OpenClaw 2026.4.27: GPU Passthrough, Codex Computer Use, and Yuanbao

OpenClaw dropped v2026.4.27 late Wednesday, and it's a dense one. The changelog spans GPU passthrough for Docker sandboxes, desktop control via Codex Computer Use, a brand-new Tencent Yuanbao channel, and a sweeping Plugin SDK refactoring that's been months in the making. Here's what's worth knowing.

Docker GPU Passthrough for Sandboxed Agents

The headline feature for local AI workloads: sandbox.docker.gpus passthrough is now opt-in. If your host Docker runtime supports --gpus, sandboxed agent containers can now access it directly. This closes #57976 and was contributed by @cyan-ember.

For anyone running local GPU-accelerated models inside sandboxed OpenClaw agents — think Ollama with a dedicated GPU, or custom CUDA workloads — this is a meaningful unlock. Previously, Docker sandbox containers were GPU-blind.

Codex Computer Use Is Now First-Class

The /codex computer-use command gets a proper setup flow in this release: status/install, marketplace discovery, optional auto-install, and fail-closed MCP server checks before Codex-mode turns start. Fixes #72094, contributed by @pash-openai.

New documentation also clarifies how Codex Computer Use, direct cua-driver MCP, and OpenClaw.app's PeekabooBridge fit together — historically a source of confusion for users setting up desktop control pipelines.

Tencent Yuanbao Channel Support

OpenClaw now officially registers the Tencent Yuanbao external channel plugin (openclaw-plugin-yuanbao) in the channel catalog. There's a new docs/channels/yuanbao.md quick-start guide for WebSocket bot DMs and group chats. The channel is contributed by @loongfay and represents a significant expansion for Chinese-language deployments.

QQ Bot also got a major overhaul in this release — full group chat support including history tracking, @-mention gating, activation modes, per-group config, a FIFO message queue with deliver debounce, and C2C stream_messages streaming via a new StreamingController lifecycle manager. Contributed by @cxyhhhhh.

Outbound Proxy Routing

Operators can now configure a managed outbound proxy (proxy.enabled + proxy.proxyUrl / OPENCLAW_PROXY_URL) with strict http:// forward-proxy validation, loopback-only Gateway bypass, and proper cleanup on exit. Contributed by @jesse-merhi and @joshavant via #70044.

This is useful for enterprise deployments where all outbound traffic must route through a corporate proxy.

Plugin SDK Overhaul

A significant chunk of this release is the Plugin SDK cleanup. Highlights:

• Manifest-backed provider catalogs: Qianfan, Xiaomi, NVIDIA, Cerebras, Mistral, Moonshot, DeepSeek, Tencent TokenHub, StepFun, BytePlus, Volcano Engine, Fireworks, and Together AI all move their model catalogs into plugin manifest modelCatalog rows. Core no longer carries bundled-provider routing tables.
• Explicit startup activation: Plugins now declare activation.onStartup in their manifests. The old implicit sidecar loading is deprecated with a compat warning and dated removal window.
• Focused SDK test subpaths: A dozen new focused test subpaths replace the old test/helpers bridges, so extension tests no longer need to import repo-internal paths.

Other Notable Changes

• iOS/Android presence events: A new authenticated node.presence.alive protocol event lets background iOS and Android wakes mark paired nodes recently alive without treating them as connected.
• Non-image attachments via chat.send: Gateway now stages non-image files as agent-readable media paths rather than silently dropping them. Fixes #48123.
• DeepInfra provider: A new bundled DeepInfra provider lands in this release.
• Matrix streaming: Tool-progress updates now stream into live Matrix preview edits by default.
• Docker TLS fix: The slim Docker runtime image now bundles the CA certificate bundle, fixing HTTPS failures on containerized gateways after the bookworm-slim base switch. Fixes #72787.

Full changelog: v2026.4.27 on GitHub.