Nous Research's Hermes Agent is still drawing developer attention, but the real story isn't a clean flight from commercial tools. It's that open agent runtimes now have a credible switching path.
The numbers still matter. Nous Research's Hermes Agent launched on February 25, 2026, under an MIT license, and GitHub mirrors and release trackers now put the NousResearch/hermes-agent repository around 220,000 stars in late July. That is not a small developer fad. It is a serious signal from people who install tools, break them, file issues, fork them, and then decide whether the thing deserves to stay in their workflow.
What makes Hermes Agent distinct isn't only the star count. Look at the architecture. The project's own line is 'the agent that grows with you,' and the product page backs that up with persistent memory, auto-generated skills, session history, scheduling, delegation, web tools, and isolated sandboxing across local, Docker, SSH, Singularity, and Modal backends. Most agent wrappers still behave like a fresh assistant every time you open them. Hermes is built around the opposite idea: the agent should remember what worked, turn it into a reusable skill. Then improve the next run. That is the sell.
It is also model-agnostic. You can point Hermes at different backends instead of building your whole agent layer around one lab's API or one subscription plan. If you're a developer, that matters because the model market is still moving too fast to swear loyalty to anyone forever. OpenAI, Anthropic, Google, DeepSeek, Qwen, Kimi, you name it. The rational position is to keep the agent layer portable.
The switching path is now real #
Nous Research pushed the desktop story forward at the end of May and early June. GitHub release listings show Hermes Agent v0.15.2 dated May 29, 2026, while outside coverage of the desktop preview placed the native app announcement on June 2. The distinction matters. The CLI patch and the desktop preview were not the same event, but together they made Hermes easier to try for people who didn't want to live entirely in a terminal.
The desktop README describes a native app for macOS, Windows, and Linux with streaming tool output, a file browser, live tool activity, voice, settings, and the same memory and skills as the CLI and gateway. That is not decoration. It moves Hermes from 'interesting repo' to a tool more people can actually test without a long setup afternoon.
The OpenClaw moment gave that testing a sharper edge. In June 2026, Anthropic's own help center described planned changes to Claude Agent SDK usage under Claude plans, then posted a June 15 update saying those changes were d. A Hacker News thread about Claude subscriptions and OpenClaw drew hundreds of comments. That correction matters. The issue was not an April restriction that permanently cut off OpenClaw; it was a messy June pricing and access fight around third-party agent use.
Hermes was ready for that fight. Its migration guide says the hermes claw migrate command imports OpenClaw and legacy Clawdbot or Moldbot setups, including memory, skills, persona files, settings, and selected credentials, with a preview before changes are applied. OpenClaw itself is still large. GitHub's OpenClaw organization page showed the main openclaw/openclaw repository above 379,000 stars in June, not 345,000. So this is not a story about a weak rival collapsing. It is about users learning that their agent state can move. That is switching friction, reduced to a command.
Nous is selling the layer around the agent #
There is a business model here, and you should look at it plainly. Nous Portal is the paid convenience layer around the MIT-licensed agent. Nous Research's Portal pages describe a single account covering model access, tool gateway usage, and Hermes Cloud hosting, with hundreds of models routed through the platform and external tools such as web search, image generation, text-to-speech, and browser automation billed through the same system.
The open core stays free. The setup, the billing, the hosted cloud instance - none of that is. That is a familiar playbook, but it is a sensible one. CrewAI has also kept an open-source core while charging around managed deployments and enterprise tooling. Hermes is different because the open project itself is the hook: memory, skills, migration, desktop, cloud, and portal all point back to one runtime that keeps getting more useful the longer it runs.
The market gives Nous room to make that bet. Gartner said in an August 2025 press release that 40% of enterprise applications could include task-specific AI agents by the end of 2026, up from less than 5% at the time. MarketsandMarkets puts the AI agents market at $7.84 billion in 2025 and projects $52.62 billion by 2030. You do not have to treat those forecasts as destiny. You should treat them as a sign that companies are budgeting for agents whether the tooling is settled or not.
For OpenAI and Anthropic, that is the uncomfortable part. Anthropic has the Claude Agent SDK, formerly the Claude Code SDK, and OpenAI has its own agent tooling. But developer loyalty in the agent layer is thinner than loyalty in the model layer. Developers pick a model for output quality. They pick an agent framework for control and portability - for how painful it is to leave. Frankly, Hermes is taking the smarter position by refusing to assume any one model provider wins forever. Stars are not revenue. They are not enterprise deployments either. But 220,000 or so GitHub stars, a working OpenClaw migration path, a desktop app, and a paid portal around an MIT-licensed runtime add up to something more solid than hype. The question now is whether Nous can turn developer attention into retained usage before the bigger labs copy the best parts and bundle them into their own stacks.
Also read: Chinese biotech is stealing the emerging market trade that AI dominated for two years • Apple became the world's most valuable company by refusing to join the AI spending race • Nvidia bets $1.5 billion on Amkor to break its chip packaging bottleneck