OpenClaw v2026.3.22: Everything New in the Biggest Release Yet

OpenClaw just dropped v2026.3.22 and this one is massive. New plugin SDK, ClawHub-first installs, SSH sandboxes, three new search providers, GPT-5.4 defaults, a full Matrix rewrite, and over 50 changes across the board. Here is everything that matters.

Breaking Changes You Need to Know

This release has more breaking changes than usual. If you are running plugins, custom skills, or the old Chrome extension relay, read this section carefully before updating.

Legacy env names removed. If you still use CLAWDBOT_* or MOLTBOT_* environment variables, they are dead now. Switch to OPENCLAW_* equivalents.

Old state directory gone. The ~/.moltbot auto-detection fallback is removed. If your state still lives there, move it to ~/.openclaw or set OPENCLAW_STATE_DIR explicitly before updating.

Image generation standardized. The bundled nano-banana-pro skill wrapper is removed. Use agents.defaults.imageGenerationModel.primary instead. One config line replaces the old skill setup.

Message plugin discovery changed. Plugin authors need to migrate to ChannelMessageActionAdapter.describeMessageTool(...). The old listActions, getCapabilities, and getToolSchema methods are gone.

ClawHub Is Now the Default

This is the big one for everyday users. When you run openclaw plugins install, it now checks ClawHub first before falling back to npm.

That means community skills, third-party plugins, and marketplace packages are now first-class citizens. You get native openclaw skills search, install, and update flows. Tracked update metadata means your agent knows when skills have newer versions available.

On top of that, this release adds Claude marketplace registry resolution. You can install plugin@marketplace directly. Codex, Claude, and Cursor bundle discovery works out of the box, with bundle skills automatically mapped into OpenClaw skills.

Three New Search Providers

Your agent just got three new ways to search the web. All bundled as plugins, all configurable from config.

• Exa brings Exa-native date filters, search-mode selection, and optional content extraction. Configure under plugins.entries.exa.config.webSearch.*.
• Tavily adds dedicated tavily_search and tavily_extract tools. If you have a Tavily API key, your agent can use it natively.
• Firecrawl exposes firecrawl_search and firecrawl_scrape tools with Firecrawl base-URL fallback and guarded endpoint fetches.

All three work alongside the existing Brave Search integration. Mix and match based on your needs. The onboarding flow will help you configure them.

SSH Sandboxes and OpenShell

Sandbox infrastructure got a major upgrade. Instead of being Docker-only, sandboxes now support pluggable backends.

The new SSH sandbox backend lets you connect to remote machines with secret-backed keys, certificates, and known_hosts. Your agent can execute code on a remote server inside a proper sandbox without Docker.

OpenShell ships as a backend with mirror and remote workspace modes. sandbox list, recreate, and prune are now backend-aware.

This is big for VPS users who want isolated execution without running Docker alongside OpenClaw. Set up an SSH key, point it at a sandbox machine, and your agent has a safe place to run code.

GPT-5.4, Claude on Vertex, and More Models

The model ecosystem expanded significantly.

OpenAI defaults updated to GPT-5.4. The default setup model is now openai/gpt-5.4. Codex uses openai-codex/gpt-5.4. Forward-compat support for gpt-5.4-mini and gpt-5.4-nano is ready.

Claude on Google Vertex AI. The new anthropic-vertex provider routes Claude through Google Vertex AI with GCP auth discovery. If you are already on Google Cloud, this can simplify billing and auth.

Per-agent thinking defaults. Each agent can now have its own thinking/reasoning/fast defaults. If one agent needs deep reasoning and another needs speed, you configure them independently.

Other model updates: Chutes provider added, MiniMax upgraded to M2.7, xAI Grok catalog synced, Z.AI GLM 4.5/4.6 families added, Mistral pricing fixed, and Xiaomi MiMo V2 Pro and Omni added.

Browser and Chrome MCP Changes

The legacy Chrome extension relay is gone. If you were using driver: "extension" and browser.relayBindHost, they no longer exist.

Run openclaw doctor --fix to auto-migrate to the existing-session or user browser profile. Docker, headless, sandbox, and remote browser flows are unaffected.

New: browser.profiles..userDataDir support means you can point Chrome DevTools MCP at Brave, Edge, or any Chromium browser through their user data directories. One config line to attach to whatever browser you already use.

Full Matrix Plugin Rewrite

Matrix got a complete rewrite backed by the official matrix-js-sdk. If you are upgrading from the previous Matrix plugin, follow the migration guide.

Key improvements: durable event deduplication across gateway restarts (no more replayed messages), bot-to-bot communication with allowBots room policy, per-account allowPrivateNetwork for internal homeservers, and proper mention-gated routing so idle sessions expire normally.

Android: Dark Theme, SMS Search, Call Logs

Android got several quality-of-life upgrades.

• System-aware dark theme across onboarding and post-onboarding screens. The app follows your device theme setting.
• SMS search through the gateway. Your agent can search text messages on your Android phone.
• Call log search. Same thing for recent call history.
• Talk audio now routes through the gateway instead of device-local ElevenLabs streaming. Secrets stay on the gateway where they belong.

Telegram: Custom Endpoints and Auto Topics

Two big Telegram additions.

Custom Bot API endpoints. If you run a self-hosted or proxied Telegram Bot API, you can now set per-account apiRoot and it works across send, probe, setup, doctor repair, and media downloads. End-to-end support for non-standard Telegram deployments.

Auto-renaming DM topics. Forum topics now get LLM-generated labels on the first message. Your DM threads get meaningful names instead of generic labels. Override per-account or per-DM with autoTopicLabel.

Also: silent error replies (off by default, enable with channels.telegram.silentErrorReplies) so bot error messages do not ping users.

Security Fixes

This release hardens several security surfaces.

• Windows media path injection blocked. Remote file:// URLs and UNC paths can no longer trigger outbound SMB credential handshakes through media loading.
• Exec sandbox hardened. Build-tool JVM injection (MAVEN_OPTS, SBT_OPTS, etc.), glibc tunable exploitation, and .NET dependency hijacks are now blocked from the host exec environment.
• Voice webhook hardened. Missing signature headers are rejected before body reads. Pre-auth body budget dropped to 64 KB / 5s. Concurrent pre-auth requests capped per IP.
• iOS pairing tightened. Setup codes are now bound to the intended node profile and reject first-use bootstrap requests with broader scopes.
• Nostr DM security. Inbound DM policy is enforced before decrypt with pre-crypto rate and size guards.
• Gateway discovery. Unresolved Bonjour and DNS-SD endpoints fail closed. TXT-only hints can no longer steer routing.

Control UI Upgrades

The Control UI got several polish passes.

• Expand-to-canvas button on assistant chat bubbles. Click to expand any response into a full canvas view.
• Roundness slider in Appearance settings. Adjust corner radius from sharp to fully rounded across all UI elements.
• Session navigation from Sessions and Cron views. Jump directly into any session without hunting.
• Usage view improved. Better styling, localization, responsive layout, and no more empty placeholder cards.

How to Update

If you installed OpenClaw via npm:

npm update -g openclaw

Then restart your gateway:

openclaw gateway restart

After updating, run the doctor to catch any config issues:

openclaw doctor --fix

Full release notes: github.com/openclaw/openclaw/releases/tag/v2026.3.22

Want a step-by-step installation guide? Check out installopenclawnow.com for the fastest way to get OpenClaw running on your machine.