Hermes Agent: The Self-Improving AI Agent, Explained
Hermes Agent is an open-source AI agent from Nous Research that runs on your own machine or server, talks to you through Telegram, Discord, Slack, WhatsApp, Signal, email, or a terminal, and gets measurably better at your work the longer you use it. It’s MIT-licensed, has passed 218,000 stars and 41,000 forks on GitHub, and ships releases at a pace most funded startups don’t match — v0.19.0 landed on July 20, 2026 with roughly 2,245 commits and 450+ community contributors since the previous version.
The pitch that separates it from every other agent: a closed learning loop. Most agents start each task cold. Hermes creates skills out of experience, refines them while using them, nudges itself to write things down, searches its own past conversations, and builds a model of who you are across sessions.
What the learning loop actually does
Four mechanisms, and they compound:
- Autonomous skill creation. After it finishes something complicated, Hermes writes the approach down as a reusable skill. Next time, it doesn’t re-derive the method. Skills keep improving as they get used.
- Agent-curated memory. Periodic self-nudges to persist what matters, rather than dumping every message into a vector store and hoping retrieval works.
- Cross-session recall. Full-text search over its own history with model summarization, so “what did we decide about the Q3 vendor thing” resolves against real prior conversations.
- User modeling. It runs Honcho dialectic modeling to build a picture of your preferences over time, and it’s compatible with the agentskills.io open standard, so skills aren’t locked to Hermes.
This is why Hermes is worth evaluating over more than one session. On a single demo task, it looks like every other agent. The difference shows up on the fifth run of something you do every week.
What it does day to day
- Lives in your messaging apps. One gateway process serves Telegram, Discord, Slack, WhatsApp, Signal, email, and the CLI, with voice memo transcription and conversation continuity across platforms.
- Runs unattended. A built-in cron scheduler takes jobs in natural language — nightly backups, weekly audits, daily reports — and delivers results to whichever platform you want.
- Delegates to itself. It spawns isolated subagents for parallel workstreams, and can write Python that calls its own tools over RPC to collapse long pipelines into a single turn.
- Runs almost anywhere. Six terminal backends — local, Docker, SSH, Singularity, Modal, and Daytona. Modal and Daytona hibernate when idle, so a persistent agent costs nearly nothing between sessions. A $5 VPS is a legitimate deployment target.
- Uses any model. Nous Portal, OpenRouter, OpenAI, or your own endpoint, switched with
hermes model. No code changes and no lock-in — the model layer is genuinely swappable.
Installing it
On Linux, macOS, WSL2, or Termux:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashOn native Windows, in PowerShell:
iex (irm https://hermes-agent.nousresearch.com/install.ps1)The installer pulls its own dependencies — uv, Python 3.11, Node.js, ripgrep, ffmpeg, and on Windows a portable MinGit that stays isolated from any system Git install. Then:
hermes # interactive CLI
hermes setup # full setup wizard
hermes model # pick provider and model
hermes gateway # start the messaging gateway
hermes doctor # diagnose problemsIf you’re coming from OpenClaw, there’s a first-class migration path: hermes claw migrate. That’s a deliberate move — the two projects compete for the same self-hosting audience, and it’s worth reading our OpenClaw comparison alongside this.
To avoid assembling five separate API keys for model, web search, image generation, TTS, and a cloud browser, Nous Portal covers them under one subscription with 300+ models and a tool gateway, wired up by hermes setup --portal. You can still bring your own key per tool — the gateway is per-backend, not all-or-nothing.
What running it actually takes
This is the part worth being clear-eyed about before you commit a weekend.
You are the operator. You supply and rotate model API keys, choose among six backends, keep the host patched, and own every failure. The project’s velocity is a genuine strength and also a maintenance surface — that release cadence means frequent updates, and the issue tracker runs deep because a repository moving this fast accumulates them faster than they close.
There are rough edges that are nobody’s fault and still your problem. Windows Defender and Bitdefender routinely quarantine the bundled uv.exe as a false positive, because ML antivirus engines flag unsigned Rust binaries that install packages; the fix is whitelisting the folder, not the file hash, since Hermes updates uv and the hash changes each version. Termux needs a curated install because the full dependency set pulls Android-incompatible voice libraries.
None of this is disqualifying. It’s simply the actual job: Hermes gives you a runtime to operate, and operating it is the price of the control and the model freedom you’re buying.
Who should run it, and who shouldn’t
Run Hermes if self-hosting is the point. If you want readable source, no vendor between you and your data, the freedom to point it at any model, and an agent that compounds knowledge on a box you control, there is currently nothing better in open source, and the learning loop is a real differentiator rather than a marketing line.
Reconsider if self-hosting was the tax rather than the goal — we cover the full field in Hermes Agent alternatives. Plenty of people install Hermes because they want an assistant that handles recurring work unattended, then discover they’ve acquired a system to administer. If that’s the trade you’re making, a managed agent gets you the same outcome without the host. Carly is that shape — you email or text it like a colleague, each agent gets its own name and email address, it connects to 200+ integrations, and there’s no VPS, no keys to rotate, and no antivirus exception to write. Free, unlimited Zapier-style workflows; AI agents from $35/month. The tradeoff is honest in both directions: you give up model choice and source access, and you get back every hour you would have spent on upkeep.
The test that settles it is a week of your real recurring work rather than a demo task — because that’s the horizon where both the self-improving memory and the maintenance burden actually show up.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
