This is our biggest release window yet — 10 releases (mc.5 through mc.14) shipping the largest AI capability expansion in MoltClaw's history alongside meaningful polish across voice, attachments, models, and reliability.

🌟 Top Highlights

1. Your AI assistant can now do ~37× more. We went from 13 HighLevel skills in mc.11 to ~483 skills live today. Ask MoltClaw to find contacts, search conversations, list opportunities, create calendar appointments, send messages, manage forms, work with social posts, handle invoicing, set up payments, manage knowledge bases, and ~470 more — all directly in chat.

2. More AI models, including Anthropic's newest flagship. Claude Opus 4.7, Opus 4.6 with 1M context, Sonnet 4.6, plus continued OpenAI / Gemini / Kimi options — pick the right brain for the job from the model picker.

3. Voice input shipped. Click the mic button in the composer and speak instead of typing. Works in Chrome and other modern browsers.

4. Images and attachments work everywhere. Send images to GPT, Claude, and Gemini; see them rendered cleanly in chat history; drop large files without hitting upload errors.

5. Three coordinated releases (mc.12, mc.13, mc.14) ship the AI skill expansion in safe waves with rollback protection — zero customer-facing downtime through the rollout.

🚀 AI Skill Expansion — Ask MoltClaw to Just Do It

Before mc.11 your assistant could chat about your CRM. Now it can actually operate it. ~483 first-class HighLevel skills are wired in across 24 domains:

How it feels in practice

"Show me all open opportunities over $5K from the last 30 days that haven't moved stage in 7+ days." "Create a calendar appointment with Sarah next Tuesday at 2pm." "Send a follow-up message to every contact tagged 'demo-scheduled' who hasn't replied." "Upload these custom field values to the contact form."

Write actions (creates/updates/deletes) always show you a confirmation card before executing — your assistant proposes the change; you approve or cancel. No silent writes to your CRM.

🧠 New AI Models

You can now pick from a richer model lineup in the composer's settings modal:

Anthropic

• Claude Opus 4.7 (NEW) — Anthropic's newest flagship. Stronger reasoning and tool use than 4.6. Available with both 200K and 1M context.

• Claude Opus 4.6 (1M context now available) — same model, longer memory.

• Claude Sonnet 4.6 — fast and capable, great default for everyday work.

• Claude Sonnet 4.5 — slightly leaner, lower latency.

OpenAI

• GPT-5.4 (default) — best overall performance and quality.

• GPT-5.2 — strong reasoning at lower cost.

• GPT-5 Mini — fast and affordable for daily tasks.

• GPT-4.1 — quick simple tasks on a budget.

Google Gemini

• Gemini 2.5 Pro & Flash.

Kimi

• Kimi 1.5 & Kimi Latest.

The model picker now groups by provider with concise "best for" labels so you can pick the right model in seconds. (GPT 5.5 and Kimi 2.6 are coming in the next release.)

🎙️ Voice Input

Click the microphone button next to the composer and start speaking — your speech transcribes directly into the message box.

• Works on Google Chrome (recommended) and other Chromium-based browsers.

• Microphone permissions are requested cleanly the first time.

• Clear error messages if your browser doesn't support it or if mic access is denied.

Heads up for Brave / Firefox users: voice input depends on the browser's Web Speech API, which Brave and Firefox restrict by default. Use Chrome for the best experience. (Browser-specific guidance in the error message is shipping in the next release.)

🖼️ Images & Attachments

A run of releases (mc.6 through mc.10) made image / attachment handling work end-to-end across every model and surface:

• Send images to any provider: Gemini, OpenAI, and Anthropic all accept images now. The right format is sent to each provider automatically.

• AskClaw composer renders attachments with previews — you can see what you're about to send before hitting send.

• Image rendering in chat history — assistant responses that include images now display them inline instead of as broken links.

• Large file uploads no longer fail with 413 errors — uploads up to the platform limit work reliably.

• Web composer thumbnails — preview thumbnails for image attachments in the composer.

• Empty-turn UX cleanup — when an attachment fails or is sanitized, the assistant no longer hangs or returns confusing empty turns.

If you've ever tried to drop a screenshot or document into a conversation and hit a wall, this release window is for you.

💬 Composer Polish

Smaller but meaningful improvements:

• Model picker dropdown is now alphabetically organized within each provider group, with the most-used models surfaced first.

• Empty-turn handling — the assistant gracefully handles edge cases where a previous message was empty or sanitized.

• Settings modal reorganization — model, reasoning effort, and other controls are easier to find.

• Thinking indicators are clearer about when the model is reasoning vs. responding.

🛡️ Reliability & Performance

A lot of the work since mc.4 is invisible-but-important infrastructure:

• Three-wave release strategy — the AI skill expansion shipped in mc.12 (reads), mc.13 (writes), and mc.14 (cleanup + telemetry) so any single release stays scoped and rollback-able.

• Automatic deprecation handshake — older versions of skills now coexist with newer versions during transition. Your existing automations don't break when we ship new skills; instead they emit telemetry that helps us know when it's safe to retire the older versions (in mc.15+).

• Hardened startup behavior — eliminated a class of intermittent startup hangs during high-traffic deploys.

• Improved background scheduler — better handling of resource cleanup so long-running locations stay healthy.

• Live deprecation telemetry — operators get visibility into which legacy skills are still in use.

• Real rollback protection — mc.12 hit a configuration anomaly during rollout, automatically rolled back to mc.11, and re-promoted cleanly within the same day. Your service stayed up.

🔐 Security & Safety

• Write actions always require user confirmation. The proposal/confirm pattern is enforced at the platform level — your assistant cannot silently mutate your CRM.

• Three high-risk write skills were quarantined during pre-release security review (user creation, payment provider configuration) — these need stronger UX guardrails than today's confirmation cards provide and will return after a redesign.

• Permission rules tightened for an internal proxy gap that affected one less-used domain.

⚙️ Behind the Scenes (For the Curious)

Numbers from the release window:

• Skill files shipped: ~445 new (across mc.11 through mc.14)

• Backend routes added: 209 read endpoints + 286 write action handlers

• Test coverage: 456 per-skill harness fixtures running on staging canaries before every release

• Rollback triggers fired and successfully recovered: 1 (mc.12 environmental flap)

• Customer-impacting incidents: 0

🔮 What's Next (Phase 28 in Planning)

Things you'll see in upcoming releases:

• GPT-5.5 and Kimi 2.6 added to the model picker

• Browser-aware voice input messaging — Brave / Firefox users get specific guidance instead of generic errors

• Live model picker — when we add a new model, it shows up in your picker without waiting for a separate UI update

• "Open MoltClaw Dashboard" quick-link from the model picker to manage advanced configuration

Questions or feedback?

If something works (or doesn't) the way you expected, please share — every release in this window had at least one improvement that came directly from a customer report. The voice input, attachment fixes, model additions, and many of the new skills were prioritized because someone asked for them.

Thanks for using MoltClaw. 🦊

Released: April 2026 · Build 2026.4.12-mc.14 · Live on all staging locations

What's New in MoltClaw — April 17, 2026

TL;DR

Talk to MoltClaw with your voice. Drop in screenshots and have it actually look at them. And meet Claude Opus 4.7 — including a version that can read a million tokens in one go.

🎤 The mic button works now

Click the mic in the chat composer, speak naturally, and watch your words appear in the box. No more silent failures or "did I click the right thing?" moments. If something does go wrong (mic permission denied, no internet, unsupported browser), you'll see a clear message instead of nothing happening.

Why it matters: Voice is faster than typing for long instructions, brainstorming, or hands-busy moments. We made it actually work.

🖼️ Drop in an image and MoltClaw will see it

Attach a screenshot, photo, diagram, or PDF page to any message and MoltClaw will actually read it. Previously the file would attach in your composer but quietly never reach the model — so you'd get an "I don't see any image" reply. That's fixed across every model in the picker (Claude, GPT, Gemini, Kimi).

Why it matters:

• Paste a screenshot of an error and ask "what's broken?"

• Drop in a competitor's landing page and ask "what would you change?"

• Share a contact's business card image and have it pulled into your CRM

• Show a flow chart and ask MoltClaw to turn it into a workflow

🧠 Meet Claude Opus 4.7 — and 1M-token context windows

The model picker now includes four new options under LeadConnector Anthropic:

ModelWhat it's forClaude Opus 4.7Anthropic's newest top-tier model. Better reasoning, sharper writing, snappier on tools. Use this as your everyday "smart mode."Claude Opus 4.7 (1M)Same model, but with a 1 million token context window. That's roughly 1,500 pages of text in a single conversation.Claude Opus 4.6Still here for anyone who's tuned prompts to it.Claude Opus 4.6 (1M)The 1M-context version of 4.6, now selectable as a separate option.

Why the 1M variants matter: You can now feed MoltClaw an entire knowledge base, a year of CRM notes, a full transcript library, or a long product spec — all in one conversation, with nothing summarized away. No more "the model lost context halfway through." It just remembers everything.

Heads up on cost: 1M-context calls use a lot more tokens, so save them for the moments when you actually need to reason across a huge document. For everyday chat, the standard 200K version is faster and cheaper.

How to try it

1. Open chat

2. Click the model dropdown → pick a new option

3. Click the 🎤 to test voice

4. Drag an image into the composer and ask MoltClaw about it

That's it — no setup, no settings, no migration. Everything is live now.

A small note on quality

We shipped these three fixes together, with each one independently reviewed by four agents (developer, QA, engineering director, and product) before it could land. We're holding ourselves to that bar so the next "feature is visible but doesn't actually work" surprise stays out of your day.

If you spot anything off, hit reply — we read every report.

— The MoltClaw Team

🧠 Memory & Context

• Active Memory plugin (2026.4.10 → 2026.4.12) — A dedicated memory sub-agent runs right before the main reply, automatically pulling in relevant preferences, past context, and user details. No more having to manually say "remember this" or "/search memory." Configurable message/recent/full context modes, live /verbose inspection, and opt-in transcript persistence.

• Dreaming system overhaul (2026.4.5) — Three cooperative phases (light, deep, REM), weighted short-term recall, a new /dreaming command, a Dreams UI, and a grounded REM backfill lane that replays old daily notes into durable memory (2026.4.9).

• Memory aging controls (2026.4.5) — Configurable recencyHalfLifeDays and maxAgeDays with verbose logging.

🤖 Model Providers & AI

• Bundled Codex provider (2026.4.10 → 2026.4.12) — codex/gpt-* models now use Codex-managed auth, native threads, model discovery, and compaction. openai/gpt-* stays on the regular OpenAI path.

• Bundled LM Studio provider (2026.4.12) — Onboarding, runtime model discovery, stream preload, and memory-search embeddings for local/self-hosted OpenAI-compatible models.

• New providers added (2026.4.5) — Qwen, Fireworks AI, StepFun, MiniMax, Ollama Web Search.

• Amazon Bedrock enhancements (2026.4.5) — Mantle support with automatic inference-profile discovery and region injection.

• Provider auth aliases (2026.4.9) — Variants can share env vars, auth profiles, and API-key onboarding without core-specific integration work.

🎨 Media Generation

• Video generation tool (2026.4.5) — Built-in video_generate with xAI Grok, Alibaba Wan, and Runway as providers.

• Music generation tool (2026.4.5) — New music_generate with Google Lyria, MiniMax, and ComfyUI workflow support.

• Seedance 2.0 integration (2026.4.10) — Video generation with duration/resolution/audio/seed metadata.

• ComfyUI plugin (2026.4.5) — Bundled local + Comfy Cloud workflow support for image, video, and music.

🎙️ Voice & Speech

• Local MLX speech provider for Talk Mode (2026.4.12, macOS) — Experimental local speech provider with interruption handling and local utterance playback.

💬 Channels & Integrations

• Android Assistant integration (2026.4.2) — App Actions metadata so OpenClaw launches from assistant triggers.

• Feishu Drive comment-event flow (2026.4.2) — In-thread replies and collaboration workflows.

• Microsoft Teams message actions (2026.4.10) — Pin, unpin, read, react, and reaction listing.

• Matrix live markers (MSC4357) (2026.4.10) — Live/typewriter animation during draft sends and edits.

• Matrix exec approvals (2026.4.5) — Native approval prompts with account-scoped approvers.

• Matrix spec-compliant mentions (2026.4.2) — m.mentions metadata across text, media, edits, and polls.

• iOS/APNs approval notifications (2026.4.5) — In-app modal interfaces.

• Telegram forum topics (2026.4.14, partially in base) — Human topic names now surface in agent context and prompt metadata.

🧩 Agents, Plugins & Skills

• ClawHub integration (2026.4.5) — Direct skill search, detail views, and installation flows in the Skills panel.

• Task Flow substrate restored (2026.4.2) — Managed-vs-mirrored sync modes, durable flow state, managed child task spawning with sticky cancel intent.

• Claude CLI integration (2026.4.5) — Loopback MCP bridge, stdin + streaming for live progress, session metadata preservation.

• Plugin config TUI (2026.4.5) — Guided onboarding flows and a --force install flag.

• Narrower plugin loading (2026.4.12) — CLI, provider, and channel activation narrowed to manifest-declared needs.

🖥️ Control UI & UX

• Control UI localization (2026.4.5) — Added Simplified/Traditional Chinese, Portuguese, German, Spanish, Japanese, Korean, French, Turkish, Indonesian, Polish, Ukrainian.

• Per-session thinking picker (2026.4.5) — In chat header and mobile settings.

• Structured diary view (2026.4.9) — Timeline navigation, backfill/reset controls, traceable dreaming summaries.

• CLI exec policy command (2026.4.12) — show, preset, set subcommands for managing tool execution approvals.

🔒 Security Hardening

• Plugin allowlist enforcement (2026.4.5) — Stricter tool access controls, owner-only modifications.

• Device pairing hardening (2026.4.5) — Non-admin sessions restricted to managing their own devices.

• Exec approval rigor (2026.4.5) — Fail-closed behavior, strict inline-eval rules.

• SSRF quarantine checks (2026.4.9) — Blocked unsafe env vars from untrusted workspace files, sanitized remote node execution event outputs.

• Browser/sandbox navigation defenses (2026.4.10) — Exec preflight validation and outbound host-media access controls.

🧪 QA & Testing

• QA multipass runner for Linux VMs (2026.4.12) — Disposable test environments.

• Matrix homeserver integration (2026.4.10) — QA lane.

• Telegram bot-to-bot verification (2026.4.10) — QA lane.

• Character-vibes evaluation reports (2026.4.9) — Model selection and parallel runs.

Customer-framing recommendation: The big three worth leading with in a customer announcement are:

1. Active Memory (automatic context recall — the most visible day-to-day change)

2. Massive model/provider expansion (Codex, LM Studio, Qwen, Fireworks, StepFun, MiniMax, Ollama Web Search, Bedrock Mantle)

3. Media generation (built-in video_generate and music_generate tools + ComfyUI)

Sources:

• OpenClaw releases (GitHub)

• Release v2026.4.12

• Release v2026.4.10

• Release v2026.4.9

• Release v2026.4.8

• Release v2026.4.5

• Release v2026.4.2

• Openclaw Release Notes — Releasebot