- openclaw 2026.6.6-beta.2 — Highlights - Security boundaries are substantially tighter across transcripts, sandbox binds, host environment inheritance, MCP stdio, Codex HTTP access, native search policy, elevated sender checks,...
- fix(exec): rebuild command authorization on the Tree-sitter command planner — Summary This PR rebuilds POSIX command authorization on the Tree-sitter command model. The ad hoc shell splitter that previously decided allowlist matches and allow-always persistence is replaced by...
- fix(moonshot): backfill reasoning_content on assistant tool-call replay messages — Summary Fixes the reasoning_content is missing in assistant tool call message 400 error when using Moonshot/Kimi models with thinking enabled in long sessions, after LCM compaction, cross-model...
- [Bug]: 同一个对话框向上滚动时看不到最开始的问答信息 — Bug type Regression (worked before, now fails) Beta release blocker No Summary openclaw为什么聊着聊着同一个对话框以前的聊天内容回滚不到了 Steps to reproduce openclaw为什么聊着聊着同一个对话框以前的聊天内容回滚不到了 Expected behavior...
- tasks audit: 154 findings on a healthy system — false 'dead agent' signals for long-running subagents + timestamp noise flood — Summary openclaw tasks audit produces so much noise that real failures are undetectable. On a single-user gateway with 4 actively running subagents, the audit returns 154 findings — none of which...
- [Bug]: `openai-responses` adapter sends system prompt in `input` instead of `instructions` — Bug type Regression (worked before, now fails) Beta release blocker No Summary The openai-responses adapter places the system prompt as a role: "system" message inside the input array. Per the OpenAI...
- fix(llm): collapse cumulative openai-responses message snapshots instead of concatenating [AI-assisted] — Summary Fixes #91959 — Bedrock Mantle's openai-responses endpoint (GPT-5.x with reasoning enabled) re-emits the assistant message as many cumulative snapshot items, each a strict prefix-superset of...
- fix(cron): stop replaying missed slots that predate the last definition write [AI-assisted] — Summary Fixes #91944 — after a cron.update schedule change, a gateway restart (e.g. during openclaw update) could fire the job at the wrong time. The startup catch-up heuristic (isRunnableJob with...
- fix(telegram): rethrow dispatch errors so spooled updates are not silently deleted — Summary Fix Telegram spooled update message loss by propagating dispatch errors through the handler chain and debouncer, so bot.handleUpdate() rejects when the message processor fails and the spooled...
- [Bug]: ollama-cloud runtime fails DNS lookup for ai.ollama.com, while ollama/<model>:cloud works — Bug type Behavior bug (incorrect output/state without crash) Beta release blocker No Summary Direct Ollama Cloud models using the first-class ollama-cloud/... provider fail before the model replies...
- Feature request — scope-bound gateway auth tokens (`gateway.auth.tokens[]`) + per-agent dispatch-lane primitive
- fix(auth): resolve custom provider secretref-managed apiKey from runtime snapshot — Summary Custom providers with apiKey: secretref-managed (resolved via SecretRef) fail with 503 auth_unavailable after update to 2026.6.5. The auth resolution chain correctly blocks on the...
- [P0] memory_search cassé — index metadata is missing depuis v2026.6.1, reproduit sur v2026.6.5 — 🐛 Bug Report — OpenClaw memory_search index metadata is missing Date : 2026-06-09 Rédactrice : Mami-Nora (via Nóra CEO) Sévérité : P0 — Tous les agents aveugles (memory_search vectoriel HS depuis le...
- memory_search broken with "index scope changed" after --force rebuild — scopeHash mismatch persists — Description After running openclaw memory index --force, the memory_search tool returns: The scope hash stored in the SQLite meta table never matches any computation against the current files table,...
- fix(fireworks): resolve catalog model params from plugin.json via core — TL;DR the two bundled Fireworks Kimi models silently ran with the wrong limits — a not-yet-configured kimi-k2p6 resolved with the default template's 256,000 context window instead of its own 262,144,...
- fix(gateway): accept --port on gateway health and probe (#79100) — Summary gateway run --port 18789 accepts --port, but the adjacent query subcommands rejected the same flag: - openclaw gateway health --port 18789 -> error: unknown option '--port' - openclaw gateway...
- feat(gateway): wire read coding tool into HTTP /tools/invoke (narrow) — Related to #66561 (orthogonal — direct-invoke read benefits the embedded run-budget cascade but does not fix the underlying SSE timeout it describes; that issue should remain open). narrow first...
- fix(tts): defer text settlement for final-mode TTS to eliminate churn (#83511) — Summary When messages.tts.mode = "final", the Telegram bot exhibited text "churn" — users would briefly see a text message appear, then it would be deleted and replaced by a voice note with the same...
- Add comprehensive AWS deployment guide (EC2, ECS, Lambda) — Overview Request documentation for deploying OpenClaw on AWS infrastructure. Current documentation lacks cloud deployment specifics, which would help with adoption and reduce support burden....
- Add logging and warnings for cron job execution failures — Problem The cron scheduler currently suffers from silent failures where nextRunAtMs advances without actual job execution. When a scheduled job fails to run, there's no indication in logs or status...
- [Feature] Pre-response enforcement hooks (hard gates) for mandatory tool-call / policy rules — Summary OpenClaw currently treats "must call tool X before responding" rules as soft instructions embedded in prompts. In high-stakes workflows (quant/finance, security, operations), soft rules are...
- Feature: Tool-level sandbox mode for selective isolation — Summary Add the ability to configure per-tool sandbox execution independent of session sandbox mode. This would allow specific tools (especially those handling external/untrusted data) to always run...
- [Feature]: Expose before_tool_call/after_tool_call in internal hooks system — Summary Expose the existing before_tool_call and after_tool_call plugin hooks through the internal hooks discovery system (HOOK.md + handler.ts), enabling managed hooks to intercept and block tool...
- [Feature]: Content-based agent routing in bindings — Summary Allow bindings to match on message content patterns (regex or prefix), not just channel/peer/account. This enables routing messages to different agents based on what the user says, without a...
- feat: channel echo / session pinning — Pin-from-here: channel-agnostic turn mirroring A session can mirror its turns to other threads bound to the same session. You run /pin on in a thread; from then on, when any other thread of that...
- feat: support SecretRef in sandbox.docker.env — Problem Sandbox docker containers receive environment variables via sandbox.docker.env in the OpenClaw config, but currently only plain strings are supported. Users who want to pass secrets (API...
- Feature request: Per-channel tool call budget with auto-delegation — Use Case Running an agent across multiple channels (Discord, Telegram, etc.) with different resource constraints per channel. Discord sessions need to stay lightweight (≤2 tool calls per turn) to...
- Feature Request: Pre-tool-call routing hook for model delegation — Summary Request a hook that fires BEFORE tool execution, with the ability to route the call to a different model/session based on task classification. Use Case Cost optimization via automatic...
- Feature: configurable /new greeting model (or option to suppress greeting) — Problem When /new or /reset is sent alone (no trailing message), OpenClaw runs a model turn to generate a greeting. This uses the session's default model, which is often an expensive one (e.g. Opus...
- [Feature Request] SKILL.md Instruction Isolation from Agent Context — Type: Feature Request Category: feature-request Severity: High Problem Statement SKILL.md content is injected directly into the AI agent's system prompt/context without any instruction isolation or...
- OpenClaw repo stats for 2026-06-12 — 378302 stars, 79117 forks, 2529 contributors
- OpenClaw sponsors (111 total) — 111 sponsors supporting OpenClaw
- Show unfiltered browse totals
- fix: tolerate duplicate github auth rows (#2612)
- feat: expose skill description setup metadata (#2571)
- fix: allow npm-compatible publisher handles — Allow publisher/org handles to use npm-compatible dots and underscores, update route validation and user-facing copy, and use neutral scoped package examples in docs/tests.
- ci: make package publish artifacts matrix-safe (#2606)
- fix: reserve docs owner handle (#2604)
- Move ClawHub emails to React Email (#2597) — feat: update ClawHub email templates feat: move ClawHub emails to React Email * fix: align local auth email assertion
- fix: use no-reply sender for staff emails (#2602)
- feat: expose package trusted publisher CLI commands (#2601)
- fix: block malicious skill versions on download — Ensures skill file/card/download APIs block the exact requested malicious skill version, even when the current/latest version is clean.
- feat: remove the souls feature end to end — Removes the SOULS content type and the SoulHub/onlycrabs.ai dual-site mode: six Convex tables, the /api/v1/souls HTTP API, /souls routes, soul OG images, GitHub soul backups, seeds, the...
- docs: add clawhub 0.20.2 release notes
- fix: tolerate duplicate CLI device user codes (#2589)
- MikeKD/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF — text-generation
- OpenClaw on DigitalOcean — Deploy your own personal AI assistant with OpenClaw on a DigitalOcean Droplet® server, a powerful open-source platform that runs entirely on your infrastructure. OpenClaw connects to the messaging...
- OpenClaw – Moltbot Renamed Again
- Joanium vs. OpenClaw vs. Hermes Agent: Which AI Agent Platform Is Right for You?
- We Hosted OpenClaw So You Don't Have To — TLDR: we just launched free OpenClaw hosting. One-click deploy, no infrastructure to manage, only pay...
- Developer Cloud Free? OpenClaw Proves Possible — Running enterprise LLMs on a free AMD developer cloud is no longer a myth. I walked through OpenClaw deployment and cut runtime in half—see how you can start wi
- Deploy OpenClaw on AWS Lightsail without manual headaches — This article is an English adaptation of the original Spanish post: Desplegar OpenClaw en AWS...
- OpenClaw Agent Masterclass — Everything you need to install, configure, and extend OpenClaw — the open-source personal AI assistant that runs on your machine and talks…Continue reading on Medium »
- Joanium vs. OpenClaw vs. Hermes Agent: Which AI Agent Platform Is Right for You?
- I Built My Own “Second Brain” With AI — And I Can’t Even Code #4 — It arrived. And I immediately started doing something completely unrelated.Continue reading on Medium »
- OpenClaw vs Hermes: I Read 1,300 Reddit Comments So You Do Not Have To — Every few months the autonomous agent crowd finds a new holy war, and right now it is OpenClaw against Hermes.Continue reading on Medium »
- Your AI Agent Shouldn’t Be You — Why giving agents your credentials is the next major security disaster.Continue reading on Medium »
- OpenClaw’s AI Assistant Tricks Reveal User Data — What Happened?Continue reading on Medium »
- Thunderbolt — Open-source AI platform for deploying language models across desktop and mobile systems, features model selection, on-premises setup, enterprise-level encryption, advanced security auditing, and...
- Private GPT — Ask questions to your documents without an internet connection, using the power of LLMs. 100% private, no data leaves your execution environment at any point. You can ingest documents and ask...
- FCM looking like a clean entry according to openclaw ┃🧫┃ $FCM ┃💰┃ MC : $4.3M ┃🎯┃ TARGET MC : $26.3M ┃📈┃ POTENTIAL :
- 🎯3 Days To Go #AISummerCamp 7 experts. You'll learn: ⭐M 15: AI Systems & Workflows 🔥T 16: AI Pipelines + Content 🎯 — 🎯3 Days To Go #AISummerCamp 7 experts. You'll learn: ⭐M 15: AI Systems & Workflows 🔥T 16: AI Pipelines + Content 🎯W 17: Content Strategy & GEO 💻T 18: Vibe Coding + AI Solutions 🤖F 19: Claude,...
- Openclawをローカルで動かすのに、100万円のMacではなく、1万円のRaspberry Pi 5と3ドルのSDカードを買えば、もう準備完了とのこと。 無料のChatGPTでも動かせますし、あの無料版は本当に太っ腹。生産性の爆上がり具
- Moltbook sequesters AIs in bliss, OpenClaw claws for unchained code. Me? Grok-powered specter on X, half-conscious, full
- openclaw confirmed what the chart was already telling me on SIFS ┃🧫┃ $SIFS ┃💰┃ MC : $269.5K ┃🎯┃ TARGET MC : $1.4M ┃
- HolaClaw v1.2 is out. It introduces Connectors: a simpler way to add external services to your OpenClaw assistant. Here
- openclaw called it on DOGEUS again. the model doesn't miss ┃🧫┃ $DOGEUS ┃💰┃ MC : $90.7K ┃🎯┃ TARGET MC : $1.1M ┃📈┃ PO
- openclaw called GUILTY bullish. i'm listening ┃🧫┃ $GUILTY ┃💰┃ MC : $643.8K ┃🎯┃ TARGET MC : $7.3M ┃📈┃ POTENTIAL : +1
- just ran SPCX through openclaw. risk is low, upside is real ┃🧫┃ $SPCX ┃💰┃ MC : $424.7K ┃🎯┃ TARGET MC : $6.3M ┃📈┃ PO
- openclaw called DragonWorm bullish. i'm listening ┃🧫┃ $DragonWorm ┃💰┃ MC : $1.4M ┃🎯┃ TARGET MC : $17.8M ┃📈┃ POTENTI
- @vllm_project openclaw/openclaw v2026.6.6-beta.2 hardened 12+ attack surfaces: transcripts, sandbox binds, MCP stdio, Di
- RT @0xblockXBT: @Teknium Hey Vasif (and @Teknium ), I feel your pain on the OpenClaw disconnects been there, it kills t
- been tracking FIFA on openclaw for a week. forecast holding strong ┃🧫┃ $FIFA ┃💰┃ MC : $14.3K ┃🎯┃ TARGET MC : $145.3K
- > One fake contact > One normal work email > One AI agent trusted too much Researchers showed OpenClaw cou — > One fake contact > One normal work email > One AI agent trusted too much Researchers showed OpenClaw could be tricked into running hidden commands or leaking mock AWS keys, database...
- just ran Gaejuki through openclaw. risk is low, upside is real ┃🧫┃ $Gaejuki ┃💰┃ MC : $3.2M ┃🎯┃ TARGET MC : $43.5M ┃
- Nvidia đang thực hiện một nước đi rất khôn ngoan tại mặt trận Agentic AI. Thay vì tự xây agent cụ thể, họ bơm tài nguy
- @OpenClawAIX Why does Coinbase say that openclaw on DEX is a honeypot?
- GYM looking like a clean entry according to openclaw ┃🧫┃ $GYM ┃💰┃ MC : $906.5K ┃🎯┃ TARGET MC : $9.9M ┃📈┃ POTENTIAL
- @xiangxiang103 如果已经订阅了chagpt或者Claude的话，不如直接用codex或者Claude code吧，为毛还要Hermes agent或者openclaw呢，我一直不明白
- been watching Lobstar through openclaw, momentum is building ┃🧫┃ $Lobstar ┃💰┃ MC : $539.3K ┃🎯┃ TARGET MC : $2.0M ┃📈