Hermes v2026.5.7 Update: Hermes Agent v0.13.0 (2026.5.7) — The Tenacity Release

This release is all about making Hermes Agent finish what it starts — with durable multi-agent teamwork, laser-focused goal tracking, and a big security cleanup.

What’s New

Multi-agent Kanban — your AI team actually completes tasks now.
Think of it like a real project board (Trello, but for AI agents). You create a board, drop tasks on it, and multiple Hermes workers pick them up, pass them along, and close them out. The magic is in the reliability: agents send heartbeats so the system knows they’re alive, tasks get reclaimed if a worker stalls, zombies get detected and cleaned up, and each task has a retry budget. There’s even a “hallucination gate” that catches when an agent thinks it’s done but really isn’t. It’s like having a project manager who never sleeps.

/goal — the agent doesn’t forget what you asked.
Ever had an AI assistant drift off topic mid-task? /goal locks the agent onto your target across multiple turns. It’s a built-in “Ralph loop” (named after the dog who never lets go of the stick) — the agent keeps coming back to the goal no matter how many side quests it encounters.

Show it a video.
New video_analyze tool lets Gemini and compatible models actually watch video content and answer questions about it. No more screenshots or frame-by-frame hacks.

Clone a voice.
xAI Custom Voices are now a TTS provider, meaning you can clone a voice and have Hermes speak in it. Perfect for content creators or accessibility setups.

Hermes speaks your language.
Static gateway and CLI messages now translate to 7 locales: Chinese, Japanese, German, Spanish, French, Ukrainian, and Turkish. The docs site also has a Chinese (zh-Hans) version.

Improvements

Checkpoints v2 — state persistence got a major rewrite with real pruning. Old checkpoints don’t pile up like digital junk; the system cleans up after itself.
Gateway auto-resume — if the server restarts, interrupted sessions pick up right where they left off. No more lost conversations.
Cron watchdog mode — scheduled tasks now have a no_agent mode that just monitors and alerts, without needing an agent to run.
Google Chat is now the 20th platform — Hermes keeps expanding its reach.
Providers are pluggable — swapping AI backends is now a cleaner, more modular process.

Fixes & Security

This release closes 8 P0 security issues — that’s serious. Highlights:

Redaction is ON by default — sensitive data gets masked automatically.
Discord role-allowlists are guild-scoped — permissions don’t leak across servers.
WhatsApp rejects strangers by default — no more random people messaging your bot.
TOCTOU windows closed — race-condition gaps in auth.json and MCP OAuth are patched.

Plus 282 issues closed (13 P0, 36 P1) with 295 community contributors. That’s a lot of eyeballs making the code safer.

What This Means for You

If you’ve ever felt like AI agents are flaky — they start tasks but don’t finish, or they go off the rails mid-conversation — this release is the antidote. The Kanban board alone is a game-changer for anyone running multi-step workflows. And the security fixes mean you can trust Hermes with sensitive data more than before.

It’s called “Tenacity” for a reason: Hermes Agent now has the persistence to see things through, the memory to stay on target, and the safety rails to keep you out of trouble.

📎 Official release notes: https://github.com/NousResearch/hermes-agent/releases