ori z blog
rss

General 5 min read

Hermes Agent vs OpenClaw vs Claude Code vs OpenCode: pick the right tool (2026)

Part 2 of 2 — The full comparison table and decision guide, built from verified primary sources. Stars, MCP, skills, Windows, free tier, and what each tool is actually for.

By

This is Part 2. Part 1 explains what Hermes Agent and OpenClaw actually are — read that first if you haven't.

The research is done. Here is the honest comparison, built from primary sources (GitHub API, npm registry, official READMEs), verified by adversarial research agents.

The four tools, side by side

Claude Code

  • Repo: anthropics/claude-code — 137K stars
  • Version: v2.1.202 — published 2026-07-06 by Anthropic
  • Install: npm install -g @anthropic-ai/claude-code
  • What it does: Edits your codebase, runs tests, makes commits, opens PRs — all from the terminal via a TUI session
  • Auth: Anthropic-shape API (ANTHROPIC_BASE_URL) — works with corp Bedrock, direct Anthropic key, or any Anthropic-compatible proxy
  • MCP: ✅ native .mcp.json workspace + ~/.claude.json user config
  • Skills: ✅ ~/.claude/skills/ via npx skills add ... -a claude-code
  • Windows: ✅ native npm binary
  • Free tier: ✅ with an API key (corp Bedrock = $0 if your employer covers it)

OpenCode

  • Repo: anomalyco/opencode — 183K stars
  • Version: opencode-ai v1.17.14 (by Dax Raad / SST team)
  • Install: npm install -g opencode-ai
  • What it does: Same as Claude Code — edits files, runs commands, makes commits
  • Auth: OpenAI-compatible endpoint — works with Groq, Cerebras, OmniRoute, Ollama, or any OpenAI-shape proxy
  • MCP: ✅ native opencode.jsonc top-level mcp key
  • Skills: ✅ ~/.config/opencode/skills/ via same npx skills CLI
  • Windows: ✅ ships opencode-windows-x64 binary
  • Free tier: ✅ Groq/Cerebras/OmniRoute all have free tiers

Hermes Agent

  • Repo: NousResearch/hermes-agent — 210K stars
  • Last pushed: 2026-07-07
  • Install: PowerShell installer (Windows); hermes-cli-mcp npm package for MCP bridge
  • What it does: General-purpose terminal agent — coding is one capability alongside persistent memory, messaging integrations (Telegram/Discord/Slack/ WhatsApp/Signal/Email), and a learning loop
  • MCP: ✅ documented integration
  • Skills: ✅ agentskills.io compatible, /skills command
  • Memory: ✅ Honcho dialectic user modelling + FTS5 session search
  • Windows: ✅ PowerShell installer
  • Free tier: ✅ BYOK

OpenClaw

  • Repo: openclaw/openclaw
  • npm: openclaw v2026.6.11
  • What it does: Routes AI through 20+ messaging platforms (WhatsApp, Slack, Telegram, Discord, LINE, Matrix, more). Has a skills/memory system and agent loop, but messaging is its primary architecture.
  • MCP: ⚠️ not prominently documented
  • Skills: ✅ own system (clawhub CLI — separate from npx skills)
  • Memory: ✅ MEMORY.md, USER.md, SOUL.md
  • Windows: ✅ Node.js binary
  • Free tier: ✅ BYOK
  • Note: Hermes Agent can migrate from ~/.openclaw with hermes claw migrate

Full comparison table

Claude CodeOpenCodeHermes AgentOpenClaw
Stars137K183K210Kmoderate
Last active2026-07-062026-072026-07-072026-06
Primary jobCode editingCode editingGeneral agentMessaging gateway
Code editing✅ dedicated✅ dedicated✅ secondary❌
Messaging❌❌✅ 6 platforms✅ 20+ platforms
MCP✅ native✅ native✅⚠️
Skills✅ npx skills✅ npx skills✅ agentskills.io✅ clawhub
Memory✅ auto-memory✅✅ Honcho + FTS5✅
Windows✅ npm✅ binary✅ PowerShell✅ Node
Free tier✅ (needs key)✅ free providers✅ BYOK✅ BYOK
Anthropic API✅ native⚠️ adapter⚠️⚠️
OpenAI-compat API⚠️ proxy✅ native✅✅
npx skills compat✅✅⚠️ agentskills.io⚠️ clawhub
AionUi listed✅✅❌✅
cc-switch listed✅✅✅✅

Decision guide

Use Claude Code when: You have corp Bedrock or an Anthropic API key and want the most polished terminal TUI with the deepest Claude/tool-use integration. The skills and MCP ecosystem is mature. This is the safe default for anyone on an Anthropic model.

Use OpenCode when: You want free-tier providers (Groq, Cerebras, OmniRoute) or you want to run Hermes 3 / Llama / Gemma locally via Ollama. Same npx skills ecosystem as Claude Code. Best cost/coverage for daily driving without a paid key.

Use Hermes Agent when: You want a single agent that handles Telegram/Discord/Slack messaging, has a persistent memory system, and covers coding as one capability. 210K stars is the largest of the four — strong community signal. Worth evaluating as a third agent if messaging + memory is part of your workflow.

Use OpenClaw when: You want your AI assistant in WhatsApp or Telegram on mobile, BYOK, no subscription. Genuine agent infrastructure (Hermes Agent migrates from it). Not optimised for terminal code editing — use Claude Code or OpenCode for that.

My current stack (after this research)

Claude Code  →  hai :6655  →  corp Bedrock       (primary, coding)
OpenCode     →  OmniRoute :20128  →  free providers  (failover, coding)

Hermes Agent is now on my radar — 210K stars with MCP + skills puts it in the same league. Will evaluate it as a messaging + memory layer on top of the coding pair. The hermes claw migrate path from OpenClaw is a thoughtful design decision that suggests a mature team.

OpenClaw serves a different need (mobile AI chat) and is well-covered by the Hermes migration path if you ever want to upgrade.

Part 1: Hermes Agent and OpenClaw — what they actually are

Read it faster

Comments

Comments are powered by giscus. Set PUBLIC_GISCUS_REPO_ID and PUBLIC_GISCUS_CATEGORY_ID in your environment to enable them.