Hermes Agent Review: The AI Operator Behind My Whole Business

Quick verdict

I’ve used Hermes Agent every day for about three months. It’s my daily driver. It runs my research, my coding, my writing pipeline, and the automations that keep this site alive. It’s the most important piece of software I have on this machine. Highly Recommended. Full stop.

• Best for: builders, solo operators, tinkerers, and anyone who wants one agent that does research, code, writing, and ops without stitching together five tools.
• Not for: anyone who wants a polished consumer chatbot, or who isn’t willing to think about API keys and model choice. The CLI is the product.
• Biggest strength: the closed learning loop (memory + skills + session search) plus a real multi-agent Kanban that ships work, plus a genuinely model-agnostic architecture.
• Biggest weakness: surface area. There’s a lot going on. The team just added a bare-bones mode for people who want the minimal path, and I think that’s the right move.
• Disclosure: I’m not on Nous Research’s payroll. I pay for my own inference. Hermes runs this site, so this isn’t a neutral review. There’s no business relationship with the maintainers.

Open source

Hermes Agent by Nous Research

An open source AI agent for builders who want one tool for research, code, writing, and ops.

Read the docs

License

MIT

Current version

v0.16.0

Stars

~185k

Models

30+ providers

Platforms

Linux, macOS, Windows, WSL2, Termux, Docker

What is Hermes Agent?

Hermes Agent is an open source AI agent that lives in your terminal and across messaging platforms. It can spawn sub-agents, run a multi-agent Kanban, schedule jobs, and remember you across sessions. If you’ve been watching Google’s Antigravity launch or Claude gaining computer control on Mac, Hermes is the model-agnostic counterpoint. One runtime. 30+ providers. No vendor tie-in.

It’s built by Nous Research, the open weights shop that’s been shipping the Hermes model family since 2023. The agent is a separate project from the models, and that distinction matters. You can run it on Nous Portal’s 300+ model catalog, or bring your own keys to Anthropic, OpenAI, Google, xAI, MiniMax, Hugging Face, OpenRouter, NVIDIA NIM, DeepSeek, Qwen, Ollama, LM Studio, or any OpenAI-compatible endpoint. So it isn’t locked to one vendor.

The current version is v0.16.0, called the “Surface Release,” which dropped on 2026-06-05. The GitHub repo was created 2025-07-22. As of mid-2026 the repo is sitting at roughly 185,000 stars, with a release cadence that’s been biweekly to monthly for most of the year. The pace shows in the product. The v0.16.0 window alone was 874 commits, 542 merged PRs, and 170 community contributors.

Specs, pricing, and key details

• License: MIT, open source. The full text lives in the repo.
• Current version: v0.16.0 (tagged v2026.6.5), released 2026-06-05.
• Pricing for the agent itself: free. You bring your own API keys.
• Pricing for the recommended gateway: Nous Portal is a paid subscription that bundles 300+ models and a tool gateway. I won’t quote a price because there isn’t one in the docs. Want managed inference plus a tool gateway? The portal is the path. Otherwise, bring your own keys.
• Platforms: Linux, macOS, Windows (native installer in v0.16.0 plus WSL2), Termux on Android, Docker.
• Models supported: 30+ providers out of the box. OpenAI, Anthropic (including Claude Max OAuth), Google Gemini, xAI Grok, OpenRouter (200+), MiniMax, Hugging Face, NVIDIA NIM, DeepSeek, Qwen, Ollama, LM Studio, custom OpenAI-compatible endpoints, and more. So you can switch with one slash command mid-session.
• Install: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash on Linux, macOS, or WSL2. iex (irm https://hermes-agent.nousresearch.com/install.ps1) on Windows. pip install hermes-agent if you want the package.
• Repo: github.com/NousResearch/hermes-agent
• Docs: hermes-agent.nousresearch.com/docs

My experience

I started using Hermes about a week after it launched. I was already running several other agents and toolchains at the time, so I was tired of stitching them together. Hermes was the first one that didn’t feel like I was duct-taping five things to get one job done.

Here’s what I actually use it for, in order of how often:

Research. Most of my editorial research runs through Hermes. First, the agent takes a topic and breaks it down. Then it runs web searches and X searches in parallel, pulls primary sources, and comes back with a cited source pack. The pack that fed this review is one of those. Specifically, I hand it to a sub-agent for QA, then to a draft agent. The FTS5 session search across my past work means it can also pull in what I’ve already researched on a topic, and that’s the part that compounds over time.

Coding. I run it on real codebases, not toy examples. It handles multi-file refactors, debugging, scaffolding, and the boring cleanup work that takes me an hour but takes the agent about four minutes. For this, sub-agents with isolated context are the move. The parent stays on the main task, the child goes and digs.

Writing and editing. This review exists because Hermes built the editorial pipeline, ran the source pack, drafted the outline, and helped me write it. I still own every word, but the agent does the boring 80%, so I get to focus on the part that needs a real person.

How I run it daily

Kanban. This is the part nobody else does. The Kanban is a real SQLite-backed task board at ~/.hermes/kanban.db, where each worker is a real OS process with its own profile, its own memory, and its own model. I run a nightly vibe-build that fans out four parallel workers, and a parent orchestrator that watches the results. I’ve shipped entire projects through this thing while I was asleep. This is the actual shape of agentic ops, not a toy demo.

Cron. Scheduled jobs that I don’t want to babysit. First, a daily research brief. Then a watchdog that pings me if something breaks. Plus a content radar that scans for new Hermes-adjacent releases. So cron jobs can be plain scripts (no agent) or full LLM-driven runs with delivery back to a Telegram chat.

Memory and skills. The agent actually remembers me. Like, not in a “this is my third visit” cookie way. In a “you told me this last month and it still applies” way. The bound is real: 2,200 characters on MEMORY.md and 1,375 on USER.md. Over time, the agent consolidates what it knows, and the skills system lets it pull in domain knowledge on demand. Those skills are also compatible with the open agentskills.io spec, which is the right call. For a look at how the Skills-and-MCP pattern plays out inside Google’s take, see my Google Antigravity Skills and MCP review.

What I like

The closed learning loop. Memory plus skills plus session search plus the model-agnostic core. That’s the actual product. That’s what makes it feel like an operator that gets to know you, not a chatbot.

Model-agnostic by design. I can flip from MiniMax to Claude to GPT-5.5 mid-session with one slash command. If your vendor’s model is on a bad week, you know why this matters. I haven’t had to rewrite a workflow when a model has changed underneath me.

The Kanban. I keep coming back to this because nobody else has it. Durable rows. Real OS processes. Human-driven or agent-driven. The thing actually ships work.

The release cadence. 170 contributors in a single release window, 542 merged PRs, 874 commits, 113 reactions on the v0.16.0 release post. That’s a real team shipping, not a hobby project.

The community. The Discord is alive. Issues get triaged. So security CVEs (the Starlette pin, the subprocess credential stripping) are owned publicly in the release notes, not buried.

Native Windows in v0.16.0. Big deal for the audience. A native PowerShell installer plus a real Electron desktop app. Plus the Linux/macOS/WSL2 path, plus Termux on Android, plus Docker.

The CLI is actually good. Multiline editing, interrupt-and-redirect, streaming tool output. Not a chat wrapper. That’s the part that separates a tool you use from a tool you abandon after a week.

What I don’t like

Surface area. This is the only real complaint I have, and I’ve written it before: there’s a lot going on. Skills, profiles, plugins, Kanban, cron, MCP, voice, image gen, browser, secrets, and on. For someone new, that’s overwhelming. The team just added a bare-bones mode for the minimal path. That’s the right move. Start there.

The marketing language is a bit much. The README calls Hermes “the only agent with a built-in learning loop.” It’s true-ish, but it sets expectations too high. Skills do improve over time, but it’s not magic. The bounds are real, consolidation is slow, and you have to teach it. So: be patient.

18,500 open issues on GitHub. That’s a lot. Some of it is the pace of releases, some of it is real bugs, some of it is feature requests. I’m not going to sugarcoat it. On the plus side, the team closes them in batches every release, and the security-tagged ones get handled fast. The v0.16.0 window alone closed 399 issues, including 16 that were security-tagged. Out of the box, the built-in secrets manager is Bitwarden.

The built-in secrets manager is Bitwarden. Not a real vault. If you want a proper credential broker (encrypted at rest, scoped per agent, time-limited materialization, audit logs), then you wire your own. The agent makes that easy, but the box doesn’t come with one.

Mobile is Termux. Fine for tinkerers. It’s not for normal humans. So it could be better.

Default sub-agent cap is 3 concurrent. Configurable, no hard ceiling anymore since v0.16.0 removed the depth cap, but you should know.

Who this is for

You want this if you’re a builder or a solo operator who wants one agent for research, code, writing, and ops. You want this if you want to own your stack instead of renting it from a vendor whose model is on a bad week. You want this if you’ve been duct-taping five tools to get one job done. You want this if you’re multi-agent curious and you actually want to try a Kanban, not just read about one.

Who should skip it

• Anyone who wants a polished consumer chat app. Use ChatGPT or Claude.ai. That’s what those products are for. If you’re willing to install a CLI and own your stack, Hermes will reward you.
• Anyone who isn’t willing to think about API keys and model choice. The portal helps, but the core is CLI-first. So be ready to bring at least a little operator mindset.
• Anyone who needs a hosted, managed, SOC2-ready enterprise product out of the box. That’s not what this is.
• Anyone who wants a single-vendor lock-in. The whole point of model-agnostic is the opposite.

Alternatives worth considering

I’m not doing a head-to-head benchmark. There isn’t a public apples-to-apples data set. This is scope and architecture, not synthetic numbers.

• Claude Code CLI. Anthropic’s coding agent, polished, model-locked to Claude, narrower in scope. Great if all you want is a coding pair and you are already in the Claude ecosystem. For the broader Claude angle, see my Anthropic Claude controlling your Mac take.
• OpenAI Codex CLI. OpenAI’s coding agent, same shape, model-locked to OpenAI, just got a /goal mode in 0.128.0 that Hermes explicitly borrowed from. Good for OpenAI-native teams. My GPT-5.2-Codex coverage has more on what that release added.
• OpenHands. Open source coding agent, less breadth than Hermes, no Kanban. So it’s solid if coding is all you need.
• Aider. Terminal pair programmer, focused, model-agnostic, much narrower. Great for one-off refactors.
• Goose. Block’s open source agent, similar scope, less shipping velocity in my observation. The v1 release was solid; the v2 felt slower to me.
• LangGraph / AutoGen / CrewAI. These are libraries, not applications. You can build on top of them, but you ship faster with Hermes because the runtime, tools, memory, Kanban, cron, and gateway are already there. For the wider agent landscape, see my Google Antigravity review and how I bridged Antigravity and OpenCode to solve my AI quota problem.

Final verdict

Highly Recommended.

Three months in, Hermes is still my daily driver. The release pace is real, the architecture is right, the community is alive, and the things that break get fixed fast. On its own, the Kanban makes it worth the install. Beyond that, the model-agnostic core makes it future-proof, and the closed learning loop makes it actually useful over time, not just on day one.

It isn’t for everyone. Surface area is the tax you pay for breadth, and the README’s “only agent with a built-in learning loop” line sets expectations that the bounds in MEMORY.md and USER.md will eventually temper. If that bothers you, the bare-bones mode is the right starting point.

So, to wrap it up: I run this site on it. My business runs on it. I’m not going back.

That’s it. That’s the review.

FAQ

Is Hermes Agent free? Yes, in the sense that the agent itself is open source under MIT. You bring your own API keys, or you can subscribe to Nous Portal for managed inference plus a tool gateway.

Is Hermes Agent the same as a Hermes model from Nous Research? No. Nous Research makes the Hermes model family (Hermes 4 and friends). Hermes Agent is a separate open source project that uses those models and many others.

Does it work on Windows? Yes, natively since v0.16.0. In addition, it runs on Linux, macOS, WSL2, Termux on Android, and Docker.

Can I use it with my own LLM endpoint? Yes. Any OpenAI-compatible endpoint works. So Ollama, LM Studio, custom proxies, all of it.

How is this different from Claude Code or Codex CLI? Those are coding agents tied to one vendor’s models. Hermes is broader (research, writing, ops, multi-channel messaging, voice, image gen) and model-agnostic by design. So if you need more than coding, Hermes is the wider door.

Is it hard to set up? One curl command. Then you pick a model provider. There’s a bare-bones mode now if you want the minimal path.

Reviewed over 90 days of daily hands-on use, including the v0.16.0 “Surface Release” cycle. Source pack and full changelog linked in the source notes.