OpenClaw v2026.4.26: Better Browser Talk, Safer Migration, and a Core That Gets Out of Your Way
OpenClaw v2026.4.26 landed on April 27, 2026 at 8:11 p.m. America/Chicago time, which was April 28, 2026 at 01:11 UTC. This is the release to grab if you care about three things: browser Talk that feels less fragile, migration tooling you can test before a real move, and fewer weird surprises in plugin-heavy installs.
The changelog is huge, but the practical story is not. Browser voice got cleaner transport plumbing and a batch of fixes around Google Live sessions. openclaw migrate showed up with dry-run, JSON output, backups, and importers for Claude and Hermes. Memory search and plugin startup both got a round of work aimed at the annoying failures that waste an afternoon.
Browser Talk is finally worth retesting
The headline feature is the new browser realtime path for Talk, including Google Live browser sessions with constrained ephemeral tokens and a gateway relay for backend-only realtime voice plugins.
You do not need to care about the transport contract to benefit from it. What matters is that browser voice now looks more like a real product surface and less like a side route. The same release also fixes several Talk-specific headaches right away: it keeps Google Live browser sessions on the WebSocket path, validates the browser endpoint, caps relay sessions per browser connection, and removes stale browser-native voice buttons that did not belong there.
If you tried browser Talk before and walked away because it felt half-finished, try it again on this build.
Migration is no longer a side script
OpenClaw now ships a bundled Claude importer and a broader openclaw migrate flow. It can preview a plan, run dry, emit JSON, create a backup before touching anything, and pull in pieces from Claude Code, Claude Desktop, and Hermes.
That matters because migration tools usually fail in boring ways. They copy some settings, miss the sharp edges, and leave you sorting out the mess by hand. This release looks more serious than that. It gives you a way to inspect the move before you commit to it.
openclaw migrate --dry-run
openclaw migrate --json
If your current setup has MCP servers, saved instructions, provider config, or credentials scattered across too many places, start here before you do anything destructive.
Memory and config-heavy setups got the useful kind of attention
The memory work is easy to miss, but I would not skip it. OpenClaw now supports separate query and document input types for OpenAI-compatible memory embeddings, and it adds retrieval prefixes for nomic-embed-text, qwen3-embedding, and mxbai-embed-large.
If you run your own embedding stack, you already know the problem: retrieval can get worse without failing loudly. These changes give you more control over how searches are shaped, which should help on installs where recall quality depends on provider quirks.
The Control UI also gets a raw config diff panel that parses JSON5, hides secrets until you reveal them, and stops firing fake raw-edit callbacks when you open the panel. That is the kind of small fix you feel every time you touch config.
There is a lot of core cleanup here
Cerebras is now bundled as a provider. Plugin manifests take on more of the provider metadata and request-family hints that used to live in core. Layered plugin stage directories are supported. Startup and auto-enable paths reuse config snapshot manifests more cleanly. Foreground gateway runs on low-memory Linux hosts avoid one ugly self-respawn path.
Most people will never read those bullets twice. They will notice the result. Fewer startup surprises. Fewer install edge cases. Less time wondering why a custom setup drifted into a weird state.
What I would test right after updating
- Run one real browser Talk session. If you use Google Live, make sure the session comes up cleanly and stays on the expected transport.
- Run
openclaw migrate --dry-run. Even if you are not migrating today, this tells you what OpenClaw can already see and import from your existing setup. - Check one memory query you know well. If you use custom embeddings, compare the result quality before and after the update.
- Open the raw config diff after a harmless edit. Make sure the panel shows the real change and keeps secrets hidden by default.
- Restart the gateway once on purpose. This release touches plugin discovery, startup, install paths, and several provider plumbing layers. A clean restart is worth verifying while you are watching.
- If you use Matrix encryption, run
openclaw matrix encryption setup. This release adds a one-flow setup path, and it is worth checking while the change is fresh.
Should you update now?
Yes if browser Talk matters to you, if a migration from Claude or Hermes is on your list, or if you run a custom memory or plugin stack.
Also yes if you maintain OpenClaw for other people. A lot of this release is cleanup work, and cleanup work is what keeps support queues short.
This is not a flashy release. It is the kind that makes OpenClaw easier to trust after the first hour.
Full changelog: View v2026.4.26 on GitHub
Generated release page: Read the structured release notes
- Fred