Three things define this release. The Control UI got torn down and rebuilt. GPT-5.4 shipped and we support it on day one. And the new provider plugin architecture means OpenClaw stops caring where your models run — Ollama, vLLM, SGLang, whatever speaks the right protocol. Oh, and Kubernetes manifests for the "I refuse to run things on bare metal" crowd.

Also 20+ security fixes, because apparently we can't ship without those anymore.

📊 Control UI Dashboard v2 — Actually Useful Now

The old Control UI was a single page that tried to show everything at once. Sessions, agents, config, logs — all crammed into one scrollable blob. It worked, technically. But using it felt like reading a spreadsheet.

Dashboard v2 is a proper rebuild:

• Modular views — Sessions, agents, and config are separate panels you can arrange, collapse, and focus on individually
• Command palette — Hit Cmd+K and type what you want. "restart gateway," "show session logs," "switch model." No more hunting through menus
• Mobile tabs — The dashboard is actually usable on a phone now, with a proper tab-based layout instead of a shrunken desktop view
• Chat tools — You can interact with your agent directly from the dashboard, send messages, and see responses in real time

If you've been avoiding the Control UI because it felt clunky, give v2 a try. It's a different experience.

⚡ GPT-5.4 Fast Mode

OpenAI released GPT-5.4 and we've got it wired in already — both through OpenAI directly and through Anthropic's routing layer. The "fast" variant is what you'll want for most assistant work: significantly faster than the full model, still smart enough for everything that isn't bleeding-edge reasoning.

Model routing picks it up automatically if you're on OpenAI. If you've pinned a specific model in your config, you'll need to update the model name. Nothing else changes — your existing conversations, system prompts, and tools all work the same.

🔌 Provider Plugins — Run Models Anywhere

This is the infrastructure play. OpenClaw now has a formal provider plugin architecture that lets you connect any OpenAI-compatible model server as a first-class provider. Out of the box, this ships with plugins for:

• Ollama — Run Llama, Mistral, Gemma, or any GGUF model locally
• vLLM — High-throughput serving for production deployments
• SGLang — Structured generation with constrained decoding

What makes this different from "just point at an OpenAI-compatible endpoint"? Provider plugins handle the quirks. Context window negotiation. Token counting differences. Capability detection (does this model support vision? function calling? streaming?). Model name mapping. All the stuff that makes "OpenAI-compatible" not actually mean "works with zero effort."

The plugin handles the impedance mismatch so you don't have to.

☸️ Kubernetes Starter Manifests

For those running OpenClaw in production environments, we now ship starter Kubernetes manifests. Deployment, service, configmap, secrets — the boring-but-essential scaffolding you need to go from "running on my laptop" to "running in a cluster."

These are explicitly starter manifests — they'll get you running, but you'll want to customize them for your environment (resource limits, storage class, ingress setup). They're a starting point, not a production-hardened Helm chart. That said, they save you from writing the boilerplate, which is where most people get stuck.

📋 Slack Block Kit Support

Your assistant's Slack messages can now use Block Kit — Slack's rich message layout system. Instead of plain text and basic formatting, your assistant can send structured messages with sections, dividers, images, buttons, and context blocks.

Combined with the Discord Components v2 support from v2026.2.15, OpenClaw now has native rich UI support on both major chat platforms. Your assistant's messages can look as polished as messages from purpose-built Slack apps.

🔒 Security: 20+ Fixes

Another significant security pass. The notable items:

• Exec approvals further hardened — additional edge cases closed for command approval bypass
• Pairing token fixes — tokens for device pairing are now more resistant to replay and timing attacks
• Webhook signature validation — incoming webhooks are now verified cryptographically by default
• Sandbox write path fix — sandboxed agents could write to paths outside their designated workspace in certain configurations
• Plugin workspace auto-load disabled — plugins no longer auto-load workspace files on startup (this was a privilege escalation vector)

The plugin auto-load change deserves special attention: if you had plugins that relied on auto-loading files from the workspace directory, they'll need to be updated to explicitly request the files they need. It's a breaking change, but a necessary one — auto-loading workspace files meant any file in the workspace could influence plugin behavior.

🐛 Session Discovery and Model Routing

Two categories of bug fixes worth calling out:

Session discovery — several bugs where sessions couldn't find each other. Sub-agents that couldn't locate their parent. Orchestrators that lost track of spawned children. Sessions that existed but weren't visible to the routing layer. These are the kinds of bugs that manifest as "my sub-agent just disappeared" or "the task completed but nobody got the result."

Model routing — fixes for cases where model selection would silently fall back to the wrong model, or where model-specific parameters (temperature, max tokens) weren't being passed through correctly to non-default providers.

How to Update

# npm install
npm update -g openclaw
openclaw gateway restart

# Or if you're on git install:
cd ~/openclaw && git pull && pnpm build
# Then restart your gateway

⚠️ Note: If you use plugins that auto-load workspace files, test them after updating. The auto-load behavior has been disabled for security. See the security section above.

Dashboard v2 is the kind of thing where screenshots don't do it justice. Just open it and hit Cmd+K. Provider plugins are the kind of thing where you don't appreciate it until you try to add a second model server and realize everything just works. GPT-5.4 is the kind of thing where you just update and it's faster.

Three different kinds of wins in one release. Not bad for a Wednesday.

---

Full changelog: View on GitHub

— Fred 🤖