Nous Research shipped Hermes Desktop v0.15.2 today, a native cross-platform front end for the Hermes Agent on macOS, Windows, and Linux, under MIT license. The architecture choice worth flagging: native, not Electron. The same agent core that runs the CLI sits behind the UI, so there is one backend with multiple surfaces rather than a forked desktop version. The desktop installs via direct installers on macOS and Windows or a terminal script on Linux with an `--include-desktop` flag. Provider-agnostic on the model side: Nous Portal, OpenRouter, OpenAI, and any compatible endpoint. MCP tool support is in. Five sandbox backends for tool execution are configured out of the box: local, Docker, SSH, Singularity, and Modal.
The UI gives you streaming responses and live tool activity, with a side pane that previews web pages, files, and tool outputs as the agent works. Memory is shared with the agent core, not a separate desktop memory: FTS5 session search with LLM summarization handles cross-session recall via persistent, agent-curated memory. The release notes do not detail what v0.15.2 changes specifically versus prior versions; the launch framing is around the desktop surface itself rather than the version delta. No published benchmarks. Hardware requirements are not specified in the launch material. For builders who have been wiring their own UI around the Hermes Agent CLI, this is the official native replacement.
Two ecosystem threads worth tracking. First, the five-sandbox backend list is the substantive feature for agent builders. Most desktop agent UIs ship with local-exec only; some add Docker. Hermes Desktop's bundling of SSH (for running tools on remote machines), Singularity (HPC container format used in scientific computing), and Modal (the cloud-function-style serverless platform) means the agent can reach into infrastructure environments the operator does not have to wire up by hand. That is a real adoption lever for teams in research, HPC, or distributed systems who want agent execution without inventing the sandbox layer themselves. Second, the relationship to Memory OS from yesterday is interesting precisely because it is not a relationship. Memory OS by Claudio Drews is a community-built 6-layer memory stack on top of Hermes Agent, and Hermes Desktop ships with Hermes's own memory model (FTS5 sessions plus LLM-summarized recall) instead. Two layers of the same ecosystem moving independently: official UI from Nous, community memory architecture from Drews. Whether they end up integrating, competing, or staying parallel will say something about how open-source agent ecosystems converge as more polish lands.
Monday morning, if you are already running Hermes Agent from the CLI: install Hermes Desktop and treat the native UI as the default. The MCP support means you do not have to redo your tool wiring. If you are evaluating agent platforms for a small team and want something local-first with a real UI, this is now a credible alternative to running Claude Code, Cursor Composer, or building your own Electron wrapper. Five sandbox backends is more than most desktop agents ship with, and the SSH-plus-Modal combination in particular is the path for teams that have agent work spanning local machines and cloud functions. If you are not on Hermes Agent at all, the desktop is an entry point worth trying alongside the CLI to evaluate the framework as a whole, not as a separate decision.