, it also jumps to definitions, finds references, renames across files, and catches errors its own edits introduce.

  • Sees its own localhost

    After starting your dev server it can screenshot the page and read the console, then fix what it finds. A live preview docks beside the chat.

  • Subagents and review

    Hand focused tasks to subagents with fresh context, and run a multi-agent review of your diff that checks correctness, security, and quality, then verifies its own findings.

  • MCP, agents, and skills

    Connect MCP servers over stdio, HTTP, or SSE, with OAuth sign-in for hosted ones. Define custom agents and skills in Markdown. Your project's own conventions are picked up too.

  • GitHub, first-class

    With the gh CLI signed in, it opens and reviews pull requests, reads and files issues, and checks CI status without leaving the chat.

  • Desktop, terminal, or headless

    Use the desktop app, an interactive terminal session, or one-shot headless runs for scripts and CI. A single-file CLI runs anywhere Node 22+ does, ACP-capable editors can run it as their agent, and a GitHub Action reviews pull requests and fixes failed CI on your own runners.

Download Houston

Windows x64

Windows 10 or newer

Download .exe

Not code-signed yet. If SmartScreen appears, choose More info, then Run anyway.

Terminal CLI

The same agent in your terminal, with no desktop app. Runs on macOS, Linux, and Windows, on x64 or Arm, with Node 22 or newer.

Homebrew macOS, Linux
brew install houston-code/tap/houston
npm any OS
npm install -g @houston-code/cli
bun any OS
bun add -g @houston-code/cli
Install script macOS, Linux
curl -fsSL https://houstoncode.ai/install.sh | sh

Then start it with houston. Prefer a single file? Download houston-cli.cjs and run it with Node.

  • Verify your download

    macOS builds are signed and notarized, Linux packages are GPG-signed, and every release has a signed SHA256SUMS. How to verify.

  • Release notes

    What changed in each version, plus every asset and past release, is on GitHub Releases.

Getting started

  1. Install

    Download for your platform and open Houston. There is no account and no sign-up.

  2. Add a model

    In Settings, add an API key for at least one provider, or point Houston at a local model.

  3. Open a project

    Pick a folder and an approval mode, then give it work. You review the actions that matter.

Frequently asked questions

Something else? Open an issue or email dev@houstoncode.ai.

What is Houston?

Houston is a cross-platform desktop coding agent. You point it at an AI model and a project folder, and it can read, edit, search, and run code, with every consequential action gated by an approval flow, and shell commands sandboxed where the operating system supports it.

Is Houston free?

Yes. Houston is free to download and run. You bring your own model API keys, or run a local model, and pay only whatever your chosen provider charges for usage. Houston itself adds no fee.

Which AI models does Houston work with?

Claude (directly, or through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry), GPT, Gemini, any OpenAI-compatible endpoint, and local models via Ollama or LM Studio. You can switch models per conversation, and the picker shows each model's context window and capabilities.

Does Houston send my code or data anywhere?

Houston runs on your device and collects no analytics or telemetry from the app. Your data leaves your machine only when you use a feature that reaches a provider you configured (sending a prompt, searching the web, or checking for updates), and only to that destination. The one usage signal we can see is the aggregate download and update-check count our download host keeps for released files, which holds no personal data and involves no in-app tracking.

Which operating systems does Houston support?

macOS 13 Ventura or newer (Apple Silicon and Intel), Windows 10 or newer (x64), and Linux x64 (glibc 2.35+). A standalone terminal CLI also runs on macOS, Windows and Linux, x64 or Arm, with Node 22 or newer.

Is it safe to let an agent run code on my machine?

You choose an approval mode (from reviewing every action, to auto-approving edits, to full autonomy), and where the OS supports it, shell commands run in a best-effort sandbox. You stay in control of what executes.