Skip to content

Hermes Agent (Nous Research)

Overview

Hermes Agent is an open-source, self-improving personal AI agent developed by Nous Research, launched in February 2026. It is the direct successor to OpenClaw and positions itself as the only agent with a built-in learning loop — creating skills from experience, improving them during use, and building a persistent model of the user across sessions. Unlike stateless coding agents, Hermes is designed as a long-lived personal assistant that runs unattended, operates across messaging platforms, and accumulates knowledge over time.

The Hermes project is built on the broader Nous Hermes model family — a series of open-source, fine-tuned LLMs (Hermes-2-Pro, Hermes-3) specifically optimized for function calling, structured output, and agentic tool use.

Key Capabilities

Self-Improvement and Learning Loop

Hermes is differentiated by autonomous, closed-loop learning:

  • Skill creation: After completing complex tasks, the agent automatically creates reusable skills from the interaction
  • Skill improvement: Skills are refined during future use, not just stored statically
  • Cross-session memory: FTS5 full-text session search with LLM summarization allows recall of past conversations
  • User modeling: Integrates Honcho for dialectic user modeling — building a persistent model of user preferences and working style across sessions

Tool Use and MCP Integration

  • 40+ built-in tools covering file operations, web search, code execution, scheduling, and data retrieval
  • MCP (Model Context Protocol) server integration for extended tool capabilities
  • Compatible with the agentskills.io open standard for skill sharing
  • Modular toolset system — add custom tools via the @tool decorator pattern

Multi-Platform Access

Interface Details
Terminal UI Multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, streaming tool output
Messaging gateways Telegram, Discord, Slack, WhatsApp, Signal, Email
Voice Voice memo transcription capabilities
Seven sandbox backends Local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox

Scheduling and Automation

Built-in cron scheduler for unattended tasks defined in natural language: daily reports, nightly backups, weekly audits. Tasks deploy to any connected platform via the messaging gateway.

Parallel Execution

Spawn isolated subagents for concurrent workstreams. RPC scripting collapses multi-step pipelines into zero-context-cost turns via Python scripts calling tools over RPC.

Hermes Model Family (Nous Research)

The agent is built on a series of OSS models fine-tuned for agentic use cases:

Model Base Key Capability
Hermes-2-Pro-Llama-3-8B Llama 3 8B Function calling via <tool_call> / <tool_response> XML tags; JSON structured output; ChatML format
Hermes-3 Llama 3.1 (8B, 70B, 405B) Goal Oriented Action Planning (GOAP) via <scratch_pad> tags; enhanced multi-turn tool use

Function Calling Architecture (Hermes-2-Pro / Hermes-3)

The Hermes models use a structured ChatML prompt format with a distinctive XML-tag tool-call protocol:

<tool_call>
{"name": "function_name", "arguments": {"param": "value"}}
</tool_call>

The Hermes-3 GOAP template adds a reasoning scratch pad:

<scratch_pad>
Goal: [restated user intent]
Actions: [planned function calls]
Observation: [summarized tool results]
Reflection: [evaluating relevance and task progress]
</scratch_pad>

Models support: - Recursive function calling (configurable depth, default 5) - JSON mode / structured outputs via Pydantic schema enforcement - 4-bit quantization for memory-efficient deployment - Few-shot example injection for domain specialisation

Architecture

Component Details
Primary language Python (88%), TypeScript (8.8%)
Installation Single-command scripts for Linux, macOS, WSL2, Windows (Windows in early beta)
Deployment $5 VPS up to enterprise GPU clusters
Model providers Nous Portal, OpenRouter (200+ models), NovitaAI, NVIDIA NIM, OpenAI, Anthropic, custom endpoints
Memory backend Persistent storage + FTS5 full-text search + Honcho user modeling
Skill storage Git-compatible, versioned skill library
Dimension Hermes Agent OpenClaw Claude Code OpenCode
Primary use case Hyper-personal assistant with learning loop Hyper-personal assistant Coding agent Coding agent
Self-improvement ✅ Automatic skill creation + refinement ❌ ❌ ❌
Cross-session memory ✅ Persistent + user modeling ✅ Persistent ❌ ❌
MCP support ✅ ❌ ✅ ✅
Messaging gateways ✅ (Telegram, Discord, Slack, WhatsApp) ✅ (WhatsApp, iMessage) ❌ ❌
Model flexibility Any provider via OpenRouter etc. Limited Anthropic only Any provider
Open source ✅ ✅ ❌ ✅
Security profile Permission-hungry (same trifecta risk as OpenClaw) Permission-hungry Sandboxed Sandboxed

Security Considerations

Hermes Agent inherits the same toxic flow trifecta risk as OpenClaw: private data access + untrusted content handling + external action capability. The more tools and integrations enabled, the more useful — and the more concentrated the permission surface. Key mitigations:

  • Run in a sandboxed Docker or Daytona environment to restrict file system and network access
  • Apply least-privilege tool selection — enable only the tools needed
  • Use the messaging gateway with a dedicated account rather than personal accounts
  • Review skills before enabling — supply chain risk from third-party skill imports
  • Migrate from OpenClaw using the built-in migration tooling (preserves settings, memories, skills)

See Toxic flow analysis for AI and Production Best Practices — Security for broader guidance.

Community and Ecosystem

The Hermes Atlas (hermesatlas.com) community directory catalogs 80+ quality-filtered repositories across 12 categories: skills, plugins, integrations, deployment templates, and forks. The ecosystem expanded rapidly after launch in February 2026.

Hermes Function Calling reference repository: NousResearch/hermes-function-calling — 1.4K stars, 187 forks (illustrates model adoption independent of the agent itself).

Suitable For

  • Teams building personal AI assistants that need persistent cross-session context and self-improvement
  • Engineers who want an OSS personal assistant on self-hosted or low-cost infrastructure
  • Developers experimenting with the GOAP reasoning pattern or Hermes model fine-tuning for tool use
  • Researchers evaluating hyper-personal assistant architectures (with careful security scoping)

Limitations

  • Hyper-personal assistant pattern requires broad permissions — inherent security tension
  • Windows support is still in early beta
  • Ecosystem is rapidly evolving; skill quality and supply chain risk require active curation
  • GOAP reasoning (Hermes-3) and skill self-improvement are newer features with limited long-term production case studies

See Also

References