GoldieBench Blog · 8 min read
Hermes Agent MCP Server: What It Is, How To Set It Up Both Ways, And The Models Behind It
Hermes Agent MCP server explained: add MCP tools to Hermes, run Hermes as an MCP server for Claude Code or Codex, and which models to run behind it.

The Hermes Agent MCP server covers two things: outside MCP servers that Hermes uses as tools, and Hermes itself running as an MCP server so Claude Code, Codex or Cursor can use it.
Hermes Agent is the free, open-source agent from Nous Research, and the CLI help describes hermes mcp as the way to "Manage MCP server connections and run Hermes as an MCP server."
MCP, the Model Context Protocol, is the shared standard that lets an agent call tools living in another program.
The server side now ships as the official mcp-serve plugin, listed in the catalog on 21 October 2026.
I'm Julian Goldie, and in this post I'll explain both directions, how to set them up and how to stay safe, and then I'll show you how the models you'd run behind your MCP agents score on GoldieBench.
Every command here comes from the official Hermes MCP docs, the plugin page or the CLI help.
The Hermes Agent MCP server at a glance
| Job | Direction | Command |
|---|---|---|
| Browse the Nous-approved catalog | Hermes as the client | hermes mcp or hermes mcp catalog |
| Install a catalog server | Hermes as the client | hermes mcp install deepwiki |
| Add any other server | Hermes as the client | hermes mcp add <name> with --url or --command |
| Add the Codex CLI as a tool | Hermes as the client | hermes mcp add codex --preset codex |
| List and test servers | Hermes as the client | hermes mcp list and hermes mcp test <name> |
| Sign in on a headless box | Hermes as the client | hermes mcp login <server> --flow device |
| Expose Hermes to other agents | Hermes as the server | hermes plugins install mcp-serve |
| Connect Claude Code, Codex or Cursor | Hermes as the server | "command": "hermes", "args": ["mcp-serve"] |
What is the Hermes Agent MCP server?
When Hermes is the client, it gains tools from MCP servers such as GitHub, databases, file systems, browser stacks and internal APIs.
When Hermes is the server, another agent can list and read your Hermes conversations, wait for new messages, send replies and answer approval requests.
MCP support ships with the standard Hermes install, so the client side needs nothing extra.
Hermes and the mcp-serve plugin are free and MIT licensed.
The problem it solves
Most people now run more than one agent, and those agents usually can't see each other's work.
A coding agent finishes a job, and then a person copies the result into a chat app by hand.
MCP lets one agent call the other directly, so the handover disappears.
In my Codex and Hermes video, Codex initialised the Hermes MCP server, listed its tools and returned my previous Hermes conversations.
How to add MCP servers to Hermes
The fastest route is the catalog, which the docs say holds servers that Nous staff have reviewed and merged.
Run hermes mcp for the interactive picker, hermes mcp catalog for a plain list, and hermes mcp install deepwiki to install an entry by name.
Catalog entries are disabled by default, so you only install what you need.
You can also ask Hermes in a chat, for example "add the Linear MCP", and a setup card appears.
In the desktop app, v0.22.0 replaced the MCP tab with a Connectors page that lists hosted apps and your MCP servers together.
For anything else, hermes mcp add <name> takes --url for a remote server or --command and --args for a local stdio server.
The docs' one built-in preset, hermes mcp add codex --preset codex, runs codex mcp-server over stdio and needs the codex CLI on your PATH.
In YAML, servers live under mcp_servers in ./.hermes/config.yaml, with command and args for local servers or url and headers for remote ones.
If you're moving from Claude Code, the docs say hermes import-agent claude-code migrates its MCP servers, and my Hermes import agent post explains it.
How to manage MCP servers in Hermes
| Command | What it does |
|---|---|
hermes mcp list | It lists configured MCP servers. |
hermes mcp test <name> | It tests one server's connection. |
hermes mcp configure <name> | It toggles which tools a server contributes. |
hermes mcp login <server> --flow device | It signs in to an OAuth server without a browser. |
hermes mcp reauth --all | It re-authenticates every OAuth server one at a time. |
hermes mcp remove <name> | It removes a server. |
/reload-mcp | It reloads servers from config inside a chat. |
Hermes names each tool mcp_<server_name>_<tool_name>, and each server gets its own toolset called mcp-<server>.
A running gateway watches config.yaml and picks up added or removed servers within about a minute.
The v0.22.3 release notes also list MCP progress notifications streamed as live tool progress.
How to run Hermes as an MCP server
Check your version first, because the command changed recently.
Builds such as v0.22.1 ship the server built in as hermes mcp serve, which the help describes as running Hermes as an MCP server to expose conversations to other agents.
The current docs ship it as the mcp-serve plugin, which needs Hermes 0.22.0 or later, and you install it with hermes plugins install mcp-serve.
Your MCP client launches it over stdio, so you add "command": "hermes" with "args": ["mcp-serve"] to the client's mcpServers config.
If a client still uses ["mcp", "serve"], the plugin page says you change it to ["mcp-serve"], and the tool names stay the same.
| Tool | What another agent can do |
|---|---|
conversations_list | It can list active conversations across platforms. |
conversation_get | It can read one conversation's details. |
messages_read | It can read recent messages. |
attachments_fetch | It can pull images and media from a message. |
events_poll and events_wait | It can poll or wait for new messages. |
messages_send | It can send a message through Hermes. |
channels_list | It can list the targets you can send to. |
permissions_list_open and permissions_respond | It can see and answer approval requests. |
Reading needs no running gateway, but sending does, because the platform connections live there.
hermes mcp-serve --tools also exposes Hermes's web, browser, vision, image generation, skills and text-to-speech tools.
The plugin page says terminal, file, code execution, process, kanban and delegation tools are never exposed, whatever the profile enables.
Real ways to use the Hermes Agent MCP server
These are the jobs where each direction earns its place, based on what the docs say each part does.
- A coding agent can post a finished build to a client's Telegram or Slack chat by calling
messages_sendthrough Hermes. - A team that lives in one coding tool can answer Hermes approval requests with
permissions_respondinstead of switching windows. - A research profile can gain code and documentation tools from catalog servers such as
deepwiki. - A profile that touches payments can keep only safe tools with an
includelist. - A profile that runs paid image or video servers can declare
tool_costsand stop at a per-chatspend_limit. - Hermes can hand coding jobs to Codex through the
codexpreset, while Codex reaches back into Hermes throughmcp-serve. - A coding agent can borrow Hermes's web search and browser through
--tools, using the providers you've already set up.
I'd add one of these at a time, because a single new link is easy to debug and five new links at once are not.
If you want the wider list of tools worth giving Hermes, my best Hermes Agent tools post covers them.
How to stay safe with MCP
The docs say installing a catalog entry runs whatever its manifest specifies, so read the source:, install.bootstrap: and transport.command: fields first.
Manifests live in the optional-mcps folder of the Hermes repo, and there is no community submission tier.
For stdio servers, Hermes passes only the env you configure plus a safe baseline.
Per-server tools: include and tools: exclude lists keep dangerous tools out of the model's sight, and the docs' Stripe example drops delete_customer.
Hermes strips invisible Unicode tag characters from tool results, which the docs call a prompt-injection channel.
A server marked trust: untrusted asks before every write-capable call, and a per-chat spend_limit caps costed tools.
The plugin page warns that an mcp-serve client can read your conversation history and send messages as your connected accounts.
If your agent also runs commands, pair MCP with a container, as I explain in my Hermes Agent sandbox post.
Limits of the Hermes Agent MCP server
- The Hermes MCP server is stdio only, and the docs suggest a separate adapter if you need HTTP.
- The command moved from
hermes mcp serveto themcp-serveplugin, so older configs need updating. - Sending needs the Hermes gateway running.
- The event queue lives in memory and starts when the server starts.
- Tool calls cost whatever your configured provider charges.
If you'd like a plan for where MCP agents fit your SEO business, you can book a free SEO strategy session with my team.
Which models to run behind your MCP agents
The Hermes MCP server doesn't ship its own model, because MCP only moves tool calls and messages between agents.
The quality of the work still depends on the brain behind each agent, whether that's Hermes, Claude Code or Codex.
GoldieBench is where I score models on real one-shot builds, and these are the current averages from the live board.
| Model | GoldieBench average | Where it fits in an MCP setup |
|---|---|---|
| Claude Opus 5 | 8.27 | It is the strongest single model on the board for the agent doing the building. |
| Hermes MoA | 8.17 | It is a Hermes Mixture of Agents panel, merged by an aggregator. |
| GPT-5.6 Sol | 8.16 | It is OpenAI's top 5.6 model and sits just behind Hermes MoA. |
| Claude Fable 5 | 8.10 | It is close behind the top three. |
| Kimi K3 | 7.89 | It is a strong open option for a Hermes profile. |
| GLM-5.2 | 7.77 | It is another solid open-weight brain. |
| Claude Opus 5.5 | 7.57 | It is the newer Opus, and it scores below Opus 5 on this board. |
| Gemini 3.6 Flash | 7.08 | It is a cheaper option for light tool-calling work. |
| Claude Sonnet 5 | 7.01 | It is a mid-tier Claude choice. |
GoldieBench measures one-shot builds, not MCP tool calling, so treat these numbers as a guide to raw build quality rather than a test of the protocol.
A sensible split is a strong model on the agent that writes code and a cheaper one on the agent that mostly reads and relays messages.
For my picks by job, read my best Hermes Agent model guide.
Verdict
The client side of Hermes MCP is mature, with a reviewed catalog, OAuth with device-code login and per-server tool filtering.
The server side now lives in the official mcp-serve plugin, and it's the cleanest way to let Claude Code or Codex reach your Hermes chats.
Also On Our Network
- Agent Operatorsthe ten-step Hermes Agent MCP server setup, both directions
- AI Income Deskseven ways a business uses the Hermes Agent MCP server
- AI Tool Verdictevery Hermes MCP feature, scored with a verdict
- agentos.guidethe Hermes Agent MCP server quick start for an Agent OS
- aiprofitboardroom.comthe Hermes MCP server explained on the AI Profit Boardroom blog
If you run an agency and want help mapping agents to client work, you can book a free strategy session with my team.
Add one reviewed server, install mcp-serve for the other direction and pick the brain from the live board, and you'll get the most out of the Hermes Agent MCP server.
FAQ
What is the Hermes Agent MCP server?
It covers two directions. Hermes connects to outside MCP servers to gain tools, and Hermes can run as an MCP server so Claude Code, Claude Desktop, Cursor or Codex can read and send your Hermes messages. The hermes mcp command manages both.
How do I add an MCP server to Hermes?
Run hermes mcp for the catalog picker, hermes mcp install deepwiki to install a catalog entry by name, or hermes mcp add with --url or --command for any other server.
How do I run Hermes as an MCP server?
On current builds, run hermes plugins install mcp-serve and point your MCP client at the command hermes with the args mcp-serve. Older builds such as v0.22.1 use the built-in hermes mcp serve.
Does the Hermes MCP server come with its own AI model?
No. MCP only moves tool calls and messages between agents. The brain is whichever model each agent uses, and on GoldieBench Claude Opus 5 averages 8.27, Hermes MoA 8.17 and GPT-5.6 Sol 8.16.
Can another agent run commands through the Hermes MCP server?
No. The mcp-serve plugin never exposes terminal, file, code execution, process, kanban or delegation tools, even with the --tools flag.
Is the Hermes Agent MCP server free?
Yes. Hermes and the mcp-serve plugin are MIT licensed. Tool calls still cost whatever your configured model or provider charges.


