OpenClaw v2026.3.22 Release Guide: 45 New Features + 13 Breaking Changes You Need to Know

OpenClaw v2026.3.22 ships 45+ new features, 13 breaking changes, and 82 bug fixes. This guide covers ClawHub skills marketplace, multi-model support, pluggable sandboxes, and a complete upgrade walkthrough.

BibiGPT Team

OpenClaw v2026.3.22 Release Guide: 45 New Features + 13 Breaking Changes You Need to Know

"They used to ship every 1-2 days — the community was totally spoiled." After a 10-day gap, the lobster just dropped its biggest update ever.

OpenClaw v2026.3.22, released on March 23, 2026, brings 45+ new features, 13 breaking changes, 82 bug fixes, 20 security patches, and 15 performance improvements — earning 294 reactions on GitHub. The Chinese tech community called it "the most powerful update in history," while Linux DO helpfully tagged it "update with caution."

This guide helps you understand: what changed, whether it affects you, and how to upgrade safely.

Related reading: OpenClaw + bibigpt-skill Complete Guide (March 2026 Update) | OpenClaw 250K Stars: How BibiGPT Wins in the Agent Era


Five Major New Features: The Highlights

1. ClawHub Skills Marketplace Goes Native

The single biggest change in this release. openclaw plugins install now prioritizes ClawHub over npm, falling back to npm only when a package isn't found on ClawHub. Full skill management commands are now available:

# Search for skills
openclaw skills search <keyword>

# One-click install
openclaw skills install <skill-name>

# Check for updates
openclaw skills update

Even more notable: cross-ecosystem compatibility. ClawHub can discover and import skill packages from Claude, Codex, and Cursor, automatically mapping them into OpenClaw format. That's 4,000+ community skills at your fingertips.

2. Agent Timeout Jumps to 48 Hours

Default agent timeout goes from 10 minutes to 48 hours with zero configuration changes. Batch jobs, long documents, complex workflows — no more mid-task disconnections.

3. Pluggable Sandbox Backends (OpenShell + SSH)

The sandbox execution environment now supports pluggable backends. The first two: SSH and OpenShell, with mirror and remote workspace modes.

In practice: your agent can now SSH into remote machines and execute tasks directly. No more workarounds for running jobs on servers.

4. Massive Model Ecosystem Expansion

This release covers virtually every major LLM provider:

  • Anthropic Vertex AI: Call Claude through Google Vertex AI (no longer limited to direct Anthropic API)
  • OpenAI: Default model upgraded to gpt-5.4, with forward-compat for gpt-5.4-mini and gpt-5.4-nano
  • MiniMax: Default upgraded from M2.5 to M2.7, with high-speed mode
  • xAI Grok: Synced to latest Pi-backed Grok model family
  • Z.AI: Added GLM 4.5/4.6 series
  • Xiaomi: Switched to v1 API, added MiMo V2 Pro and MiMo V2 Omni
  • Mistral: Updated pricing and model metadata
  • GitHub Copilot: Forward-compatible dynamic model IDs

Bottom line: use whatever LLM you want — OpenClaw handles the integration.

5. Built-in Web Search Triple Pack

Exa, Tavily, and Firecrawl ship as first-party built-in plugins. Your agent finally has reliable native search capabilities without third-party configuration headaches.

Try pasting your video link

Supports YouTube, Bilibili, TikTok, Xiaohongshu and 30+ platforms

+30

13 Breaking Changes Explained — Read Before Upgrading

If you're wondering "should I update?", this table has your answer:

#ChangeWho's AffectedWhat to Do
1ClawHub over npmPlugin installsUsually transparent; npm fallback still works
2Chrome MCP legacy path removedBrowser integration usersRun openclaw doctor --fix
3Plugin SDK → openclaw/plugin-sdk/*Skill developersFollow migration docs
4nano-banana-pro removedImage generation usersSwitch to agents.defaults.imageGenerationModel
5CLAWDBOT_* / MOLTBOT_* env vars removedAll legacy config usersRename to OPENCLAW_*
6.moltbot directory removedEarly adoptersMigrate to ~/.openclaw
7Matrix plugin rewrittenMatrix usersFollow migration guide
8Discord command deployment changeDiscord usersUsually transparent
9Sandbox env var injection blockedDevelopersMAVEN_OPTS/SBT_OPTS/GRADLE_OPTS no longer pass through
10Message discovery API changePlugin developersUse describeMessageTool(...)
11Image generation tool standardizedImage generation usersUse image_generate tool
12Voice-call webhook hardenedVoice call usersEnsure provider signature headers are correct
13Agent media reply refactorPlugin developersUse details.media path
# 1. Update OpenClaw
openclaw update

# 2. Auto-fix most breaking changes
openclaw doctor --fix

# 3. Verify everything works
openclaw status

Important: If you're still using the .moltbot directory or CLAWDBOT_* / MOLTBOT_* environment variables, you must migrate now — there's no fallback path.


20+ Messaging Platform Enhancements

Nearly every major platform got improvements:

  • Matrix (new): Official matrix-js-sdk plugin with bot conversations, mention gating, bot-to-bot chat, and private networks
  • Telegram: Custom Bot API endpoints, auto topic labels, topic editing, silent error replies
  • Feishu/Lark: Structured approval cards, ACP support, reasoning stream rendering
  • Discord: Carbon reconcile command deployment, component auth fixes
  • Slack: @slack/bolt startup stability fixes
  • WhatsApp: Reconnect mechanism fixes, credential handling improvements
  • Android: System-aware dark theme, call log search, SMS search, camera memory optimization
  • Mattermost: Thread reply mode off support, DM retry mechanism
  • WeChat/WeCom/QQ/DingTalk: Community plugins listed in official docs

Security + Performance Highlights

Security (20+ Fixes)

The largest security release in OpenClaw history:

  • Windows SMB credential protection: Blocks remote file:// URLs and UNC paths
  • Unicode hidden text detection: Escapes blank Hangul filler in exec approval prompts
  • Nostr DM security: Enforces DM policy before decryption with rate/size guards
  • Device pairing hardening: Binds iOS setup codes to intended node profiles
  • SSRF policy enforcement: Comprehensive hardening across browser and network layers

Performance (15+ Improvements)

  • Gateway startup: Loads from compiled dist, WhatsApp cold-start drops from tens of seconds to seconds
  • Provider discovery lazy-loading: No more preloading all providers at startup
  • Model catalog caching: Caches by config state to avoid repeated startup work
  • Telegram message latency: Normal replies no longer rebuild the full model catalog
  • CLI startup optimization: Lazy-loading for help and channel-add paths

Top 5 ClawHub Skills to Install After Upgrading

With ClawHub now built in, discovering and installing skills has never been easier. Here are 5 worth grabbing right after your v2026.3.22 upgrade:

1. bibigpt-skill — AI Video Summary

Supports 30+ platforms including Bilibili, YouTube, TikTok, Xiaohongshu, and podcasts. OpenClaw's native summarize only covers YouTube — bibigpt-skill fills the gap for everything else. Trusted by over 1 million users with over 5 million AI summaries generated.

openclaw skills install bibigpt-skill

# Use it right away
@OpenClaw summarize this video with bibi https://www.youtube.com/watch?v=xxx

2. Context7 — Live Library Docs

Fetches up-to-date documentation and code examples while you code. No more outdated APIs.

3. web-search — Enhanced Search

Works alongside built-in Exa/Tavily/Firecrawl for deeper information retrieval.

4. notion-connector — Bidirectional Notion Sync

Sync your agent's work directly to Notion for seamless knowledge management.

5. github-pr-review — Automated PR Review

Let your agent review code and catch issues before they ship.

See BibiGPT's AI Summary in Action

Bilibili: GPT-4 & Workflow Revolution

Bilibili: GPT-4 & Workflow Revolution

A deep-dive explainer on how GPT-4 transforms work, covering model internals, training stages, and the societal shift ahead.

总结

本视频深入浅出地科普了ChatGPT的底层原理、三阶段训练过程及其涌现能力,并探讨了大型语言模型对社会、教育、新闻和内容生产等领域的深远影响。作者强调,ChatGPT的革命性意义在于验证了大型语言模型的可行性,预示着未来将有更多更强大的模型普及,从而改变人类群体协作中知识的创造、继承和应用方式,并呼吁个人和国家积极应对这一技术浪潮。

亮点

  • 💡 核心原理揭秘: ChatGPT的本质功能是"单字接龙",通过"自回归生成"来构建长篇回答,其训练旨在学习举一反三的通用规律,而非简单记忆,这使其与搜索引擎截然不同。
  • 🧠 三阶段训练: 大型语言模型经历了"开卷有益"(预训练)、"模板规范"(监督学习)和"创意引导"(强化学习)三个阶段,使其从海量知识的"懂王鹦鹉"进化为既懂规矩又会试探的"博学鹦鹉"。
  • 🚀 涌现能力: 当模型规模达到一定程度时,会突然涌现出理解指令、理解例子和思维链等惊人能力,这些是小模型所不具备的。
  • 🌍 社会影响深远: 大型语言模型将极大提升人类群体协作中知识处理的效率,其影响范围堪比电脑和互联网,尤其对教育、学术、新闻和内容生产行业带来颠覆性变革。
  • 🛡️ 应对未来挑战: 面对技术带来的混淆、安全风险和结构性失业等问题,个人应克服抵触心理,重塑终身学习能力;国家则需自主研发大模型,并推动教育改革和科技伦理建设。

#ChatGPT #大型语言模型 #人工智能 #未来工作流 #终身学习

思考

  1. ChatGPT与传统搜索引擎有何本质区别?
    • ChatGPT是一个生成模型,它通过学习语言规律和知识来“创造”新的文本,其结果是根据模型预测逐字生成的,不直接从数据库中搜索并拼接现有信息。而搜索引擎则是在庞大数据库中查找并呈现最相关的内容。
  2. 为什么说大语言模型对教育界的影响尤其强烈?
    • 大语言模型能够高效地继承和应用既有知识,这意味着未来许多学校传授的知识,任何人都可以通过大语言模型轻松获取。这挑战了以传授既有知识为主的现代教育模式,迫使教育体系加速向培养学习能力和创造能力转型,以适应未来就业市场的需求。
  3. 个人应该如何应对大语言模型带来的社会变革?
    • 首先,要克服对新工具的抵触心理,积极拥抱并探索其优点和缺点。其次,必须做好终身学习的准备,重塑自己的学习能力,掌握更高抽象层次的认知方法,因为未来工具更新换代会越来越快,学习能力将是应对变革的根本。

术语解释

  • 单字接龙 (Single-character Autoregressive Generation): ChatGPT的核心功能,指模型根据已有的上文,预测并生成下一个最有可能的字或词,然后将新生成的字词与上文组合成新的上文,如此循环往复,生成任意长度的文本。
  • 涌现能力 (Emergent Abilities): 指当大语言模型的规模(如参数量、训练数据量)达到一定程度后,突然展现出在小模型中未曾察觉到的新能力,例如理解指令、语境内学习(理解例子)和思维链推理等。
  • 预训练 (Pre-training): 大语言模型训练的第一阶段,通常称为“开卷有益”,模型通过对海量无标注文本数据进行单字接龙等任务,学习广泛的语言知识、世界信息和语言规律。
  • 监督学习 (Supervised Learning): 大语言模型训练的第二阶段,通常称为“模板规范”,模型通过学习人工标注的优质对话范例,来规范其回答的对话模式和内容,使其符合人类的期望和价值观。
  • 强化学习 (Reinforcement Learning): 大语言模型训练的第三阶段,通常称为“创意引导”,模型根据人类对它生成答案的评分(奖励或惩罚)来调整自身,以引导其生成更具创造性且符合人类认可的回答。

Want to summarize your own videos?

BibiGPT supports YouTube, Bilibili, TikTok and 30+ platforms with one-click AI summaries

Try BibiGPT Free

Wrapping Up

OpenClaw v2026.3.22 is the biggest update to date. ClawHub transforms the skill ecosystem, 48-hour agent timeout unlocks long-running tasks, pluggable sandboxes with SSH open up remote execution, and full multi-model coverage means you're never locked into a single provider.

Yes, 13 breaking changes sounds scary — but openclaw doctor --fix handles most of them. If you've been putting off migrating from .moltbot and legacy env vars, now's the time.

After upgrading, take ClawHub for a spin — 4,000+ skills are waiting.

Related reading:

Experience BibiGPT now

Ready to try these powerful features? Visit BibiGPT and start your intelligent audio/video summarization journey!

Get started