Table of contents
How Hermes Talks to the Outside World
Beyond the LLM itself, Hermes integrates with five categories of external service: AI inference, MCP tool servers, web search, voice backends, and the OpenAI-compatible API mode. Each one is driven by environment variables and either a config file entry or an interactive CLI command. This lesson walks through the shape of each integration so you can decide what to wire up and what to leave for later.
Primary source: the integrations index at https://hermes-agent.nousresearch.com/docs/integrations/. Refer there for the canonical, per-service configuration steps.
AI and Model Providers
Hermes works with OpenAI, Anthropic, Google, Nous Portal, OpenRouter, Ollama, and any OpenAI-compatible endpoint. Provider selection isn't just 'which company are we billing' — it's also about routing:
- Provider Routing — pick which provider handles each request based on cost, latency or quality.
- Fallback Providers — primary goes down, Hermes automatically retries against the backup.
- Credential Pools — rotate across multiple keys to spread load or tolerate a revoked key without downtime.
Interactive management is via hermes model and hermes auth. These handle OAuth flows and API key storage, and always write secrets to ~/.hermes/.env rather than to the YAML config.
MCP: Tools Without Writing Tools
The Model Context Protocol is the reason you don't have to write a Python integration every time you want the agent to touch a new system. An MCP server exposes capabilities over a standard protocol; Hermes subscribes to it. There are already MCP servers for GitHub, databases, file systems, browser stacks, and a long tail of internal APIs.
Manage MCP from the CLI:
hermes mcp list
hermes mcp add <server>
hermes mcp test <server>
hermes mcp configure <server>
hermes mcp serve # run Hermes itself as an MCP serverThat last one is worth underlining — you can expose Hermes to other MCP-compatible agents, turning it into a tool other systems can call.
Web Search: Four Backends, Auto-Detected
Web search is a commodity Hermes doesn't try to reinvent. Four backends are supported:
- Firecrawl — the default.
- Parallel — fast, cost-focused.
- Tavily — answer-oriented.
- Exa — neural search.
Each is selected by its own API key environment variable (FIRECRAWL_API_KEY, TAVILY_API_KEY, etc.). If you don't set web.backend explicitly, Hermes picks automatically based on which credentials it can find. Once configured, the built-in web_search and web_extract tools route through your chosen backend transparently.
Voice Backends: TTS and STT
Text-to-speech providers: Edge TTS (free, default), ElevenLabs, OpenAI, MiniMax, NeuTTS. Speech-to-text providers: local Whisper, Groq, OpenAI. Edge TTS is the recommended starting point because you can validate the whole voice pipeline without spending anything.
Once a voice provider is configured, voice message transcription works across every messaging platform that supports it — Telegram, Discord, WhatsApp, Feishu, QQ Bot, etc. You don't reconfigure voice per platform; you reconfigure once and every supported gateway uses the same stack.
OpenAI-Compatible API Server
Running hermes as an OpenAI-compatible HTTP server turns the whole agent into a drop-in backend for any OpenAI-client app — Open WebUI, LobeChat, your own Next.js prototype. The port is configurable, and the server speaks the standard /v1/chat/completions shape. For teams that already have a frontend they like, this is the shortest path to 'agent-powered, same UX'.
External Memory Providers
Beyond the local MEMORY.md / USER.md files, Hermes can back memory with seven external providers including Honcho and Mem0. Each handles user profiles and conversation history with its own persistence model — useful when you want memory to follow a user across machines, or when you need cross-session personalisation at scale.
Management commands:
hermes memory setup
hermes memory status
hermes memory off # disable external memory, fall back to files
hermes honcho # Honcho-specific peer identity + recall tuningPlugins
The plugin system is the escape hatch: custom tools, lifecycle hooks, alternative memory providers, custom context engines — all loaded from ~/.hermes/plugins/ without touching core code. If you're integrating something bespoke that doesn't fit any of the above categories, a plugin is usually the cleanest answer.
Next Steps
Lesson 6 is the practical 'how to actually use this well' guide — prompt discipline, AGENTS.md, keyboard shortcuts, operational security. Lesson 7 is for developers who want to contribute, extend, or ship their own provider. If you're a UK team and the integration count is starting to feel daunting, that's exactly what SetupClaw's managed service is for.