Nous Research launches Hermes Marketplace: plugins, skills, and MCP in one catalog

If you have ever assembled an agent by hand, you know the pain: one MCP server is installed from a GitHub repository, skills live in another directory, configuration files are copied manually, and a week later everything breaks after an update. Nous Research has addressed the problem in a mature way by launching Hermes Marketplace, the official catalog for its Hermes Agent framework.

What's inside

The marketplace brings together three types of extensions: plugins, agent skills, and MCP connectors. Everything is installed with a single CLI command: hermes plugins install. There is no need to hunt down repositories or figure out where to put the files.

The catalog is built around the Portable Agent Plugins v1 format. A package combines an mcp.json configuration file with a skills/ directory. The idea is simple: the instructions for connecting to an external tool and the instructions teaching the agent how to use it live side by side and are installed together. The plugin author writes the package once, while the user receives a ready-made combination of "tool plus the knowledge required to use it."

Modules included in the release

The most intriguing module is Claude Subscription Direct SDK (claude-subscription-directsdk). It routes Hermes Agent execution steps through the local Claude Code CLI, using the Agent SDK path. In practice, this means the agent can consume a Claude Pro/Max subscription quota instead of paying for every token through the API. For users running long agent loops, the budget difference can be substantial. There is an important caveat, however: Anthropic's subscription terms may change, and building a production workflow around this approach without checking the rules is risky.

Two connectors have been added for 3D work and procedural video:

• Blender Lab MCP is an official stdio server. It installs a portable version of Blender 5.1+ without administrator privileges and verifies the SHA-256 hash of the downloaded archive.
• TouchDesigner MCP uses the Streamable HTTP protocol at the local address 127.0.0.1:40404, so it is not exposed externally.

Here is an example workflow: you ask the agent to "build a scene with three cubes and a rotation animation." Through MCP, it opens Blender, creates the objects, and returns the result. Previously, you would have had to write the bridge between the LLM and Blender yourself.

There are also oversight tools. Memory Wiki is a local dashboard for auditing the agent's persistent memory, namely the MEMORY.md and USER.md files. It addresses a long-standing problem: the agent has "remembered" something about you, but you cannot tell exactly what it stored or where the information came from. Telegram Business Mode turns the agent into a bot secretary that prepares draft replies and sends them only after you approve them.

Limitations and conclusion

This is still an early release, and the source says nothing about the quality of third-party plugin reviews. Any marketplace offering executable code and file access creates an attack surface: before installing someone else's plugin, you should inspect its mcp.json file and the permissions it requests. It is also worth remembering that the post is a retelling published in a Telegram channel, so details should be checked against the marketplace's own page.

What does this mean in practice? If you build agents, consider Portable Agent Plugins v1 as a candidate packaging standard. Even if you do not need Hermes, the "mcp.json plus skills/" format is a useful model for your own extensions. If you work on a product, pay attention to the trend: the value of agent platforms is shifting away from the model itself and toward an ecosystem of ready-made integrations. Whoever assembles the catalog fastest will win the users. Founders should also consider whether it is time to release their own connector in such a catalog: it is an inexpensive distribution channel.