OpenClaw v2026.4.24: Google Meet agents, full-agent voice, and DeepSeek V4 land in one release

TL;DR

OpenClaw v2026.4.24 (tagged Apr 24, 2026) is the release where the open-source personal AI agent stops being a chat box and starts being a participant. Three things actually matter: a bundled Google Meet plugin that lets agents join meetings with personal Google auth, a full-agent voice loop across Talk Mode, Voice Call and Meet, and the DeepSeek V4 Flash + V4 Pro models in the bundled catalog — with V4 Flash now the onboarding default. Browser automation gets coordinate clicks and a 60s action budget, startup is materially lighter, and there is one breaking Plugin SDK change every plugin author needs to know about.

What's new

• Google Meet bundled participant plugin — personal Google OAuth, Chrome and Twilio realtime transports, paired-node Chrome support (Parallels/BlackHole/SoX), artifact + attendance exports (recordings, transcripts, smart notes), and recovery tooling like googlemeet doctor --oauth and recover_current_tab.
• Full-agent voice consults — Talk, Voice Call and Google Meet now expose openclaw_agent_consult, so realtime voice loops can hand off to the full OpenClaw agent with tool access instead of a thin voice-only model.
• DeepSeek V4 in the box — deepseek/deepseek-v4-flash (default) and deepseek/deepseek-v4-pro ship in the bundled catalog with the thinking/replay contract fixed for follow-up tool-call turns.
• Realtime transports — browser WebRTC sessions backed by OpenAI Realtime, plus a new Gemini Live realtime voice provider for backend Voice Call and Meet bridges with bidirectional audio and function calls.
• Browser automation — viewport coordinate clicks (openclaw browser click-coords), browser.actionTimeoutMs with a 60s default, per-profile headless overrides, smarter tab reuse and recovery.
• Lighter startup — static model catalogs, manifest-backed model rows, lazy provider dependencies, and external runtime-dependency repair for packaged installs. node-llama-cpp is no longer installed by default.
• Other upgrades — new Gradium bundled TTS for voice notes and telephony, refined Tool Access panel, a Steer action on queued messages, hybrid memory search exposing raw vectorScore/textScore, and broad client fixes across Slack, Telegram, Discord, WhatsApp and MCP.

Why it matters

For two years the agent conversation has been stuck on text chat with bolted-on tool calling. v2026.4.24 quietly removes that ceiling: voice loops are no longer a separate, dumber surface — they consult the same agent that runs in your terminal, with the same tools and memory. Combined with a participant plugin that can actually sit in a Meet call, this is the first OpenClaw release where the right mental model isn't “a chatbot with shell access” but “a coworker that can show up to your meetings.”

The DeepSeek V4 default flips the cost equation too. Before this release the path of least resistance for a strong agent backend was a Claude or GPT key. Now an open-source default runs on a 1M-token context window with delivery quality approaching Claude Opus 4.6 non-thinking mode — for a fraction of the API spend.

Technical facts

Architecture: V4 ships a novel attention mechanism — token-wise compression combined with DeepSeek Sparse Attention (DSA) — that holds the 1M context window without the linear compute and memory blow-up of vanilla attention.

Other concrete numbers in this release:

• Browser action budget: browser.actionTimeoutMs default raised to 60s so healthy long waits no longer fail at the client transport boundary.
• Bundled Pi packages bumped to 0.70.2.
• Codex sessions: openai-codex/gpt-5.5 aligned with Codex's 272K runtime cap inside a 400K native context window.
• Voice Call telephony: ElevenLabs requests now use pcm_22050 for PCM telephony synthesis instead of getting MP3 back.
• The OpenClaw repo crossed roughly 364k stars / 74.5k forks by release time.

Comparison

vs the previous OpenClaw line: voice was passive TTS/STT — now it's a true realtime loop with full tool access. Browser RPA used to die at the 60s transport boundary — tunable now, and tab recovery survives crashed pages.

vs Claude / ChatGPT desktop apps: those are still conversational front ends. OpenClaw runs locally and executes — files, shell, browser, APIs — across messaging surfaces (WhatsApp, Telegram, Slack, Discord, Signal, Matrix, Feishu, and more).

vs coding-only agents (Claude Code, OpenCode, CodeBuddy): DeepSeek V4 was specifically tuned for those four agent harnesses, but OpenClaw is the universal one — coding is one skill family among 100+.

Use cases

• Meeting agent. Agent joins a Google Meet via the participant plugin, listens through the realtime voice loop, fetches answers via tools mid-meeting, then exports the recording, transcript and smart notes when it ends.
• Phone IVR replacement. Voice Call bridges Twilio/Telnyx/Plivo to a V4-Flash agent that qualifies leads, answers billing questions, and escalates via tools — with a dry-run voicecall smoke command to check provider readiness before placing a live call.
• Repo-scale code work. Hand the agent an entire monorepo; V4-Pro's 1M-token window lets it reason across the whole codebase before a refactor or debug pass.
• Unattended browser RPA. Checkout automation, KYC capture, and dashboard ops with coordinate clicks, the new 60s budget, and tab recovery covering crashed pages.
• Multi-channel ops bot. One agent across Slack threads, Discord DMs, Telegram topics, and WhatsApp voice notes — voice notes are auto-transcribed before agent dispatch.

Limitations & pricing

Limitations. DeepSeek V4-Pro still trails Claude Opus 4.6 thinking mode on Agentic Coding. V4-Flash has gaps on high-difficulty agent tasks vs Pro. V4 thinking sessions require the prior reasoning_content to be replayed on tool-call follow-ups — OpenClaw handles this internally, but custom integrations must comply.

Security. OpenClaw needs broad system access by design, which keeps it on the radar of cybersecurity teams — Cisco AI security previously flagged a malicious third-party skill, and Chinese authorities banned OpenClaw on state-enterprise office computers in March 2026. Treat skills like you treat npm packages: vet what you install.

Breaking change. Plugin SDK — the Pi-only api.registerEmbeddedExtensionFactory(...) path is gone. Tool-result rewrites must move to api.registerAgentToolResultMiddleware(...) with contracts.agentToolResultMiddleware declaring the targeted harnesses, so transforms run consistently across Pi and Codex app-server dynamic tools. The /models add chat action is also deprecated.

Pricing. OpenClaw itself is free under MIT, self-hosted. DeepSeek V4 API: Flash is the highly economical default, Pro is higher cost/latency for complex agent work. Both support 1M context and Thinking / Non-Thinking modes through OpenAI ChatCompletions or Anthropic-compatible endpoints.

What's next

Two dates worth pinning:

• Jul 24, 2026, 15:59 UTC — the legacy deepseek-chat and deepseek-reasoner endpoints are retired permanently. They currently route to V4-Flash; migrate explicitly to deepseek-v4-pro or deepseek-v4-flash before then.
• Foundation handoff — original creator Peter Steinberger joined OpenAI in February 2026, with stewardship moving to a new non-profit foundation. The 2026.4.24 changelog already lays groundwork: OTEL trace-context carriers are scaffolded for “future span correlation”, with opt-in diagnostics.otel.captureContent controls.

How to update:

openclaw update openclaw --version
npm install -g openclaw@latest

Docker users: pull the latest image and restart.

Sources: OpenClaw v2026.4.24 release notes, DeepSeek V4 Preview, OpenClaw DeepSeek docs, Blockchain.News analysis.