Hermes v2026.3.28 Update: Hermes Agent v0.5.0 (v2026.3.28)

This release is a major hardening and expansion update that brings over 400 new AI models, better plugin support, and a security overhaul — think of it as giving your AI assistant a bigger toolbox, a cleaner workshop, and a locked front door.

🆕 What’s New

More Models, One Door
Previously, you had to juggle multiple providers to try different AI models. Now, Hermes Agent connects directly to Nous Portal (400+ models) and Hugging Face as first-class providers. The setup wizard walks you through configuration, and a live /models command lets you see what’s available without leaving your chat. It’s like going from a single grocery store to a whole farmers’ market — same basket, way more produce.

Telegram Private Chat Topics
If you use Telegram, you can now create project-based topics within a single chat. Each topic can have its own set of skills (tools) bound to it. So you can have a “Work” topic that can access code tools, and a “Personal” topic that can’t — keeping workflows isolated and organized. It’s like having separate folders on your desk instead of one messy pile.

Native Modal SDK Backend
The old Modal integration relied on a third-party wrapper (swe-rex) that felt like using a walkie-talkie to call your neighbor. Now, Hermes Agent talks directly to Modal’s SDK, which means faster, more reliable cloud sandboxes — no more tunnel headaches.

Plugin Lifecycle Hooks (Finally!)
Plugins can now hook into the agent’s core loop: before/after every AI call, and when a session starts or ends. This lets developers add custom logging, caching, or even alter prompts mid-conversation. It’s like giving plugins a seat at the control panel, not just a peephole.

🔧 Improvements

GPT Models Now Actually Use Tools
OpenAI’s GPT models sometimes “described” what they wanted to do instead of just doing it. Hermes now includes a guidance prompt that pushes them to make real tool calls. We also strip stale budget warnings from conversation history — these were confusing the model into avoiding tools across turns. Think of it as clearing the old sticky notes off your monitor so you can see the task clearly.

Anthropic Output Limits Fixed
Previously, Anthropic models were capped at 16,000 tokens per response, even if they could handle more. That limit is now per-model, so longer outputs won’t get cut off mid-sentence.

Nix Flake Support
For the Linux/Nix crowd, Hermes now ships as a proper Nix flake with a NixOS module. It even auto-generates config keys from Python source, so you don’t have to guess the right setting names. It’s like getting a pre-assembled IKEA desk instead of a bag of screws.

🛡️ Fixes & Security

This release is a hardening milestone. We removed the compromised litellm dependency, pinned every dependency version, regenerated lock files with hashes, and added CI scans for supply chain attacks. We also patched several CVEs. Bottom line: if you were worried about using open-source AI infrastructure, this update tightens the bolts significantly.

🎯 What This Means for You

• More choice — try niche or frontier models without changing your setup.
• Cleaner workflows — keep projects separate in Telegram, and let plugins do more.
• Fewer surprises — GPT models behave, Anthropic outputs don’t truncate, and your supply chain is cleaner.

Whether you’re a developer, a power user, or just someone who likes their AI to work without babysitting, this release makes Hermes Agent more capable, more reliable, and safer to use.

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