Hermes v2026.3.23 Update: Hermes Agent v0.4.0 (v2026.3.23)

This release turns Hermes Agent into a full-fledged AI platform — think of it as adding a universal remote control, a bigger toolbox, and a better memory to your existing AI assistant.

🎉 What’s New

1. Talk to Hermes from Anywhere (OpenAI-Compatible API)

Previously, Hermes was like a great chef who only cooked in their own kitchen. Now, we’ve added a standard takeout window: any app that speaks OpenAI’s language (like many existing AI tools) can call Hermes directly via /v1/chat/completions. Plus, there’s a new REST API for scheduling cron jobs — perfect for automating daily reports or reminders. We’ve also added safety guards like input limits and CORS protection so you don’t get unexpected guests.

2. Six New Messaging Adapters

Hermes now speaks Signal, DingTalk, SMS (via Twilio), Mattermost, Matrix, and Webhooks — on top of the existing Telegram, Discord, and WhatsApp. Think of it as adding six new phone lines to your office. And if one line drops, the gateway now auto-reconnects with a smart “wait and retry” strategy (exponential backoff), so you won’t lose messages during hiccups.

3. @ Context References (Like Mentioning a File in a Chat)

You can now type @file or @url in the CLI to pull in a document or webpage as context — with tab completion. It’s like saying “read this before answering” without copy-pasting. Huge time-saver for debugging or summarizing long docs.

4. Four New AI Providers

Hermes now works with GitHub Copilot, Alibaba Cloud (DashScope), Kilo Code, and OpenCode Zen/Go. If you already pay for one of these, you can plug it in as the brain behind Hermes. No need to switch ecosystems.

5. MCP Server Management (with OAuth 2.1)

The new hermes mcp command lets you install, configure, and authenticate MCP servers — including full OAuth 2.1 PKCE flow. Think of it as a “plug-and-play” hub for external tools, but with proper security keys instead of just duct tape.

6. Gateway Prompt Caching = Lower Costs

For long conversations, Hermes now caches the AI agent per session, preserving Anthropic’s prompt cache. This is like keeping a book open on the right page instead of re-reading the whole chapter each turn — dramatically cheaper for ongoing chats.

🔧 Improvements

Context compression overhaul: Long chats now get structured summaries with “tail protection” (so the last few messages aren’t lost). You can even set a custom summary endpoint and a fallback model if the main one fails.
Streaming by default: Responses now stream token-by-token out of the box — feels faster and more natural, like watching a typist rather than waiting for a printer.

🐛 Fixes & Reliability

This release includes 200+ bug fixes — a “spring cleaning” pass. We fixed memory leaks, race conditions in the gateway, edge cases in adapter reconnection, and SQLite response persistence issues. If you’ve ever had a message vanish or a connection hang, there’s a good chance it’s fixed here.

🎯 What This Means for You

If you’re a developer: You can now expose Hermes as a drop-in replacement for OpenAI-compatible apps, manage MCP servers from the CLI, and save money on long-running agents.
If you’re a power user: You get more ways to connect (Signal, SMS), faster responses (streaming), and smarter context handling (@ refs, compression).
If you’re a team: The new adapters and auto-reconnect make Hermes a more reliable hub for multi-platform workflows.

This is our biggest “platform” release yet — we’ve moved from “a great agent” to “a great agent that plays well with everything.”

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