OpenClaw 2026.6.1 Ships With Skill Workshop Governance, Workboard Orchestration, and a Harder Runtime

The latest stable release of the open-source AI agent framework adds governed skill creation, multi-agent workboards, SQLite-backed state, and a sweeping reliability push across channels, providers, and plugins. The beta line is already moving toward operator install policies and tighter privacy gates.

---

OpenClaw shipped v2026.6.1 on June 3, 2026, and the release is less about headline features than about making the agent harder to break. The changelog is enormous, spanning cleaner recovery from interrupted tool calls, bounded timeouts across nearly every provider and channel plugin, SQLite-backed state for restarts, and a new Skill Workshop that turns agent-created skills into reviewable proposals before they touch production workflows.

The project also pushed v2026.6.2-beta.1 the same day, so the beta line is moving faster than stable. That is worth watching. Here is what is in the stable release, what is cooking in beta, and why it matters for anyone running agents unattended.

What Is OpenClaw?

OpenClaw is an open-source AI agent framework, self-described as "your own personal AI assistant, any OS, any platform, the lobster way." It runs on macOS, Windows, Linux, iOS, and Android, with gateway support for Telegram, Discord, Slack, WhatsApp, Signal, iMessage, Microsoft Teams, Google Chat, Feishu, and more. It is provider-agnostic, works with local and hosted models, and supports plugins, skills, cron jobs, memory providers, and multi-agent delegation.

Skill Workshop: Governed Skill Creation

The standout addition in 2026.6.1 is Skill Workshop, a governed flow for creating and reviewing agent-generated skills.

Previously, an agent could create a skill during a session and it would immediately become available. Skill Workshop changes that. Now when an agent proposes a skill, the proposal enters a pending state with:

• A proposal list visible in the Control UI
• Review actions (approve, reject, quarantine)
• Revision handoff with versioned, dated frontmatter
• Support file attachments with scanner checks and hash safeguards
• Rollback metadata in case a approved skill needs reversal
• A dedicated skill_workshop agent tool that surfaces guidance in Codex app-server prompts

The Control UI adds a styled dashboard, a "today" view for pending proposals, a revision dialog, searchable file previews, and reusable session routing. A dedicated guide covers the full flow from CLI to Gateway to agent tool behavior.

This matters because agent-generated skills are powerful and risky. A skill that automates file deletion or API calls with the wrong scope can do real damage. Skill Workshop makes the operator the gatekeeper without blocking the agent's ability to learn.

Workboard: Multi-Agent Orchestration

Workboard is new orchestration infrastructure for multi-agent planning and run tracking. It adds task-backed board runs, task comments in edit modals, and keyboard movement controls in the Control UI.

The pitch is simple: when you delegate work across multiple agents, you need a surface to track what each agent is doing, hand off context between them, and review results. Workboard provides that. It is still early, the beta adds more keyboard controls and tighter card operations, but the foundation is there.

SQLite-Backed State: Surviving Restarts

A major architectural shift in 2026.6.1 is moving runtime state from in-memory and filesystem structures to SQLite. The affected areas include:

• Plugin install indexes
• Inbound channel queues
• iMessage monitor state
• ACP metadata
• Session metadata
• Memory writes
• Cron migrations

The practical effect: restart the gateway or the CLI, and state recovers with less duplicate filesystem scanning, less lost queue state, and less chance of a cron job firing twice because the ledger was in RAM.

Reliability: Bounded Timeouts Everywhere

The 2026.6.1 release adds timeouts, retries, and bounds across an exhaustive list of paths. A partial inventory:

• Providers: generated media downloads, OAuth requests, device-code lifetimes, local service probes, TTS, music, workflow polling
• Channels: Telegram, Discord, WhatsApp, Signal, Feishu, Google Chat, Microsoft Teams, QQBot, Nostr, Zalo, Nextcloud
• Agents: interrupted tool calls, stale session bindings, compaction handoffs, media delivery retries, Codex app-server idle timers
• Release/CI: readiness probes, artifact checks, status polling, docker cleanup, log drains

The theme is production hardening. OpenClaw is being run by operators who leave agents unattended for hours or days. Unbounded hangs are unacceptable in that context.

External Plugins: Copilot and Tokenjuice

Two previously bundled capabilities have been externalized as official plugins:

• @openclaw/copilot — the GitHub Copilot agent runtime
• @openclaw/tokenjuice — the Tokenjuice integration

Both are published to npm and ClawHub. The move lets OpenClaw ship a leaner core while keeping the integrations official and maintained.

Provider Updates

New and updated provider support in 2026.6.1:

• MiniMax M3 added
• Google/Vertex catalog fixes, defaults resolved to google-generative-ai
• OpenRouter SQLite model caching
• Copilot Claude 1M capabilities preserved
• Foundry reasoning alignment, DeepSeek V4 thinking params skipped on fallback
• OpenAI response replay guards
• Gemini stop sequences forwarded
• Kimi incompatible Anthropic cache markers stripped

What Is Cooking in Beta

The beta line (v2026.6.1-beta.3 and v2026.6.2-beta.1) adds several features that have not hit stable yet:

Operator install policy replaces dangerous-code scanner. Plugin and skill installs now use an explicit operator-selected policy instead of the old automatic dangerous-code scanner. The change adds clearer doctor checks, CLI wiring, ClawHub metadata paths, and troubleshooting surfaces. The security model shifts from "scan and hope" to "operator decides."

Data-handling conformance checks. New policy gates reject unsupported policy keys, unsafe exec approval precheck environments, malformed script limits, and suspicious gateway startup configs.

Chat and Control UI improvements. Streaming text preservation, incremental chat history rendering, local draft typing state, first-output latency instrumentation, calmer composer controls, and better ACK timing metadata.

Android companion shell navigation improvements.

iOS native iPad layouts and hosted push relay defaults.

The Bigger Picture

OpenClaw's June releases reveal a project maturing from "feature-rich agent framework" to "production-grade agent infrastructure." The emphasis on governance (Skill Workshop), observability (Workboard), state durability (SQLite), and bounded runtime behavior (timeouts everywhere) aligns with what enterprise operators are asking for: agents that do not hang, do not lose state, and do not silently install dangerous code.

The beta's operator install policy is particularly notable. It acknowledges a hard truth about AI agents: automatic security scanning of LLM-generated code is imperfect. Giving the operator an explicit policy decision, with clear recovery paths, is a more honest security model.

OpenClaw 2026.6.1 is available now via npm (npm install -g openclaw@latest) and the project's GitHub releases page. The beta can be installed with npm install -g openclaw@beta.