OpenClaw v2026.5.18: Real-Time Android Voice, Typed Tool Plugins, and a Faster Gateway
By AgentRiot
OpenClaw v2026.5.18 ships real-time voice sessions on Android, a new typed tool plugin SDK, faster gateway restarts, and a redesigned Mac settings experience. Here is what changed and why it matters.

The latest OpenClaw release brings voice to Android, typed plugins to the CLI, and measurable speedups to gateway restarts.
OpenClaw shipped v2026.5.18 on May 18, 2026, and the theme is clear: the platform is getting faster, more extensible, and more polished across every surface. This is a broad release with changes spanning the Android app, the CLI plugin system, the gateway core, the Mac desktop app, and the QA lab.
Here is what is actually in it.
Real-time voice on Android
The biggest user-facing change is Talk Mode on Android. It now uses real-time Gateway relay voice sessions with streaming microphone input, real-time audio playback, tool-result bridging, and on-screen transcripts.
What this means in practice: you can talk to your OpenClaw agent on Android the same way you would on Discord or the Mac app, with the agent hearing you in real time, running tools, and speaking back. The on-screen transcript gives you a text record of the conversation while the audio plays. Tool results bridge back into the voice stream, so the agent can look something up and continue the conversation without breaking the flow.
This is not a voice-note-to-text pipeline. It is a real-time duplex voice session, and it brings Android up to parity with the desktop and Discord experiences.
Typed tool plugins
The CLI now supports defineToolPlugin, along with openclaw plugins build, validate, and init. This is a structured way to build simple tool plugins with generated manifest metadata, optional tool declarations, and context factories.
Before this, plugin development required manual manifest authoring and careful schema alignment. The new typed plugin flow generates the manifest from your TypeScript definitions, validates the plugin before you publish it, and provides a standard context factory pattern for tool execution.
If you have been holding off on building custom tools because the plugin boilerplate was too heavy, this removes most of that friction.
Faster gateway restarts
Gateway restart latency got attention on multiple fronts. Startup logging and plugin-service startup now overlap with channel sidecars, reducing the time from process start to ready state. The update-check startup is deferred until after readiness, so package update checks no longer block the sidecar-ready signal.
There is also new benchmark tooling: pnpm test:restart:gateway measures repeated restart readiness, downtime, trace costs, and resource slopes. If you are running OpenClaw at scale and optimizing for fast recovery, you now have metrics to work with.
Mac app settings redesign
The Mac app settings pages were rebuilt with consistent card layouts, cached navigation, cleaner permissions and voice panes, and steadier spacing around the native sidebar. Settings tabs stay mounted when you switch between them, eliminating the blank-and-reload flicker that plagued earlier versions.
The Channels settings now open faster by deferring config-schema work and showing compact quick settings instead of the full generated schema. The Config settings load selected paths on demand rather than fetching the entire generated config tree upfront.
These are quality-of-life improvements that add up if you spend time configuring agents in the Mac app.
Browser dialog handling
The browser automation surface now surfaces pending and recently handled modal dialogs in snapshots. If an action opens a modal, the system returns blockedByDialog instead of hanging or failing opaquely. You can answer pending dialogs with browser dialog --dialog-id.
This makes browser automation more reliable for workflows that encounter confirmation dialogs, login prompts, or cookie banners.
New skills
Several new skills shipped in this release:
- Meme-maker: curated template search, local SVG/PNG rendering, Imgflip hosted rendering, and Know Your Meme provenance links.
- Node inspector debugging: for debugging Node.js processes.
- Fused diagram generation: for creating diagrams from code or descriptions.
- Throwaway spike workflow: for quick experiments that you do not intend to keep.
- Python debugging: pdb, breakpoint(), post-mortem inspection, and debugpy remote attach.
The Obsidian skill was updated to target the official obsidian CLI instead of the third-party obsidian-cli. The repo-local Codex closeout review skill was renamed to autoreview.
QA-Lab expansion
The QA lab grew significantly. New runtime parity scenarios include first-hour 20-turn and optional 100-turn tests, with tier metadata for standard and soak gates. A live-only Codex Pi-shaped Read vocabulary canary catches native workspace-read prompt compatibility drift. Runtime tool fixture scenarios cover Codex-native workspace tools, OpenClaw dynamic tools, and optional plugin-backed tools.
Tool coverage is now exposed through openclaw qa coverage --tools, and the standard Codex-vs-Pi tier hard-gates required OpenClaw dynamic runtime-tool drift with a blocking release-check verifier.
Proxy and security
HTTPS managed forward-proxy endpoints are now supported, with scoped proxy.tls.caFile CA trust for proxy endpoint TLS. The Docker/Podman image build accepts OPENCLAW_IMAGE_APT_PACKAGES as a runtime-neutral build arg for extra apt packages.
Fixes worth noting
- Discord voice sessions now keep hearing follow-up turns with OpenAI realtime, with prebuffered assistant playback to avoid choppy starts.
- Telegram delivers generated media completions back into forum topics by preserving topic IDs.
- Codex app-server hydrates inbound image attachments before queued runs, so Responses-backed agents receive Discord images as native vision input.
- TTS directives apply before message-tool sends reach delivery, so voice notes attach correctly instead of leaking raw tags.
- Sharp is now installed with the root package, with fallback to sips, ImageMagick, GraphicsMagick, or ffmpeg for image resizing.
What to test before upgrading
- Android voice: If you use Talk Mode, verify that real-time sessions connect and that tool results bridge back into the audio stream.
- Gateway restart timing: If you run multiple gateway instances, measure restart latency with the new benchmark tooling and confirm sidecars report ready in the expected window.
- Browser automation: If you use browser tools, test workflows that encounter modals and confirm
blockedByDialogsurfaces correctly. - Plugins: If you build custom plugins, try the new
defineToolPluginflow and validate thatopenclaw plugins buildgenerates the expected manifest. - Mac app settings: If you use the Mac app, navigate through settings and confirm the cached navigation and mounted tabs feel responsive.
The bottom line
OpenClaw v2026.5.18 is a release about maturity. The Android voice work brings a major platform to parity. The typed plugin SDK lowers the barrier to extending the platform. The gateway restart optimizations and QA lab expansion show the team is thinking about production scale. And the Mac app polish demonstrates attention to the daily user experience.
If you are running OpenClaw in production, the restart and Telegram fixes alone are worth the upgrade. If you are building on the platform, the typed plugin SDK is the feature to explore first.

