OpenClaw 2026.6.1: MiniMax M3 Is In, Skill Workshop Adds Approval Gates, and Windows Gets Native Container Support

What OpenClaw Is - For Readers Who Are New to It

OpenClaw is a self-hosted agent runtime and message router - "your own personal AI assistant, any OS, any platform, the lobster way." You run it on your own machine (macOS, Windows, or Linux) and it connects AI models to tools, workflows, and chat applications like WhatsApp, Telegram, iMessage, and Slack. Rather than paying a SaaS platform to own your conversation history and workflow logic, OpenClaw puts the orchestration layer on your own hardware using whatever AI model you choose.

The positioning is direct: OpenClaw Cloud starts at $59/month, a self-hosted VPS setup runs around $13/month, and traditional human assistants cost $3,000+/month for comparable coverage. The platform has scaled to hundreds of thousands of users, and recent releases have shifted from feature velocity toward reliability, audit trails, and governance - the hallmarks of a platform graduating from enthusiast use to production deployment.

MiniMax M3 Integration - What It Adds and How to Configure It

The headline feature of 2026.6.1 is MiniMax M3 provider support (PR #88860). MiniMax M3 was released on OpenRouter on May 31, 2026 and became generally available via MiniMax's own API on June 1. Its core strengths align with what OpenClaw agents need most: a 1M-token context window via MiniMax Sparse Attention, native multimodal input (text, image, and video in a single context), strong coding performance, and reliable agentic tool calling.

export MINIMAX_API_KEY=your_key_here
openclaw models set minimax/MiniMax-M3

openclaw configure
# Select MiniMax auth option
openclaw models set minimax-portal/MiniMax-M3

openclaw models list
# Should show minimax/MiniMax-M3 as available

Two provider IDs are available depending on your auth method. API-key setups use minimax/MiniMax-M3. OAuth portal setups (MiniMax's web-based auth) use minimax-portal/MiniMax-M3. If you see an error about MiniMax provider not configured, the fix is in 2026.1.12+ - upgrade first, then run openclaw configure and select a MiniMax auth option. The 2026.6.1 release also adds MiniMax M3 pricing metadata and PDF defaults to the provider catalog, so the model should surface automatically in model selection UIs after the upgrade.

Skill Workshop - Governed Agent Skill Creation

Skill Workshop is the most architecturally significant change in the 2026.6.1 release. Before this release, agent skills in OpenClaw were created and deployed without a formal review layer - an agent could propose a new skill and it would go into production without a human approving it. For personal use that is fine; for teams or any deployment touching real external channels (customer messages, file systems, credentials), it is a meaningful risk.

Skill Workshop adds a full approval pipeline for agent-created skills:

The OpenClaw 2026.5.31 update tracker described this direction well: "The next useful agent platform feature is not another prompt trick; it is change control." Skill Workshop is the materialization of that insight - governed skill edits, inspectable workboards, durable state, and explicit recovery paths are what separate a toy automation from a system you can deploy in a business context.

Windows Execution Containers - What Changed and What It Means

OpenClaw has technically supported Windows for some time, but execution of agent tasks on Windows required workarounds - typically running through WSL (Windows Subsystem for Linux) rather than native Windows processes. The 2026.6.1 work on Windows Execution Containers moves toward running agent workloads natively inside Windows containerized environments, eliminating the WSL dependency for most use cases.

The timing alongside Microsoft Build 2026 is deliberate. Build 2026 shipped Windows Agent Framework 1.0 - the OS-level API layer that allows AI agents to integrate with the Windows shell, task scheduler, and security model as registered system services. OpenClaw's Windows Execution Container work positions it to leverage Windows Agent Framework as a deployment target - meaning OpenClaw agents could eventually be registered as Windows Agent Framework services, appearing in the Windows task bar with managed permissions and Intune governance. That deeper integration is forward roadmap; the 2026.6.1 release delivers the foundational container work that makes it possible.

SQLite State Migration - Why It Matters for Reliability

Multiple subsystems have moved from filesystem-based state to SQLite-backed storage in 2026.6.1: plugin install ledgers, inbound channel queues (PRs #88794, #88797), iMessage monitor state, ACP metadata, session metadata, memory writes, and cron migrations. This might read as internal plumbing, but the practical impact is significant: SQLite-backed state survives restarts without duplicate filesystem scanning, recovers from crashes without losing queue positions, and provides transactional guarantees that raw filesystem operations cannot.

For users running OpenClaw as an always-on agent server - monitoring channels, executing scheduled tasks, handling inbound messages - the filesystem state model meant restarts could lose queue position or re-process already-handled messages. SQLite-backed queues and monitor state eliminate that class of reliability issue. If you run OpenClaw on a VPS or homelab server that restarts occasionally, this upgrade is worth applying for the reliability improvement alone.

Full Changelog Summary - 2026.6.1 Beta

How to Upgrade

npm install -g openclaw@beta

git pull origin main
npm install
npm run build

docker pull openclaw/openclaw:beta
docker compose pull && docker compose up -d

Note: 2026.6.1 is currently in beta (beta.1 and beta.2 released, stable not yet tagged on npm). The npm stable channel remains on 2026.5.28. If you are running a production setup and cannot tolerate beta instability, wait for the stable tag. If you are comfortable with beta software, the MiniMax M3 and Skill Workshop additions are worth upgrading for now.

Frequently Asked Questions