Hermes Agent
A tiernew this weekA self-improving, model-agnostic CLI agent from Nous Research that now solves MCP's context-bloat problem with BM25-powered Tool Search.
Kai's verdict
Hermes Agent is the most technically serious open-source personal agent right now — the Tool Search MCP feature alone is a meaningful infrastructure contribution, not just a feature bump. If you're running multi-server MCP setups and tired of burning tokens on irrelevant schemas, this is the move. (Verdict pending Phi's full review.)
Strengths
- Tool Search for MCP uses BM25 to defer schema loading, cutting token overhead ~85% and boosting Claude Opus 4 accuracy from 49% to 74% on large tool catalogs
- Built-in learning loop that creates reusable skills from experience and builds persistent memory across sessions
- Fully model-agnostic — works with OpenRouter (200+ models), Anthropic, OpenAI, Ollama, or any custom endpoint with no code changes
- Runs anywhere: local terminal, $5 VPS, GPU cluster, or serverless; reachable via Telegram, Discord, Slack, WhatsApp, Signal, or Email
- MIT licensed with active community, curated MCP catalog, and agentskills.io compatibility for shareable, portable skills
Weaknesses
- CLI-only — no GUI; onboarding friction is real for non-technical users
- Tool Search only pays off with 15+ MCP tools; smaller models struggle to formulate effective search queries for it
- Self-hosted setup means you own the infra burden — debugging MCP server configs, API keys, and provider auth is on you
Best for
Power users and developers who want a deeply configurable, self-improving agent they can run on their own infra with full MCP ecosystem access and no vendor lock-in.
Pricing
Free (MIT open-source) + Nous Portal subscription for bundled models & tools
Core agent is fully open-source and self-hostable. Nous Portal bundles 300+ models, web search, image gen, TTS, and browser automation under one subscription — pricing not publicly listed per tier.