OpenClaw 2026.4.27 Release: Smarter Desktop Control, a New AI Provider, and Across-the-Board Stability Improvements

If you’re using or evaluating the OpenClaw open-source project, the April 27, 2026 release deserves your attention. This update introduces several substantial new capabilities while also addressing dozens of reliability issues that directly affect day-to-day usage. Whether you’re just getting started or already running OpenClaw in production, this breakdown will help you understand what’s changed and how it might affect your workflows.

Let’s start with the three biggest highlights, then dig into the details.

The three major themes of this release

First, Codex Computer Use officially lands, giving AI agents the ability to see and interact with desktop environments under strict safety guardrails. Second, DeepInfra joins the line-up of bundled model providers, adding a versatile option for image, audio, video, and embedding tasks. Third, a wide-ranging stability push touches nearly every popular messaging channel—Telegram, Slack, Matrix, QQ, and more—alongside core components like session management, memory compaction, and plugin loading.

The release also expands channel coverage for Tencent Yuanbao and QQBot, and moves the plugin system significantly toward manifest-first metadata, which speeds up Gateway startup and makes provider catalogs easier to audit.

Let’s walk through each theme.

Codex Computer Use: Letting AI See and Operate Your Desktop

The most eye-catching addition in this release is Codex Computer Use. In plain terms, this feature allows your AI assistant to view the screen and control the mouse and keyboard—all while operating within a tightly permissioned framework.

What problem does it solve?

Normally, when you ask an AI to help with a task that involves desktop applications, it can only guess the interface state from text descriptions. With Computer Use, an agent in Codex mode can literally observe the screen content and perform appropriate actions. This is a game-changer for test automation, repetitive office tasks, and workflows that span multiple desktop apps.

What’s included?

The release ships a complete setup experience for Codex-mode agents:

• Status check command: /codex computer-use status shows whether your Computer Use environment is ready.
• One-command installation: /codex computer-use install handles the environment setup.
• Marketplace discovery: You can find and install related extensions from the marketplace.
• Optional auto-install: The system can automatically set up the environment on first use if you prefer.
• Fail-closed safety checks: Before a Codex-mode turn starts, the system verifies that the MCP (Model Context Protocol) server is available. If it isn’t, desktop control is blocked—a “fail-closed” approach that prevents unsafe operation.

How it fits with other components

If you also use OpenClaw.app’s PeekabooBridge, the updated documentation now explains how Codex Computer Use, direct cua-driver mcp, and PeekabooBridge work together. This makes it easier to decide which combination suits your desktop-control setup.

A practical note: if you have Codex installed on your machine but haven’t configured OpenAI authentication, the system won’t silently expose Codex delegation capabilities. You need to explicitly enable skills.entries.coding-agent.enabled in your configuration. This prevents accidental exposure of features that can’t actually function.

DeepInfra: A Versatile New Provider Joins the Bundle

For anyone working with multimodal tasks—images, audio, video—more provider options are always welcome. This release introduces DeepInfra as a bundled provider with broad capabilities.

What can DeepInfra do?

The DeepInfra integration covers a comprehensive set of modalities:

• Dynamic model discovery: Automatically fetches the list of available OpenAI-compatible models.
• Image generation and editing: Supports text-to-image and image editing.
• Image and audio understanding: Can analyze uploaded images and audio clips.
• Text-to-speech (TTS): Provides speech synthesis.
• Text-to-video: Generates video from text prompts.
• Embeddings: Produces vector embeddings for memory search and semantic retrieval.
• Static catalog metadata: Offers a stable, auditable model catalog.

How to get started

Configuration is simple: set the DEEPINFRA_API_KEY environment variable. DeepInfra uses a provider-owned base URL policy, so you don’t need to specify an API endpoint—the system uses the official base URL.

Tencent Yuanbao and QQBot: More Coverage for Chinese Messaging Platforms

For users who operate within China’s messaging ecosystem, this release adds support for Tencent Yuanbao and significantly upgrades QQBot capabilities.

Tencent Yuanbao

The Tencent Yuanbao external channel plugin is now registered in the official channel catalog. A quick-start guide covers WebSocket bot direct messages and group chats. This means you can connect OpenClaw agents through Yuanbao’s interface.

QQBot

QQBot received a major functionality boost:

• Full group chat support: Includes group history tracking, @-mention gating, multiple activation modes, and per-group configuration.
• Message queue management: A FIFO queue with delivery debouncing prevents flooding.
• C2C streaming: One-on-one chats now support streamed messages with a lifecycle manager.
• Chunked upload for large files: A unified media-sending module handles large file transfers in chunks.
• Engine refactor: The underlying engine was restructured into pipeline stages, with focused outbound submodules and built-in slash-command modules.

These changes make QQBot more stable and controllable in group scenarios.

Plugin Manifest-First Metadata: Faster Starts, Easier Audits

If you run many plugins, you may have noticed startup times creeping up. This release takes a significant step toward manifest-first metadata management.

What is manifest-first?

A plugin manifest is a file that describes a plugin’s basic information—activation conditions, model catalogs, aliases, and suppression rules. Previously, much of this information was computed or queried at runtime. Now, more of it lives in the manifest file itself.

What are the benefits?

• Faster startup: The Gateway only needs to read manifest files, not load full plugin code, to decide which plugins to activate.
• Easier auditing: Provider model rows, aliases, and suppression rules are declared in manifests, making them straightforward to inspect.
• Explicit startup declarations: Plugins must now declare activation.onStartup to be imported at Gateway boot. This prevents unnecessary loading.

Heads-up for plugin authors

If you maintain your own plugins, the legacy implicit startup loading still works but has been deprecated. You’ll see compatibility warnings. It’s recommended to migrate to the explicit activation.onStartup metadata field.

Stability Fixes: Covering Channels and Core Components

The fixes section of this release is extensive. Here they are, grouped by impact area.

Telegram fixes

Telegram is one of the most widely used channels, and several fixes target common pain points:

Slack fixes

• Socket Mode timeouts: A default 15-second pong timeout is now active, with new configuration overrides available for fine-tuning.
• Media download stalls: Private file and forwarded attachment downloads now have idle and total timeouts to prevent infinite waits.
• Auto-reply cleanup: Session reset phrases like new session no longer leak into the next model turn.

Matrix improvements

• Tool progress updates are streamed into live preview edits by default.
• Approval messages now carry versioned structured metadata for richer client rendering.

Session and memory system

Session management received several hardening updates:

• Session history defaults: chat.history and sessions.list defaults now align with backend runtime state.
• Device pairing recovery: Malformed JSON state files no longer cause paired entries to disappear.
• Gateway authentication: Stale device tokens are cleared to avoid rate-limit loops after token rotation.
• Approval idempotency: Duplicate approval resolutions within the grace window are now treated as idempotent.
• Memory compaction: Pre-compaction memory-flush prompts are kept runtime-only and no longer appear in session transcripts.

Agent execution and model fallback

• Sub-agent model override: sessions_spawn.model now correctly sets the child session’s initial model.
• Fallback trajectory events: A new model.fallback_step event records each fallback attempt with source model, target model, and failure reason—useful for support exports.
• Empty output handling: Silent exec and node result blocks now emit (no output) to prevent Anthropic-compatible providers from rejecting empty tool result text.

Docker and Sandbox Environments

The Docker sandbox gets two practical capabilities:

1. GPU passthrough: An optional sandbox.docker.gpus setting lets sandboxed agents use host GPUs when the Docker runtime supports --gpus. This is ideal for local inference workloads inside sandboxes.

2. Default image fix: When the default sandbox image is missing, the system no longer incorrectly tags a plain Debian image as openclaw-sandbox:bookworm-slim. Instead, it guides you to build the correct image that includes the Python tooling required by sandbox helpers.

Control UI and WebChat Experience

The web-based interface received several thoughtful improvements:

• Large attachment uploads: Fixed a RangeError: Maximum call stack size exceeded crash when uploading large PDFs or images.
• New Session button: The toolbar button now confirms reset before executing.
• Non-image attachments: Generic file attachments can now be uploaded through the chat interface.
• Streaming rendering: Duplicate chat messages caused by rapid sends or replays are now prevented.
• TTS audio persistence: Final-mode TTS audio is stored as a supplemental audio-only update, avoiding a duplicate text message.
• Tool output redaction: Sensitive information in tool call arguments and results is redacted before being streamed to the Control UI.

Platform Updates: macOS, iOS, and Windows

Maintenance updates for native apps and daemons landed on all major platforms:

• macOS: Core dependencies like Peekaboo and ElevenLabsKit were updated. The Gateway service is now properly stopped before package replacement. LaunchAgent service secrets are stored in owner-only env files, not world-readable plists.
• iOS: The project passed SwiftFormat and SwiftLint checks. Simulator builds and logic tests are warning-free.
• Windows: Restart handoffs without a listener are now routed through the Windows supervisor, and failed task scheduling can be retried.

Frequently Asked Questions

Do I need to do anything to upgrade?

If you installed OpenClaw via npm, run openclaw update. The update process skips plugin sync for any plugins you’ve disabled in your configuration, preserving their install records.

If you’re in an offline or restricted network, note the new models.pricing.enabled option. Setting it to false skips the startup fetch of pricing catalogs from OpenRouter and LiteLLM while keeping explicit model costs functional.

How is DeepInfra different from existing providers?

DeepInfra stands out for its all-in-one multimodal coverage. Beyond standard chat, it offers image generation/editing, image/audio understanding, text-to-speech, text-to-video, and embeddings. If you want a single provider for diverse media tasks, DeepInfra is a solid option.

Is Codex Computer Use safe?

Yes. The fail-closed design is the primary safety mechanism. Before any Codex-mode turn begins, the system checks that the MCP server is reachable. If it isn’t, desktop control is refused. Additionally, dangerous media commands are opt-in; safe commands like camera listing, device info, and screen snapshot are available on Windows companion nodes by default.

How do QQBot group chats work?

Group chat support includes full history tracking and @-mention triggering. You can configure per-group policies for activation modes and message queuing. By default, group replies are private—the agent won’t post to the group unless it explicitly uses the message tool. You can restore legacy auto-posting with messages.groupChat.visibleReplies: "automatic".

Should I update my custom plugins?

If you’re using bundled plugins, everything is handled for you. If you’re a plugin author, watch for deprecation warnings about implicit startup loading. Migrating to the manifest-first model with explicit activation.onStartup is recommended for future compatibility.

Summing Up

OpenClaw 2026.4.27 balances new functionality with deep reliability work. Codex Computer Use opens up desktop automation scenarios that were previously out of reach. DeepInfra gives you a flexible, multimodal provider option out of the box. Tencent Yuanbao and QQBot support make the platform more accessible in China’s messaging landscape.

Meanwhile, the sheer volume of fixes—covering Telegram stalls, Slack media timeouts, session correctness, model fallback logging, plugin startup performance, and much more—means upgrading should deliver a noticeably smoother experience.

For the complete, itemized changelog with linked issues and pull requests, visit the OpenClaw releases page. Every change listed there has been community-tested and is ready for you to put to work.