Skip to content

OpenClaw vs Hermes Agent vs Claude Code

Summary

Three widely used AI agents in 2026, each built around a different idea. Claude Code (Anthropic, proprietary) goes deep on codebases. OpenClaw (OpenClaw Foundation, MIT) is a self-hosted Gateway that puts one assistant in 32 messaging channels. Hermes Agent (Nous Research, MIT) runs a learning loop that writes and patches its own skills. All three now overlap: Claude Code has auto memory and chat channels, OpenClaw drafts skills through its Skill Workshop, and Hermes codes over ACP. The table below separates what is verified from community estimates.

Compared topics: OpenClaw · Hermes Agent · Claude Code (no topic in this vault; facts below come from code.claude.com, checked 2026-09-25).

  • Claude Code: "Make me indispensable to your codebase."
  • OpenClaw: "Become the automation layer of your life."
  • Hermes Agent: "Grow into whatever you need, and improve every time you use it."

High-Level Comparison

Dimension Claude Code OpenClaw Hermes Agent
Creator / steward Anthropic Peter Steinberger; OpenClaw Foundation (US 501(c)(3), July 2026) Nous Research
License Proprietary (npm: "SEE LICENSE IN README.md"; Anthropic commercial terms) MIT, no enterprise edition MIT
Latest version 2.1.283 (2026-09-25, npm latest); stable channel 2.1.274 (2026-09-16) 2026.9.6 (2026-09-23); extended-stable 2026.7.35 0.21.5, tag v2026.9.24 (2026-09-24)
GitHub stars 112k (2026-04, issue-tracker repo) ~390k (2026-09, per the Tools Catalogue) 210k (2026-07)
Primary focus Deep coding Universal personal assistant Self-improving agent
Surfaces / channels Terminal, VS Code (and Cursor), JetBrains, desktop app, web, mobile app, Slack; Channels (Telegram, Discord, iMessage) in research preview 32 channels (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, Matrix...) plus macOS/iOS/Android/Windows apps ~28 platforms (Telegram, Discord, Slack, WhatsApp, Signal, Teams, Email, SMS...) plus TUI, desktop app, API server
Runtime Native installer, Homebrew, WinGet, npm TypeScript on Node.js 24.16+ or 26.1+ Python 3.14 (plus Node for TUI/desktop)
Models Claude models via the Anthropic API, Amazon Bedrock, Google Cloud, or Microsoft Foundry Any provider via plugins; swappable runtimes (built-in, Codex, claude-cli, ACP) 40+ providers, including local endpoints
Self-learning Partial: auto memory records corrections and preferences Partial: memory "dreaming" and Skill Workshop drafts, both human-curated Yes: agent-created, self-patched skills plus curator
Memory CLAUDE.md (you write) + auto memory (Claude writes) USER.md + MEMORY.md + daily notes, memory_search Bounded MEMORY.md + USER.md, FTS5 session_search, optional external provider
Skills Agent Skills (SKILL.md), plugins, plugin marketplaces SKILL.md skills: bundled, ClawHub registry, Skill Workshop SKILL.md skills: agent-created, plus hubs (skills.sh, ClawHub, official)
Security record Permission modes, opt-in OS sandbox, managed settings. 31 GitHub security advisories published on the repo, 2025-06 to 2026-09 (advisories, checked 2026-09-27) High-severity CVEs (CVE-2026-25253 CVSS 8.8, CVE-2026-32922 CVSS 9.9); 647 advisories by 2026-08-27; 135k+ exposed instances (2026-02) At least 10 CVEs published April-September 2026 (auth, approval-guard, cross-user sessions, SSRF, path traversal)
Stability Not measured here No published measurement; 2026.9.x releases added restart and unfinished-work recovery No comparable figure; fast release churn
Usage limits Plan limits per 5-hour session (Pro, Max 5x, Max 20x) or API billing Your provider's limits (self-hosted) Your provider's limits (bring your own key or OAuth subscription)

Correction (2026-09-25)

Earlier versions of this page said Hermes Agent had "0 CVEs", listed an official "OpenClaw Cloud" plan at $59/month, and gave OpenClaw 24+ channels and one terminal backend. All four were wrong or stale: Hermes has at least ten published CVEs, the OpenClaw Foundation has no paid tier or hosted service, OpenClaw's catalog lists 32 channels, and it has five sandbox backends.

Architecture Comparison

The diagram shows each agent's main path from the user to the model and to where tools run, using each project's real component names.

flowchart TB
    subgraph CC["Claude Code"]
        CC_S["Terminal, IDE, desktop,<br/>web, Slack, Channels"] --> CC_L["Agent loop"]
        CC_L <--> CC_API["Claude API<br/>(Anthropic, Bedrock, Google Cloud, Foundry)"]
        CC_L --> CC_T["Tools: Read, Edit, Bash,<br/>subagents, MCP, hooks"]
        CC_T --> CC_SB["Permission modes +<br/>opt-in Bash sandbox"]
        CC_M["CLAUDE.md + auto memory"] -.-> CC_L
    end

    subgraph OC["OpenClaw"]
        OC_CH["32 channels + native apps"] --> OC_GW["Gateway<br/>127.0.0.1:18789"]
        OC_GW --> OC_RT["Agent runtime<br/>openclaw, codex, claude-cli, ACP"]
        OC_RT <--> OC_LLM["Model provider plugins"]
        OC_RT --> OC_EX["Tools on host or sandbox<br/>Docker, Podman, SSH, OpenShell, Crabbox"]
        OC_HUB["ClawHub skills + plugins"] -.-> OC_GW
        OC_MEM["USER.md, MEMORY.md,<br/>daily notes"] -.-> OC_RT
    end

    subgraph HA["Hermes Agent"]
        HA_CH["~28 platforms, TUI,<br/>desktop, API :8642"] --> HA_AG["AIAgent loop"]
        HA_AG <--> HA_LLM["40+ providers"]
        HA_AG --> HA_TB["7 terminal backends<br/>local, docker, ssh, modal,<br/>daytona, vercel_sandbox, singularity"]
        HA_AG -.->|"post-turn review"| HA_SK["skills/ + curator"]
        HA_AG <--> HA_MEM["MEMORY.md, USER.md,<br/>state.db FTS5"]
        HA_EVO["Offline DSPy + GEPA<br/>self-evolution repo"] -.->|"PR for review"| HA_SK
    end

Detailed Feature Matrix

Memory and Persistence

Feature Claude Code OpenClaw Hermes Agent
Long-term memory CLAUDE.md files (managed, user, project) written by you MEMORY.md (durable facts), loaded at session start MEMORY.md (2,200 chars), frozen snapshot per session
Agent-written memory Auto memory, on by default, per repository under ~/.claude/projects/<project>/memory/; first 200 lines or 25KB loaded Background "dreaming" sweep distills daily notes into MEMORY.md Background review after each turn saves memory
User model Auto memory stores preferences and corrections USER.md (stable preferences) USER.md (1,375 chars); optional Honcho dialectic modeling
Session search Resume earlier sessions memory_search over daily notes; conversation search (2026.8.1) session_search over all sessions (SQLite FTS5, no LLM calls, ~20 ms)

Autonomy and Self-Improvement

Feature Claude Code OpenClaw Hermes Agent
Skill creation Human-written skills and plugins Human-written; agents draft skills in the Skill Workshop for human approval Automatic via skill_manage, nudged every 15 tool-calling iterations and by the post-turn review
Skill improvement Manual Manual Patched in place during use; curator marks stale after 14 unused days, archives after 30
Prompt / skill optimization None None Offline DSPy + GEPA pipeline evolves SKILL.md files and opens a PR (Phase 1: skills only)
Review gate You edit files Workshop approval Optional write_approval

Deployment

Feature Claude Code OpenClaw Hermes Agent
Local install Native installer (macOS, Linux, WSL, Windows), Homebrew, WinGet, apt/dnf/apk Installer script, npm, Windows install.ps1 Installer script, install.ps1, desktop bundles (pip unsupported since v0.20.0)
Docker Dev containers Official images ghcr.io/openclaw/openclaw Official image nousresearch/hermes-agent
Hosted / cloud Claude Code on the web (cloud sessions), Routines No official hosted service; third-party hosting only; cloud workers for sessions Serverless backends: Modal, Daytona, Vercel Sandbox
Mobile Claude mobile app (web sessions, Remote Control) iOS and Android node apps Android/Termux (aarch64, Tier 2)
Sandbox Built-in Bash sandbox (Seatbelt on macOS, bubblewrap on Linux/WSL2), opt-in Off by default; backends Docker, Podman, SSH, OpenShell, Crabbox; NemoClaw for a full OpenShell deployment Container backends (Docker, Singularity, Modal, Daytona, Vercel Sandbox); dangerous-command approvals on local and ssh
Execution backends Local (plus cloud sessions) Host or 5 sandbox backends 7 terminal backends

Multi-Agent

Feature Claude Code OpenClaw Hermes Agent
Sub-agents Yes (subagents, background agents) Yes (sub-agents, Swarm tool, Code Mode) Yes (delegate_task, live subagent control since 0.21.0)
Multi-instance Parallel sessions (desktop, web, background agents) claworc (community) or per-tenant gateways Isolated profiles (hermes profile)
Orchestration Dynamic workflows, Agent SDK Lobster pipelines + Task Flow; community ClawTeam Multi-profile kanban, cron, hermes peer (A2A)
Hosting other agents n/a Runs Claude Code, Codex, Gemini CLI, OpenCode as runtimes (claude-cli, ACP) Imports from OpenClaw, Claude Code, Codex CLI (hermes claw migrate, hermes import-agent)

Benchmarks

Benchmarks measure models, not harnesses

SWE-bench and OSWorld score the underlying model, not the agent. Earlier versions of this page gave Claude Code "80.8% SWE-bench" and "72.5% OSWorld". Those were Claude Opus 4.6 model results from early 2026, since superseded by newer models. Neither OpenClaw nor Hermes Agent publishes harness benchmarks, and both can run the same Claude models. For current model scores, see Anthropic's models overview. Choose on workflow fit, not on a leaderboard.

Cost Comparison

Agent Entry cost Typical monthly Heavy use
Claude Code Pro $20/month ($17/month billed annually); the Free plan excludes Claude Code $20-100 (Pro or Max 5x) Max 20x at $200/month, or API pay-per-token
OpenClaw Free (MIT) + model costs Model/API costs only; no official estimate Model costs only; no official paid tier
Hermes Agent Free (MIT) + model costs ~$5-10 VPS + model/API costs Heavy API use, or a Nous Portal plan ($20/$100/$200 per month)

Plan prices: claude.com/pricing (checked 2026-09-25). OpenClaw and Hermes publish no usage-cost estimates; the VPS figure is the Hermes README's own minimum example, and Nous Portal plan prices come from the plans page (checked 2026-09-27).

Which One Should I Pick?

The flowchart gives a starting point; the lists below add the conditions that change the answer.

flowchart TD
    START{"Main job for the agent?"} -->|"Write, review, refactor code"| CC["Claude Code"]
    START -->|"Personal assistant in chat apps"| MULTI{"Many channels, devices,<br/>or scheduled automations?"}
    START -->|"Research or ops that should<br/>improve with use"| HA["Hermes Agent"]
    MULTI -->|"Yes, widest reach"| OC["OpenClaw"]
    MULTI -->|"A few platforms, learning matters more"| HA
    OC --> HARD{"Can you run a hardened Gateway<br/>(sandbox on, loopback, fast patching)?"}
    HARD -->|"No"| MANAGED["Use Claude Code Channels or a managed<br/>assistant, or NemoClaw/OpenShell"]
    HARD -->|"Yes"| OC_OK["OpenClaw with sandboxing<br/>and exec approvals on"]
    HA --> SHARED{"Several users talk to it?"}
    SHARED -->|"Yes"| HA_APPR["Hermes with allowlists<br/>and write_approval on"]
    SHARED -->|"No"| HA_OK["Hermes default setup"]
    CC --> CHAT{"Also need chat-app access?"}
    CHAT -->|"Yes"| COMBO["Claude Code + OpenClaw or Hermes,<br/>or Claude Code Channels"]

Choose Claude Code When

  • Your main need is writing, reading, and refactoring code
  • You work in the terminal, VS Code, JetBrains, the desktop app, or on the web
  • You want a vendor-supported tool with managed settings for teams
  • You accept a Claude subscription or API billing and Claude-only models

Choose OpenClaw When

  • You want one assistant across the widest set of messaging platforms (32 channels) and paired devices
  • Breadth of integration matters more than self-learning
  • You need workflow automation (Lobster pipelines, cron, webhooks)
  • You want the largest community skill registry (ClawHub), and you will vet skills before installing
  • You prefer self-hosting under a foundation-governed MIT project, and can keep up with frequent security releases

Choose Hermes Agent When

  • You want an agent that writes and maintains its own procedures over time
  • Long-term memory and cross-session recall are critical
  • You want the most execution backends (7, including serverless sandboxes that cost little when idle)
  • You are comfortable with a pre-1.0, fast-moving Python project, and will stay on the latest tag given its 2026 CVE record

Combine Them

Combination Use case
Claude Code + OpenClaw Coding plus personal automation; OpenClaw can drive Claude Code as a runtime
Claude Code + Hermes Development plus research workflows that learn
OpenClaw + Hermes Channel breadth plus a learning agent; each can import the other's memory

Timeline

Date Event
2025-02-24 Claude Code research preview
2025-05-22 Claude Code generally available (with Claude 4)
2025-11 Clawd/Clawdbot (future OpenClaw) first published
2026-01-27 Clawdbot renamed Moltbot (Anthropic trademark request)
2026-01-30 Renamed OpenClaw
2026-02 OpenClaw passes 100k GitHub stars about a week after its late-January launch wave (Panto's OpenClaw statistics dates it 2026-02-02)
2026-02 CVE-2026-25253 (CVSS 8.8), ClawHavoc malicious skills, 135k+ exposed instances; Steinberger joins OpenAI
2026-03-16 NVIDIA NemoClaw early preview (GTC 2026)
2026-03-18 to 21 OpenClaw CVE cluster (reported as nine CVEs), incl. CVE-2026-32922 (CVSS 9.9), fixed in 2026.3.11
2026-04-13 Hermes Agent 0.9.0
2026-04 to 09 At least ten CVEs published against Hermes Agent
2026-07 OpenClaw Foundation launches as a US 501(c)(3)
2026-09-23 OpenClaw 2026.9.6
2026-09-24 Hermes Agent 0.21.5
2026-09-25 Claude Code 2.1.283

Sources