OpenClaw Codex Agents Can Now Control Your Desktop

OpenClaw's Codex harness can now drive desktop applications. The self-hosted AI platform shipped Computer Use integration this week, giving Codex-mode agents the ability to see screens, move cursors, and click through native applications — turning an AI chat assistant into something closer to a hands-on operator.

What Computer Use Actually Does

Computer Use bridges OpenClaw's agent system with the Codex app-server's native plugin infrastructure. When enabled, the agent gains access to screen capture, mouse control, and keyboard input through MCP tool calls — the same protocol OpenClaw uses for its other tool integrations.

The boundary is deliberately narrow: Codex owns the native MCP plugins and tool execution, while OpenClaw handles configuration and validation on the app-server side. First-time setup is automatic. OpenClaw discovers the available marketplace, syncs the Computer Use plugin, installs it, and verifies readiness before the agent's first turn executes.

Two Ways to Set It Up

For most users, the default path is hands-off. OpenClaw calls the app-server's plugin list, waits for the marketplace to sync, selects Computer Use, and installs it automatically. Turn-start auto-install uses previously discovered marketplaces, so subsequent sessions skip the sync wait entirely.

Power users get override options. Configuration fields under plugins.entries.codex.config.computerUse let you specify a custom marketplace source, path, or name. There are also two new CLI commands: /codex computer-use status to check the current state and /codex computer-use install for explicit installation when you want full control over the process.

Already Validated in the Wild

The team validated the feature with a live test: a Codex agent launched macOS Calculator, performed an arithmetic operation, and returned the correct result — all through Computer Use tool calls. The agent harness confirmed zero fallback usage, meaning the operation ran entirely through the native Codex pipeline without falling back to alternative execution paths.

Plugins Get a Final-Answer Review Gate

Alongside Computer Use, OpenClaw also shipped a new lifecycle hook for plugins: before_agent_finalize. This hook fires at the natural final-answer boundary — the moment an agent has composed its response but before the harness commits it.

Plugins can return one of three decisions: continue to accept the response, revise to send the agent back for another pass, or finalize to force completion immediately. For Codex specifically, the hook maps Codex's native Stop events into OpenClaw's finalization workflow — a decision: block becomes revise, and continue: false becomes finalize.

The practical use case is quality control. A plugin could intercept an agent's final answer to check for compliance, run a secondary validation, or enforce output formatting — without the user seeing the intermediate response. Non-bundled plugins need the allowConversationAccess permission to use this hook, which keeps the security surface tight.

Why This Matters

Computer Use turns OpenClaw from a chat-and-tools platform into something that can interact with any desktop application. Combined with the new finalize hook — which lets plugins review and correct agent output before it reaches users — the Codex harness is becoming a genuine automation runtime, not just a conversation interface.

For teams already running OpenClaw as their self-hosted AI backbone, these updates mean agents can now handle workflows that previously required human screen interaction: filling out forms, navigating legacy applications, or operating tools that don't expose APIs.