Open-source · Linux, Windows & macOS
Latch
A coding agent for your terminal and browser.
Written in Rust, with mandatory command sandboxing, durable sessions, and validation tied to the workspace it checked. Use your own model provider.
Installation
Full install guide →Precompiled binaries for Linux x86_64/ARM64, Windows x86_64, and Apple Silicon macOS. The installers verify SHA256 checksums.
Linux / macOS Bash · automatic architecture selection
curl -fsSL https://jinshuo-li.github.io/latch/install.sh | bashLinux requires Bubblewrap, ripgrep, and Git; Apple Silicon macOS uses Seatbelt, ripgrep, and Git. Both install to ~/.local/bin; follow the installer’s PATH instructions.
Windows PowerShell
& { $ErrorActionPreference='Stop'; [Net.ServicePointManager]::SecurityProtocol=[Net.ServicePointManager]::SecurityProtocol -bor [Net.SecurityProtocolType]::Tls12; irm -DisableKeepAlive https://jinshuo-li.github.io/latch/install.ps1 | iex; $env:Path="$env:LOCALAPPDATA\Programs\Latch\bin;$env:Path"; latch --version }Requires Git, ripgrep, and a user-owned NTFS workspace. The native sandbox is included. Add the install directory to your user PATH.
Two interfaces, the same kernel
Web interface guide →The TUI and local Web interface share sessions, tools, approvals, and validation.


How Latch works
- Sandboxed commands
- Bubblewrap on Linux; native AppContainer on Windows. Command execution stops if the sandbox is unavailable. Boundaries and limits →
- Durable sessions
- Session events persist in SQLite. Working-history rotation keeps the original record available; resume restores task state and evidence. Sessions and continuity →
- Recorded validation
- A pass applies to the workspace state it checked. Later writes make it stale, and active managed processes block certification. Validation semantics →
- Multi-agent support
- A root agent can delegate to independent child sessions. Child reports and group coordination do not certify the root task. Agent coordination →
Further reading
Agent Skills and MCP are available with explicit permission and lifecycle limits. Runtime capabilities distinguish implemented and deferred work.
The 25-task benchmark report records results, failures, and methodology from one exploratory comparison. It does not establish general performance guarantees.