Independent register of AI agents & harnesses · no sponsored placementsEdition 2026-09 · 107 entries · evidence to 26 Sept 2026

Compare / Hermes Agent vs ZeroClaw

Hermes Agent vs ZeroClaw

Hermes Agent rates 69 against ZeroClaw's 61 in this edition. The table lines up every dimension we record; the leading value in each numeric row is marked. Add a third or fourth entry.

Dimension
Hermes AgentNous Research
ZeroClawZeroClaw Labs
Rating
AgentsWisdom score6961
Confidencehighmedium
Rank in categoryNo. 1 of 3No. 3 of 3
Pillars
Adoption & momentum7971
Experts (independent reviews)6467
Crowd (community sample)6547
Community posts sampled219
Experts vs crowdCrowd warmer by 1Crowd cooler by 20
Adoption
GitHub stars249k33k
npm weekly downloadsn/en/e
PyPI monthly downloads139kn/e
VS Code installsn/en/e
Reported usersn/en/e
Reported customer organisationsn/en/e
Releases, last 90 days183
Benchmarks (shown, not scored)
Terminal-Bench 2.0not listednot listed
SWE-benchnot listednot listed
Independent reviews83
Trust & safety
Overall gradeBB
Permission modelBA
Data access scopeBA
Data storageAA
Data retention & trainingn/en/e
Incident historyn/eC
ComplianceDD
TransparencyAA
Price
PriceFreeFree
Plans
  • Open source (self-hosted): Free
  • Hermes Cloud (hosted instances via Nous Portal): Custom
  • Open source (self-hosted): Free
Product
CategoryPersonal & desktop agentsPersonal & desktop agents
MakerNous ResearchZeroClaw Labs
LicenceMITMIT OR Apache-2.0
Use it viaCLI, Desktop app, Chat apps, MCP server, APICLI, Chat apps, Web app
Runs onmacos, linux, windowslinux, macos, windows
Models it drivesNous Portal models, OpenRouter, GPT, Claude (Anthropic-compatible endpoints), local via llama.cpp/vLLM/Ollama, any OpenAI-compatibleClaude, GPT, local via Ollama / LM Studio / llama.cpp / vLLM, any OpenAI-compatible
Key features
  • Persistent memory with full-text search across past sessions
  • Autonomous skill creation and refinement after complex tasks
  • Messaging gateway for 20+ channels (Telegram, Discord, Slack, WhatsApp, Signal, email, Matrix)
  • Natural-language cron scheduling and subagent delegation
  • Seven terminal backends: local, Docker, SSH, Modal, Daytona, Vercel Sandbox, Singularity
  • Dangerous-command approval modes (smart/manual/off) plus a hardline blocklist
  • Desktop app with built-in Bot Mode for multi-agent group chats (v0.21.0)
  • Single static Rust binary; targets low-memory hardware such as Raspberry Pi
  • 30+ channels including Telegram, Discord, WhatsApp, Slack, Signal, Matrix and email
  • Default 'supervised' autonomy: low-risk runs, medium-risk asks, high-risk blocked
  • Auto-detected OS sandbox (Landlock/Bubblewrap/Firejail, Seatbelt, AppContainer, Docker)
  • Workspace-only file and shell boundaries with forbidden paths
  • HMAC tool receipts and an emergency stop command
  • Gateway with web dashboard for chat, memory, config and cron
  • SOP engine, WebAssembly plugin host, A2A discovery, ZeroRelay/ZeroRouter (v0.8.x)
Verdicts
Praised for
  • Users report it is more reliable than OpenClaw across updates (Reddit comparison threads)
  • Practical everyday automations such as price watching, business task tracking and homelab maintenance (HN)
  • Works with local and hosted models; widely covered by large YouTube creators
  • Small resource footprint compared with OpenClaw (independent reviews, community reports)
  • Security-by-default posture: supervised autonomy, sandboxing and workspace scoping
  • Some users report it as stable and easy to run on a small VPS
Criticised for
  • Setup and secure operation require technical judgment (The Rundown AI)
  • Some users find the UI/UX slow or unclear, and report slow task completion versus OpenClaw
  • Rapid AI-assisted development cadence with dozens of third-party CVE reports in 2026
  • Several users report bugs, tool-calling failures and freezes, especially with local models
  • Smaller plugin/integration ecosystem than OpenClaw; Rust makes extension harder for some
  • Limited visibility into running tools in early versions (HN)
  • Impersonation sites and forks caused confusion; the official repo was briefly unavailable (404) in late March 2026
Evidence
Evidence collected26 Sept 202626 Sept 2026
Sources cited44