Overview

OpenLatch is the AI Operations Platform for the enterprise — a runtime control plane deployed inside every AI agent. It reads every action the agent takes — file writes, shell commands, network requests — for cost and risk, and is built to enforce before an action executes.

The Problem

AI coding agents are powerful, but they operate with broad permissions in your development environment. A compromised prompt, a malicious dependency, or an adversarial injection can turn your helpful assistant into an attack vector.

The Solution

OpenLatch runs inside the agent, not as a proxy in front of it — the only vantage point from which cost and risk are the same behavioral signal. Every action the agent attempts is captured as a behavioral event and forwarded to the control plane.

• Capture — A thin client intercepts agent tool calls before execution
• Forward — Each action is wrapped in an envelope and sent to the control plane
• Enforce — The control plane is built to return a verdict before the action executes

Current state: the client captures and forwards events today and fails open — it does not block. The enforcement core that produces verdicts is being rebuilt; until it lands, OpenLatch gives you the behavioral event stream, agent inventory, and audit trail.

Supported Agents

OpenLatch works with any AI coding agent that supports tool call hooks:

• Claude Code
• Cursor
• GitHub Copilot
• Windsurf
• Custom agents via the OpenLatch SDK

Design Principles

• Zero configuration — npx @openlatch/client init is all you need
• Fail-open — If the cloud is unreachable, agents continue working
• Privacy-first — Your code stays local; only action metadata is analyzed
• Open source — The client is Apache-2.0-licensed and fully auditable