Openclaw — the open-source personal AI assistant sitting at 371,074 GitHub stars with 76,742 forks — unleashed a torrent of five beta releases over May 10-11, collectively branded v2026.5.10-beta. The weekend release cycle shipped changes across almost every subsystem: agent-to-agent protocols, Slack integration, Discord voice, Fal image generation, Fly Machines, the Control UI, and the build system’s migration to pnpm 11.
The cadence is remarkable even by Openclaw’s standards: beta.1 on May 10 at 13:16 UTC, beta.2 at 17:37, beta.3 on May 11 at 03:05, beta.4 at 15:40, and beta.5 at 16:13. That’s five releases in roughly 27 hours, each carrying production-quality changelogs spanning dozens of pull requests.
Agent-to-Agent: Deeper Conversation Chains
The most architecturally significant change in the beta cycle is the relaxation of agent-to-agent limits. Previously capped at 5 ping-pong turns, session.agentToAgent.maxPingPongTurns can now be configured up to 20 (default remains at 5) — building on the initial Openclaw controller design. This matters for complex multi-agent workflows where agents need sustained back-and-forth — collaborative debugging sessions, iterative code reviews, or negotiation tasks between specialized sub-agents.
Two companion features reinforce this direction:
tools.message.crossContextper-agent overrides — sandboxed and public agents can now restrict message sends to the current conversation without changing the global bot policy. This is crucial for multi-tenant gateways where different agents need different isolation guarantees.tools.message.actions.allowper-agent overrides — agents can expose and enforce send-only message tools granularly, giving operators fine-grained control over what each agent can communicate.
The message routing changes alone touch on safety patterns that enterprise deployments have been requesting since Openclaw crossed 300K stars.
Slack Gets Unfurl Control and Thread Broadcast
Three Slack improvements landed in beta.3 that together make Openclaw a much more polished Slack citizen:
| Feature | Description | Issue |
|---|---|---|
unfurlLinks / unfurlMedia |
Per-account config to suppress link and media previews in bot replies without workspace-wide settings | #48435 |
replyBroadcast |
Agents can opt into Slack’s parent-channel reply_broadcast behavior for thread replies |
#64365 |
| Mention metadata | Bot now distinguishes direct mentions from implicit thread wakes that mention someone else | #79025 |
The canonicalization of outbound DM routes in beta.3 also fixes a subtle but painful bug: message.send calls to D... targets no longer split the same Slack DM thread into separate channel sessions. (#80091)
Discord Voice Diagnostics and Telegram QA Automation
The beta cycle brought significant quality-of-life improvements for voice agents:
- Discord/voice realtime diagnostics — speaker turn timing, playback resets, barge-in detection, and audio cutoff analysis. These diagnostics make it feasible to debug voice interactions in production, a notoriously difficult problem.
talk.realtime.instructions— operators can now append realtime voice style guidance while preserving Openclaw’s built-in agent-consult defaults. (#79081)- Opus codec handling — test and source installs default to the pure-JS
opusscriptdecoder, avoiding slow native addon compiles outside dedicated voice-performance lanes. An opt-in native install script is available for production voice deployments.
On the QA side, a new Telegram live PR evidence automation system leases Convex credentials, captures Crabbox transcripts, generates motion GIF previews, and posts inline PR comments — a taste of Openclaw dogfooding its own platform for development workflows.
Infrastructure: Fly Machines, pnpm 11, and Local Models
Three infrastructure changes deserve attention:
-
Fly Machines container detection — Openclaw now detects Fly Machines from runtime environment variables and adjusts gateway bind and Bonjour defaults accordingly. (#80209) This makes deploying Openclaw on Fly.io genuinely zero-config.
-
pnpm 11 migration — the workspace package manager was upgraded to pnpm 11, with all Docker, install, update, and release workflows aligned to the new config surface. (#79414, #80588) Thanks to @altaywtf for both.
-
Local model server support — a new
localServicestartup option at the provider level allows on-demand local model servers to spin up before OpenAI-compatible requests, including one-shot model probes. This is a meaningful step toward fully offline agent operation.
Fal Image Generation Gets GPT Image 2 and Nano Banana 2
Providers/fal received a substantial update: GPT Image 2 and Nano Banana 2 reference-image edit requests are now routed to /edit with proper image_urls arrays. Nano Banana 2 edit geometry is enforced using aspect_ratio and resolution params, and input-image caps are lifted to 10 for GPT Image 2 and 14 for Nano Banana 2. Aspect-ratio hints are now allowed in edit mode. (#77295)
Control UI Recovery and Code Quality
The Control UI now shows a plain HTML recovery panel when the app module never registers — a small but critical fix for users who land on a blank dashboard. (#44107)
On the code quality front, the beta cycle enabled stricter Vitest lint rules (focused, disabled, conditional, hook, matcher hazards), additional oxlint rules for promise and TypeScript footguns, and stricter tsc compiler checks for implicit returns, side-effect imports, overrides, and unused production code. Logging saw targeted additions for model transport, payload, SSE, and code-mode diagnostics with redacted URL handling.
The 371K Milestone
At 371,074 stars, Openclaw has added approximately 2,600 stars since May 7 — part of the open-source agent ecosystem 2,600 stars since May 7 (when it was at 369,246). The growth rate shows no signs of slowing — the project now averages over 500 new stars per day, making it one of the fastest-growing open-source projects on GitHub by any measure.
The ecosystem around Openclaw continues to expand as well. The awesome-openclaw-skills list (VoltAgent) has passed 48,500 stars with 4,750 forks, and the official ClawHub skill directory sits at 8,574 stars with 1,324 forks — both indicators that the plugin and skill ecosystem is maturing alongside the core project.
Looking Ahead
With the v2026.5.10 beta cycle concluding and the project pushing toward a stable release, the trajectory is clear: Openclaw is evolving from a Claude Code harness into a full personal AI operating system with its own build system, deployment targets, voice infrastructure, and agent communication protocols. The agent-to-agent depth improvements, combined with per-agent message isolation and local model support, suggest the project is quietly preparing for a world where users run fleets of specialized agents rather than a single general-purpose assistant.