Show HN: Assign tasks to 7 AI agents with -mentions, autonomous mode, OpenClaw

[bahaAbunojaim]

I posted Mysti here a couple months ago and got a lot of feedback that shaped where I took it. Quick recap: it's a VS Code extension that lets you use multiple AI coding agents through one interface, including having them collaborate on problems.

Three things I want to highlight in this release.

@-mentions for task delegation. You can now assign work to specific agents inline. Something like: @claude write the API handlers then @gemini review them for security issues then @claude fix what gemini found. The agents execute sequentially and each one gets the output of the previous agent as context. It's basically a pipeline — you describe the chain of work in natural language and Mysti figures out the task graph. If a sub-agent fails, it auto-retries once and keeps going. I've been using this a lot for the pattern of "write with one model, review with another" which catches things a single agent misses.

Autonomous and semi-autonomous modes. This is the one I'm most excited about. You can set a goal — "refactor the auth module to use JWT" — and the AI just works. It auto-approves safe operations (file reads, file creates, running tests) and only interrupts you for genuinely uncertain decisions. There's a safety classifier with three levels: conservative (read-only auto-approved), balanced (normal dev operations auto-approved), and aggressive (everything except hardcoded dangerous stuff like rm -rf and git push --force). The system learns your preferences over time — if you keep approving TypeScript file creation, it stops asking. Semi-autonomous is the gentler version: you stay in the loop, but if you don't respond within 30 seconds, it makes the call for you based on what it's learned. I use balanced mode daily now. The number of permission dialogs during a big refactor went from dozens to maybe two or three.

OpenClaw integration. This one's worth calling out specifically. OpenClaw runs a local daemon with a WebSocket gateway — so instead of spawning a CLI process for every message, Mysti maintains a persistent connection with real-time streaming. OpenClaw also has an active mode where it connects to messaging channels — WhatsApp, Telegram, Slack, Discord — and you can manage all of it from VSCode. I think this is going to be a big deal.

Mysti now supports 7 providers total: Claude Code, Codex, Gemini, Copilot, Cline, Cursor, and OpenClaw. All of them work with brainstorm mode (pick any 2 to collaborate), all of them support the @-mention routing, and each tab runs independently so you can have different agents working in parallel across panels.

Apache 2.0 licensed. Uses your existing CLI installations — no middleman, no proxy.

Website: https://deepmyst.com/Mysti GitHub: https://github.com/DeepMyst/Mysti VS Code Marketplace: https://marketplace.visualstudio.com/items?itemName=DeepMyst...

[clemenshelm]

Multi-agent orchestration is where permissions become critical. Who can use which agent? What can each agent access? I'm building Pinchy (https://heypinchy.com) for exactly this: enterprise governance layer on OpenClaw with RBAC and audit trails. AGPL, self-hosted.

[bahaAbunojaim]

Happy to discuss further