Agent Platform Guides

Setup and configuration guides for Hermes Agent, OpenClaw, and Pi Coding Agent — the three most-used self-hosted AI agent platforms in 2026.

June 9, 2026
hermes-agentopenclawpiagent-platformsself-hostedcoding-agents

Agent Platforms

Three open-source platforms dominate the self-hosted agent space in 2026. Each takes a different approach to autonomy, security, and extensibility.

Hermes Agent (Nous Research) is the current #1 on OpenRouter — 271 billion tokens processed daily across 19 messaging platforms. Seven-layer security model, broad LLM provider support, and a skills system. Self-hosted, runs on your infrastructure.

OpenClaw holds the record for fastest GitHub star growth ever (250K+). NVIDIA built NemoClaw on top of it. Skills & Tools model, massive community plugin ecosystem. Runs as the engine for some of the largest production agent deployments.

Pi (Earendil Inc.) is the minimal agent harness. No MCP, no sub-agents, no plan mode built in — everything is an extension. OpenClaw uses Pi's SDK as its underlying engine. For developers who want full control over context engineering.

Which Platform?

Platform Comparison

Hermes Agent
Best for teams that need cross-platform messaging (WhatsApp, Telegram, Slack, Discord) with strong security. Multi-user, multi-channel, enterprise-ready. #1 usage on OpenRouter.

Values: Free (self-hosted)

OpenClaw
Best for developers who want the largest plugin/extension ecosystem. Skills & Tools marketplace with thousands of community contributions. Most GitHub stars of any AI framework.

Values: Free (self-hosted)

Pi Coding Agent
Best for developers who want a minimal CLI agent with no bloat. Extensible via TypeScript extensions and packages. Used as the engine inside OpenClaw.

Values: Free (MIT)

Note:

All three are free and self-hosted. You provide the LLM API keys or run models locally. Costs depend on which models you use, not the platform itself.

Provider Wiring

All three platforms support the same major LLM providers. Wire them to your preferred model:

  • Anthropic (Claude Opus, Claude Sonnet) — best for coding, complex reasoning
  • OpenAI (GPT-4o, GPT-4o-mini) — best for general tasks, function calling
  • Google (Gemini 3.5 Flash, Gemini 3 Pro) — best for long context, multimodal
  • OpenRouter — unified API for 200+ models, pay-per-token
  • Ollama — run Llama, Qwen, DeepSeek, Hermes 3 locally. Zero API cost, full privacy