Hermes Agent Adapter

​

Manifest

runtime:
  adapter:
    type: "hermes"
    # Entry point can be an AIAgent instance, a class, or a factory function
    entry_point: "src.main:agent"
  capabilities:
    manages_own_history: true

runtime:
  lifecycle: "daemon"
  capabilities:
    manages_own_history: true

​

Best For

• Agents that learn and improve across sessions.
• Long-running daemon agents with persistent state.
• Multi-provider LLM workflows (OpenRouter, OpenAI, Anthropic, local endpoints).
• Agents needing built-in tools (terminal, browser, code search).

​

How It Works

1. Persistence — HERMES_HOME is redirected to CHARM_WORKSPACE_DIR/.hermes, so SQLite session databases, FTS5 search indexes, auto-generated skills, and memory files all persist across daemon restarts.
2. Streaming — Hermes’s stream_delta_callback is wired to Charm’s SSE emitter for real-time token output in the Store frontend.
3. Headless mode — The agent runs with quiet_mode=True to suppress TUI spinners and terminal chrome.
4. Tool tracking — Hermes’s tool_start_callback and tool_complete_callback feed into Charm’s tool usage analytics.
5. Threaded Streaming — Because Hermes executes synchronously, the adapter wraps run_conversation in a background thread and bridges stream_delta_callback to Charm Runner via a thread-safe Queue, ensuring real-time token output.
6. Lazy Configuration — If entry_point points to a factory function, the adapter automatically injects environment variables like model (from CHARM_HERMES_MODEL) before instantiation.

​

Plugin Compatibility Notes

• State and Memory Plugins: Hermes manages its own memory system natively using a localized SQLite FTS5 database (CHARM_WORKSPACE_DIR/.hermes/memory.db). It does not support and will completely ignore any custom memory plugins configured in charm.yaml (e.g., memory: { provider: "postgres" }).
• Output Renderers (Custom UI): The adapter automatically parses _charm_render_type JSON objects outputted by Hermes, allowing you to render dynamic UI components in the Charm frontend.

​

Environment Variables

​

Runtime Image

runtime:
  custom_image: "ghcr.io/charm-ai-labs/charm-runner-hermes:latest"

​

Operational Notes

• Use lifecycle: "daemon" for agents that need persistent memory across sessions.
• Always set LLM API keys via charm.yaml environment_variables section, not hardcoded in source.
• Test with charm validate before publishing.
• The adapter is distributed as a separate pip package (charm-adapter-hermes). Install it in your runtime image or let the runner install it from requirements.txt.