Hermes

Hermes (Nous Research) is an agent that runs on a server and answers through gateways such as Telegram and Slack, plus its CLI, TUI and desktop app.

  • Plugin: log has a Hermes plugin, named log, in this repo under hermesPlugin/.
    • One plugin: it used to be two, a portable log (tools, skills) and a native logHooks (hooks), because Hermes loads a folder with a plugin.yaml as native and ignores its plugin.json, only a portable package can declare an MCP server, and only a native plugin can register hooks (hermes_cli/plugins_discovery.py, plugins_loader.py; Portable Agent Plugins). A config.yaml server wins over a package's anyway, so log's MCP server now comes from config.yaml and the plugin is one native folder, hermesPlugin/.
    • Install: log itself first, with logMcp on the PATH. log launch --setup (pick Hermes) copies the plugin into <Hermes home>/plugins/log with the log install's path and removes an old plugins/logHooks that log wrote; then hermes plugins enable log and hermes mcp add log --command logMcp --env "LOG_ACCOUNTS=${LOG_ACCOUNTS}". The Hermes home is HERMES_HOME, else ~/.hermes, else %LOCALAPPDATA%\\hermes on Windows (LOG_HERMES_DIR overrides all three).
    • After setup or update: log setup and log update say two things when they copy the plugin. The first time the Desktop half lands, switch log on in Hermes Settings > Plugins (a package's Desktop half ships switched off). When dashboard/plugin_api.py changed, restart the Hermes gateway (hermes gateway restart; when the gateway runs by hand, stop it and start it again), because the gateway mounts the plugin's routes once at start and restarting the app alone keeps the old ones. They also name a running gateway that started before the routes on disk changed.
    • What it adds: every log play as a log:<play> skill (generated from docs/useCases/ by bun run skills:build), log's GenUI tools and recipes, a read-only log tab in the dashboard and the Desktop half that draws blocks and views.
    • The router hook: pre_llm_call runs scripts/hooks/event.ts UserPromptSubmit --host hermes with the message (and SessionStart first on a session's first turn) and returns { "context": … }, cut to 7,500 characters per event. It finds log by itself: LOG_ROOT, then the log_root that log launch --setup writes, then Bun's global node_modules/@listeningkit/log (BUN_INSTALL or ~/.bun), then npm's global folders. When config.yaml has no mcp_servers.log, it adds a line saying so with the command that fixes it.
    • What it doesn't do yet: it has no hosted option; the hook runs only where log is installed on the same machine as Hermes.
  • Connects:
    • Config: mcp_servers in ~/.hermes/config.yaml. A remote server takes url and auth: oauth; Hermes handles registration, PKCE and refresh, and keeps tokens in ~/.hermes/mcp-tokens/<server>.json. A local server takes command. (Hermes: MCP, MCP config reference)
    • Tool names: Hermes names tools mcp_<server>_<tool>, so log's tools appear as mcp_log_report and so on.
    • Reloading: after a config change, run /reload-mcp.
  • Draws MCP Apps: Hermes's MCP page doesn't mention MCP Apps or ui:// resources, and Hermes isn't in the support matrix. Treat it as text.
    • Views: in Hermes Desktop, report's answer names a ::logview line that log's Desktop plugin draws as the view; Telegram and the CLI get it as text.
    • Show me the browser: for an account that is signed in already, sessionsSignIn { account, open: true } opens its browser through the same gates, hidden (no window or taskbar entry), and browserView streams it into the chat.
    • Sign-in: in Hermes Desktop, sessionsSignIn opens the account's browser and browserView's answer names a ::logbrowser{account="..."} line; the agent writes it on its own line and the Desktop plugin draws the live browser there, on the owner's machine too, so the owner signs in in the chat. Telegram and the CLI sign in in the log console; a streamed browser has no window of its own unless LOG_BROWSER_WINDOW=visible.
  • Guidance:
    • Prompt router: with the plugin enabled, pre_llm_call names the play before each model call. Without it the model routes each request with getGuide({ ask: "<the owner's words>" }) (as mcp_log_getGuide).
    • Server instructions: Hermes's MCP page doesn't mention them.
    • Resources and prompts: Hermes does wrap a server's resources and prompts as tools (list_resources, read_resource, list_prompts, get_prompt), so log's log://agents resource can be read.
    • What's left: guidance comes through the tool descriptions and getGuide.
    • Skills: the plugin's skills are named log:<skill> by Hermes. Whether Hermes loads them as Claude Code does isn't checked by running it.
  • Approvals: Hermes routes MCP form elicitation to its approval prompts, which show as buttons on Telegram and Slack. URL elicitation is declined as unsupported.

On this page

No Headings