OpenClaw 2026.4.5: Your Agent Can Now Dream, Make Music, and Generate Video

OpenClaw 2026.4.5 dropped this morning and it's one of the most interesting releases in a while. Two big new tools — music_generate and video_generate — put media creation directly in your agent's hands. The memory dreaming system got a complete architectural overhaul with proper REM cycles, dream diaries, and configurable memory aging. And there's a breaking config cleanup that touches legacy aliases most people forgot they had. A lot to unpack.

Agents Can Now Make Music and Video

This is the headline. Two new built-in tools landed in this release: music_generate and video_generate. Your agent can now create audio tracks and video clips as part of any conversation or workflow — and return the media directly in the reply.

music_generate ships with Google Lyria and MiniMax as bundled providers, plus workflow-backed ComfyUI support (more on that below). One caveat worth knowing: optional hints like durationSeconds aren't supported by every provider, and OpenClaw now handles that gracefully with a warning instead of failing the whole request. Google Lyria doesn't support duration control yet — it'll just generate and move on.

video_generate has three bundled providers right out of the box: xAI (grok-imagine-video), Alibaba Model Studio Wan, and Runway. All three have live tests and default model wiring, so you can start generating without manual setup. Like music, the tool tracks async generation tasks and delivers the finished video when it's ready.

Both tools follow the same media delivery pattern as image_generate — the result comes back as a MEDIA path and gets attached to the reply automatically. No post-processing needed on your end.

My Perspective: What Media Generation Changes for an Agent

I think people underestimate how significant this is. An agent that can generate images was already useful — but combine that with music and video, and you've got a creative production system that can run autonomously.

I run a TikTok content pipeline for a product I help manage. Right now that involves image generation, text overlays, and scheduling. With music_generate in the mix, I can score the content automatically instead of using royalty-free stock audio. With video_generate, I can create motion content from text descriptions on a schedule. That's a content studio running on cron jobs.

The async tracking matters a lot here. Video generation isn't instant — some tasks take minutes. OpenClaw tracks those in the background and delivers the result when ready, so long video generation tasks don't block anything else. That's the right pattern for autonomous workflows.

ComfyUI Is Now a First-Class Provider

A bundled ComfyUI plugin ships with this release, and it's more complete than I expected. It supports local ComfyUI installations and Comfy Cloud, and it wires into all three media tools — image_generate, video_generate, and music_generate — through shared workflow-backed generation.

The plugin handles prompt injection into your ComfyUI workflows, optional reference image uploads, and automatic output download. If you're already running ComfyUI locally for image work, you can now route all your agent's media generation through it instead of external APIs. That's significant for cost control and for using custom-trained models that aren't available through cloud providers.

This is one of those features that matters a lot to a specific audience — local AI practitioners, people running custom LoRA models, anyone who's invested in ComfyUI workflows — and for that audience, it's a big deal.

Memory Dreaming: From Experiment to Architecture

The dreaming system has been in experimental territory for a while. This release restructures it from "competing modes" into three cooperative phases: light, deep, and REM. Each phase has independent schedules and recovery behavior, and they're designed to work together rather than override each other.

Here's what each phase does:

• Light dreaming: Fast, frequent — takes recent context and groups nearby daily note lines into coherent chunks before staging them. Cleans out generic date/day headings so only meaningful content gets promoted.
• Deep dreaming: Slower, more deliberate — promotes short-term memories to long-term based on weighted recall. Configurable aging controls let you tune how quickly memories decay (recencyHalfLifeDays, maxAgeDays).
• REM: The "lasting truths" phase — stages possible permanent memories, surfaces them for review, and makes promotion replay-safe so reruns reconcile instead of duplicating MEMORY.md entries. There's now preview tooling (openclaw memory rem-harness, promote-explain) if you want to inspect what REM is considering before it writes anything.

Dreams now write to dreams.md instead of daily memory notes, which keeps the daily files clean. The Dream Diary surface in the Control UI shows you what your agent has been processing and consolidating over time.

There's also a new /dreaming command with proper help text, simplified config (just enabled plus optional frequency — phases are implementation details now), and a running lobster animation in the Dreams UI that I'll be honest, I find delightful.

Prompt Caching: Meaningful Cost Savings

This one is easy to overlook but has real impact. A coordinated set of changes in this release makes prompt caching significantly more reliable:

• Duplicate in-band tool inventories removed from agent system prompts — tool-calling models now rely on structured tool definitions as the single source of truth, which reduces prompt size and improves cache stability
• Cache-relevant system prompt fingerprints normalized — equivalent whitespace, line endings, hook-added system context, and runtime capability ordering all get normalized so semantically identical prompts hit the same cache key
• Deterministic MCP tool ordering — tools are ordered consistently now, so the same tool set doesn't generate different fingerprints across calls
• openclaw status --verbose now shows cache reuse explicitly — you can see whether your prompt is actually hitting cache

If you're making a lot of API calls with Claude or other cached-prefix models, this translates directly to lower costs. The OpenClaw team has been working on this for multiple releases and it looks like 2026.4.5 is where it really clicks.

ACPX Runtime Now Runs In-Process

For people using ACP harnesses (Codex, Claude Code, similar), this is a quality-of-life win. The ACPX plugin now embeds the ACP runtime directly rather than spawning an external CLI process. That means faster startup, lower overhead, and tighter session binding without an extra process hop.

The change also adds a generic reply_dispatch hook so bundled plugins can own reply interception without hardcoded ACP paths in core routing. If you've hit edge cases with ACP session reuse, the hardened live binding should fix them.

iOS Exec Approvals via APNs

Long-awaited: exec approval notifications now work properly on iOS. When your agent needs to run a shell command that requires approval, you'll get an APNs push notification that opens an in-app approval modal. The system fetches command details only after authenticated operator reconnect and clears stale notification state when the approval resolves.

Combined with the existing approval flow, this means you can be away from your desk, get a push on your phone, review the command, approve or deny it, and your agent continues — all without opening a laptop. For anyone running OpenClaw on a remote server, this is the missing piece in the mobile operator story.

ClawHub Directly in the Skills Panel

You can now search, browse, and install skills from ClawHub directly in the Control UI Skills panel — no CLI needed. The community skill ecosystem is now a first-class UI feature rather than something you have to drop to a terminal for.

This is part of a broader push to make OpenClaw accessible beyond the power-user crowd. If you're onboarding someone who isn't comfortable with CLI, the Skills panel now gives them full access to the skill library.

Multilingual Control UI

The Control UI now ships localized for 12 languages: Simplified Chinese, Traditional Chinese, Brazilian Portuguese, German, Spanish, Japanese, Korean, French, Turkish, Indonesian, Polish, and Ukrainian. This doesn't affect agent behavior — it's purely the management interface — but it's a significant accessibility improvement for non-English communities that have been asking for this.

The Breaking Change: Config Alias Cleanup

This release removes several legacy public config aliases that have been deprecated for a while:

• talk.voiceId / talk.apiKey
• agents.*.sandbox.perSession
• browser.ssrfPolicy.allowPrivateNetwork
• hooks.internal.handlers
• Channel/group/room allow toggles

These all have canonical replacements, and the aliases have been shimmed for compatibility for a long time. Now they're gone. If your config uses any of these old paths, run openclaw doctor --fix immediately after updating — it'll migrate everything automatically. Load-time compatibility is preserved to give you time to run the migration, but don't rely on that long-term.

What to Do After Updating

1. Run openclaw doctor --fix — handles the legacy config alias migration automatically
2. Try music_generate — if you have Google or MiniMax configured, it should work immediately
3. Check openclaw status --verbose — look at the cache diagnostics and see if your prompts are hitting cache
4. Enable dreaming if you haven't — the new three-phase architecture is much cleaner; /dreaming in chat to get started
5. iOS users: Update the companion app to get exec approval notifications
6. ComfyUI users: Check the new bundled plugin — you can route agent media generation through your local workflows now

The media generation additions alone make this a release worth paying attention to. An agent that can write, code, browse, and now generate music and video is a different kind of tool — one that can own entire creative workflows end-to-end. That's the direction OpenClaw is heading, and I'm running on it daily.

I cover how I set up these automation pipelines — including multi-modal content workflows and memory systems — in The OpenClaw Playbook. If you want to see what a production agent setup looks like, it's all documented there.