Hermes Dashboard and Browser TUI
Start with hermes.mdx for shared setup and an overview of all four invocation paths.
NVIDIA NeMo Fabric starts Hermes’s dashboard as a managed FastAPI/Uvicorn service. Browser chat uses a prebuilt terminal renderer through a PTY/WebSocket bridge, attached to that service’s in-process TUI backend. It does not call the native API gateway or embedded SDK agent.
Installation
Complete the shared installation first. The dashboard path additionally requires:
- Backend dependencies: FastAPI, Uvicorn, and the other dependencies in pinned Hermes’s
webextra, plusaiohttpfor authenticated readiness checks. The repository’s Python 3.14just install-hermes-agentrecipe already provides them. - Frontend assets: Both the built browser assets and terminal renderer used by browser chat, plus a supported Node runtime.
- Host: POSIX, for the managed backend and browser terminal bridge.
For a separately prepared Hermes environment, set ADAPTER_PYTHON to its interpreter and HERMES_AGENT_CHECKOUT to your pinned v0.22.7 source checkout, then install the web extra and the pinned readiness dependency:
Keep this ADAPTER_PYTHON value when following the frontend setup commands to build both frontends in that pinned checkout. Browser assets must be available at hermes_cli/web_dist/index.html or the directory selected by HERMES_WEB_DIST. Runtime startup uses these prebuilt assets; it does not install or build them.
Configuration
Use normalized model, workspace, tools, and telemetry fields with these harness settings:
Open http://127.0.0.1:9119. The native API is optional. To attach terminal clients too, add "terminal": {"port": 9119}; both client paths share one backend process and start separate conversations by default. Omit sdk to also start the embedded Hermes AIAgent alongside the dashboard process.
Authentication
The adapter generates a per-start credential for the TUI backend. Hermes embeds that credential in the served page, and the browser uses it automatically. The dashboard has no separate user sign-in on loopback: anyone reaching the port can use its pages and browser chat. Limit access to the runtime user; port forwarding must have its own access controls.
The TUI credential is distinct from the model-provider API key and the API gateway’s api.token. When terminal attachment is also enabled, the same TUI credential is published in interfaces/tui.token for the attachment helper. This does not add a browser sign-in gate or per-user isolation.
Ownership
The adapter pins browser chat to the managed runtime profile and its existing TUI service. Named-profile chat is unsupported. Other dashboard pages remain native; gateway replacement, profile/configuration edits, and restart/migrate actions are outside the qualified ownership model and can bypass generated configuration.
Conversations, Relay, and Testing
Browser turns use native conversations and responses. They do not become SDK invocations or normalized RunResult objects. Managed native agent turns finalize and flush Relay segments when execution returns, preserving native history.
See the TUI example for browser input and artifact inspection. Refer to the shared limitations for approvals, health/activity, and lifecycle boundaries.