The first time I noticed Hermes Agent, it was already too late to be early.
It was a Thursday — March 12, 2026. No keynote, no countdown, no launch thread on X. Nous Research just pushed a git tag, flipped the GitHub repository to public, and left a single message in their Discord: "this exists now." By Friday morning the repo had over a thousand stars and people in twenty time zones were arguing about which install path to use.
In the two weeks before that tag, sixty-three contributors who had never met in person had pushed 216 pull requests into the project, closed 119 issues, and grown the test suite from almost nothing to 3,289 tests. None of them worked at Nous Research. They argued through GitHub comments and on day fourteen, they shipped something that actually worked.
So what was actually in the box?
One process, seven front doors
The headline of v0.2.0 is the Multi-Platform Messaging Gateway. A single Hermes process listens on Telegram, Discord, Slack, WhatsApp, Signal, IMAP/SMTP email, and Home Assistant — at the same time. Same session manager, same memory, same tool registry. Per-platform you can configure which skills are available and how attachments are handled, but the agent on the other end is one agent.
This is more interesting than it sounds, because the standard alternative — seven separate bots, each with their own state — is terrible. Memory diverges. Tools drift out of sync. Hermes makes the gateway itself the integration point and lets the agent stay singular. You install it on a $5 VPS and reach it from whichever app you happen to be in.
Native MCP, not bolted on
Right next to the gateway is a full Model Context Protocol client. Both stdio and HTTP transports. Reconnection, resource and prompt discovery, server-initiated sampling. For readers who are not deep in agent-land: MCP is the open standard Anthropic published for letting LLMs talk to external tools in a consistent way. Most frameworks bolted MCP on later as an adapter over an older tool-calling system. Hermes wired it straight into the core on day one — every tool that speaks MCP works without a wrapper.
Skills as first-class units
v0.2.0 ships with seventy-plus bundled skills across fifteen categories, backed by what the project calls a Skills Hub: conditional activation (skills only load when prerequisites are present), prerequisite validation, and community discovery. The Hub later becomes agentskills.io. Day-one skills include image analysis, sandboxed Python execution, file search, web fetch, and a couple dozen others.
The technical decision here is that skills are declarative units with manifests, dependencies, and activation conditions — not Python functions registered at import time. That is why an agent can carry seventy of them at once without melting the prompt.
The provider router and the safety net
Two more architectural calls in v0.2.0 shape everything that comes after.
The first is a centralized provider router. A single call_llm() / async_call_llm() API replaces scattered provider logic across vision, summarization, compression, and trajectory saving. Every auxiliary consumer goes through one code path with automatic credential resolution. It sounds boring until you try to swap providers — at which point you change one file, not eleven.
The second is the safety pair: git worktree isolation (hermes -w launches each session inside an isolated worktree, so the agent cannot touch your real code by accident) and filesystem checkpoints with rollback (snapshots before destructive operations, undone with /rollback). The agent is allowed to be bold because you can actually take it back. This is the difference between "AI assistant that is careful" and "AI assistant that is brave because the system is careful for it."
And the editor side
One last thing that gets buried in the release notes but matters: ACP server support. Through the Agent Communication Protocol, Hermes integrates natively with VS Code, Zed, and JetBrains. It stops being "a terminal thing" and starts living inside the editor you actually use.
---
I keep coming back to that Thursday in March. There was no announcement, no deck, no investor call — just a git tag, a public flip, and sixty-three people who happened to be inside when the doors opened. If the rest of this blog has a thesis, it is that the speed of the people building Hermes turned out to matter more than the speed of any individual feature.