Introduction

Latch is an open-source coding agent written in Rust for Linux, Windows, and Apple Silicon macOS. It provides a terminal interface, a local Web interface, and a machine CLI. You choose a model provider and supply its credential.

Start here¶

Install Latch, open a terminal in your project, and run latch. Use /setup to configure a provider and model, then run latch doctor to check local prerequisites without contacting the provider.

latch
# Or open the local browser interface:
latch --web

See terminal and CLI usage and the Web interface guide.

Execution and evidence¶

Commands run inside a mandatory platform sandbox: Bubblewrap on Linux, Seatbelt on macOS, or the embedded native AppContainer runner on Windows. If the sandbox is unavailable, command execution fails. Modes, safety policy, and approval handling control which operations may proceed. Sandboxing does not establish that a change is correct; recorded validation provides evidence for the workspace state it checked. Later writes make earlier passes stale.

Sessions persist as durable events in SQLite. Working-history rotation bounds model requests while keeping the original history available. Resume restores session state and evidence. Root agents can delegate to independent child sessions; child reports do not certify the root task.

Read about working safely, continuity, and multi-agent coordination.

Integration status¶

Agent Skills and MCP tool servers are implemented in the current release. Skills are instructions loaded on demand, not permission grants. MCP supports tool discovery and execution over sandboxed stdio or Streamable HTTP. Remote HTTP services run outside Latch's local sandbox. Windows stdio servers reconnect for each call; stateful Windows servers need Streamable HTTP.

MCP resources, prompts, sampling, elicitation, subscriptions, OAuth flows, and asynchronous tasks are not exposed. Computer use and several replaceable runtime ports remain deferred. See Skills and MCP for supported behavior and runtime capabilities for the implemented and deferred boundary.

Supported platforms and limits¶

Precompiled binaries target x86_64 and ARM64 Linux (glibc 2.35+), x86_64 Windows, and Apple Silicon macOS. Linux needs Bubblewrap, ripgrep, and Git. Windows needs Git, ripgrep, and a focused user-owned NTFS workspace. The native Windows sandbox is embedded in the binary. Apple Silicon macOS uses the system Seatbelt sandbox, Git and ripgrep; macOS 15 is natively tested with a macOS 14 deployment baseline.

Windows qualification does not cover every host or workspace: an entire live home directory and hostile host races remain unqualified. Read the Windows findings and boundary status before relying on behavior beyond the tested scope.

Source and recorded results¶

Latch is MIT licensed. Source and issues are on GitHub. The 25-task comparison records one attempt per task with a specific model and host; it is an exploratory measurement, not a general performance guarantee.