Hermes vs OpenClaw: Pick a Path (Not a Fandom)

Hermes is an agent harness, not a web host or git deployer. OpenClaw is the same shape, not a privacy-first or local-only system. If you also need a site, see the Cloudflare Tunnel setup guide.

The problem

“Which agent framework should I use?” gets answered with fanboy threads. That’s useless when your real constraint is setup time, where it reaches you, or whether you can debug it at 2am. This is a decision aid, not a verdict. We run Hermes daily. For OpenClaw we rely on canonical public docs and community reports and say so.

The rule / frame

Pick on friction and fit, not feature count. The right question is “which one gets me a capped, observable run with the least new concepts?” — not “which has more stars.” Framework choice is premature until you’ve closed one loop. If you haven’t run any agent end-to-end yet, start with the Start Here hub and come back.

What each is, concretely

Hermes Agent (tested by ABS)

Hermes is a self-hosted agent you drive from a chat surface (CLI, Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, and 20+ other platforms per the canonical docs). It ships with a Skills System, 60+ built-in Tools, MCP integration, an OpenAI-compatible model-provider client (works with Nous Portal, OpenRouter, OpenAI, or any OpenAI-compatible endpoint), a closed learning loop with persistent memory, and a static-site deploy hook (this site runs on it). Maintained by Nous Research (not by ABS). We run it daily on Linux. MIT-licensed (© 2025 Nous Research).

OpenClaw (inferred — we have not run it)

OpenClaw is a self-hosted personal-AI assistant daemon built around a single always-on Gateway process (Node, default port 18789, loopback bind by default). It works with hosted and local model providers; onboarding prompts for an Anthropic/OpenAI/Google API key. It ships 25+ documented chat channels (Discord, Feishu, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp, Zalo, plus China/SEA channels), a SKILL.md skill system, a plugin/ClawHub extensibility layer, Markdown workspace memory with optional backends and “Dreaming” consolidation, per-channel pairing/allowlist policies, and security audit / doctor commands. One trust boundary per gateway — not multi-tenant. Native apps ship for macOS, iOS, Android, and Linux (verified via the repo’s apps/ tree). MIT-licensed (© 2026 OpenClaw Foundation).

Side-by-side (where both are sourced)

Dimension	Hermes Agent	OpenClaw	Evidence class
Self-hosted install	curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash; PowerShell equivalent.	curl -fsSL https://openclaw.ai/install.sh | bash; PowerShell equivalent.	[CONFIRMED] both
First-success	hermes --version → hermes chat -q "Reply with OK."	openclaw onboard --install-daemon → openclaw gateway status shows port 18789 → openclaw dashboard opens Control UI.	[CONFIRMED] both (different surfaces)
Maintainer	Nous Research (per the docs).	OpenClaw Foundation.	[CONFIRMED] both
Runtime	Python-based; Hermes Desktop for macOS/Windows; CLI for Linux/WSL2/Android Termux.	Node 22.22.3+ / 24.15+ / 25.9+ (26 recommended).	[CONFIRMED] both
Model providers	Nous Portal (one OAuth, bundles web/image/TTS/browser), OpenRouter, OpenAI, “any endpoint.”	27+ hosted providers (Alibaba, Amazon Bedrock, Anthropic, Cohere, OpenAI, Google, Mistral, Moonshot, OpenRouter, xAI, Z.AI, etc.) + OpenAI-compatible proxy for Ollama/vLLM/LM Studio/llama.cpp.	[CONFIRMED] both
Local-model path	Hermes is the client of an OpenAI-compatible local endpoint (Ollama, vLLM, LM Studio, llama.cpp). It does not itself expose an OpenAI-compatible local-model server.	OpenClaw can be configured for local inference through OpenAI-compatible proxies; hosted providers are the default onboarding.	[CONFIRMED] both — neither is “local-first”
Messaging channels	”20+ platforms”: CLI, Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Home Assistant, Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, Weixin, BlueBubbles (iMessage), QQ, Yuanbao, Microsoft Teams, LINE, ntfy, SimpleX, Buzz, IRC, Raft, Photon.	25+ documented: bundled/core (Telegram, WebChat, Reef), official plugins (Discord, Feishu, iMessage, Matrix, MS Teams, Signal, Slack, SMS, Synology Chat, Twitch, Tlon, Nostr, etc.), external (WeChat, WeCom, Yuanbao, Zalo ClawBot).	[CONFIRMED] both — comparable breadth; OpenClaw’s list leans China/SEA more; Hermes has phone-centric extras (BlueBubbles, Photon).
Memory model	Bounded, curated memory: MEMORY.md (2,200 chars) + USER.md (1,375 chars) in ~/.hermes/memories/, injected as a frozen snapshot at session start; FTS5 session search; optional external providers; Honcho dialectic user modeling.	Markdown on disk: USER.md, MEMORY.md, memory/YYYY-MM-DD.md, DREAMS.md; optional SQLite/QMD/Honcho/LanceDB backends; “Dreaming” background consolidation. Can import memory from Codex, Claude Code, and Hermes.	[CONFIRMED] both; OpenClaw’s budgets are filesystem-backed; Hermes enforces strict char caps to preserve LLM prefix cache.
Extensibility	Skills (SKILL.md in ~/.hermes/skills/) compatible with agentskills.io open standard; Skills Hub; /learn to generate skills; external skill directories.	Skills (SKILL.md loaded into prompt) + Plugins (packaged runtime extensions, tools/skills/channels/providers/hooks); ClawHub marketplace + npm/git/local/archives.	[CONFIRMED] both
Voice	Native real-time voice in CLI, Telegram, Discord, Discord VC.	Voice Call is an official plugin (Plivo / Telnyx / Twilio).	[CONFIRMED] both — Hermes ships voice more natively; OpenClaw pushes voice to a plugin.
Automation	Built-in cron with delivery to any platform; gateway ticks every 60s.	Automations (precise cron), Heartbeat (every 30 min default), Background Tasks ledger, Task Flow, Hooks, Standing Orders.	[CONFIRMED] both
DM security default	DM pairing available.	dmPolicy: pairing default.	[CONFIRMED] both
License	MIT — © 2025 Nous Research. (LICENSE file retrieved 2026-08-07.)	MIT — © 2026 OpenClaw Foundation. (LICENSE file retrieved 2026-08-07.)	[CONFIRMED] both
Maintainer maintenance signal	Active. ~2–3 week major-release cadence; latest v2026.8.3 / Hermes Agent v0.20.0 on 2026-08-03 (Nous Research).	Active. Latest v2026.7.1-2 on 2026-08-04; commits within the hour.	[CONFIRMED] both — apples-to-apples.
Stars / forks / sponsors	Not used as a proof point. Volatile, not advertised in the README.	Not used as a proof point. 385k+ stars and 2,876 contributors exist but are not safe copy.	[REJECTED as evidence]
Volatile specifics	”Latest release tag” is the safer active-maintenance signal.	Same.	[REJECTED as copy]
“OpenAI-compatible server”	Hermes exposes an OpenAI-compatible API Server as a documented surface; Hermes is also a client of any OpenAI-compatible endpoint.	Gateway serves /v1/models, /v1/chat/completions, /v1/responses, /v1/embeddings.	[CONFIRMED] both expose it; both are also clients of others.
Hosting websites	Not a feature — use Cloudflare Tunnel.	Not a feature.	[CONFIRMED] both
Multi-tenant	Not a documented feature.	Explicitly not supported (“mutually untrusted users on one Gateway” forbidden).	[CONFIRMED] OpenClaw; [INFERENCE for Hermes].
Cost cap (built-in)	ABS operator guide provides the pattern. We have not confirmed Hermes ships a built-in ceiling.	Provider plugins own usage reporting; no documented built-in monthly spend cap.	[INFERENCE both]
OpenAI-compatible local endpoint (server-side)	Not a documented feature.	Not a documented feature.	[CONFIRMED] neither.
MCP integration	First-class: documented in the Hermes docs landing as a built-in surface.	First-class: mcp.servers.* config, openclaw mcp add/doctor CLI, Streamable HTTP / SSE / Stdio transports, tool filters via toolFilter.include/exclude. Source: docs.openclaw.ai/tools/mcp.	[CONFIRMED] both
OTel / observability surface	ABS 1-line observability hook is an ABS operating pattern. Hermes’s own native OTel export surface was not retrieved in this pass.	First-class plugin @openclaw/diagnostics-otel over OTLP/HTTP; metrics, traces, and structured logs; default sample rate 0.2. Source: docs.openclaw.ai/gateway/opentelemetry.	[CONFIRMED for OpenClaw]; [INFERENCE for Hermes] — verify before publishing an OTel parity claim.

Choose Hermes if…

• You want a working chat-driven agent tonight with minimal new concepts.
• You value a cost cap and an observability hook as first-class patterns, even if you wire them yourself using our guides.
• You want a closed learning loop named as the headline feature — autonomous skill creation, skill self-improvement, FTS5 session search, Honcho dialectic user modeling.
• You want memory with hard caps (MEMORY.md 2,200 chars; USER.md 1,375 chars) that protect system-prompt budget and preserve LLM prefix cache.
• You want native real-time voice in CLI/Telegram/Discord without a telephony plugin.
• You want first-class MCP integration as a documented, official surface.
• You want six terminal backends including serverless (Daytona, Modal) that hibernate when idle.
• You want one OAuth (Nous Portal) to bundle model + web search + image gen + TTS + browser.
• You want agentskills.io-compatible skills (open standard) rather than a project-specific format.
• You want this site (and our guides) to integrate with your stack.

Choose OpenClaw if…

• You want the widest single-install channel catalog — especially China/SEA (Feishu, WeChat via external plugin, WeCom, QQ bot, Yuanbao, Zalo, Zalo personal, Tlon, LINE, plus iMessage via the bundled imsg plugin).
• You want a mature plugin ecosystem with a separate MIT-licensed registry — ClawHub (https://clawhub.ai, repo openclaw/clawhub) — not just a skill system.
• You want per-channel and per-agent sandboxing primitives with explicit mode: off | non-main | all semantics.
• You want to import existing memory from Codex, Claude Code, or Hermes into a single home.
• You want the OpenAI-compatible HTTP API (/v1/models, /v1/chat/completions, /v1/responses, /v1/embeddings) on the same Gateway port for direct integration with Open WebUI, LobeChat, LibreChat.
• You want memory that is filesystem-backed and budget-truncated only, not char-capped at session start.
• You want native macOS, iOS, Android, and Linux clients that ship with the project (the OpenClaw apps/ repo tree confirms this; macOS has a separate MLX TTS companion).

Choose neither / use a managed service if…

• You do not want to operate a long-running daemon on a host you own. Both Hermes and OpenClaw are self-hosted personal-assistant products; neither surfaces a hosted chat in the canonical docs.
• You want zero CLI / config / token rotation. Both projects assume you will run a Gateway or messaging service and manage keys.
• You want a single multi-tenant shared workspace for an adversarial team. OpenClaw docs explicitly forbid that shape.
• You need FedRAMP / SOC2 / HIPAA out of the box. Neither project is presented as compliance-certified in the docs we reviewed.

What this comparison does NOT decide

• Hosted vs self-hosted vs on-device. All three are choices you make on top of the framework.
• Provider choice (Nous Portal, OpenAI, Anthropic, OpenRouter, local). Both frameworks accept a wide range.
• Single user vs small team. Both frameworks are documented as personal-assistant products; small-team usage is not a documented primary use case for either.
• Cost economics. Both charge for the underlying model calls; neither ships a built-in monthly budget.
• OpenClaw vs Hermes vs in-house. If neither is the right shape, a managed product (ChatGPT, Claude for Teams, Notion AI, or a SaaS workflow tool) may be.

Related guides (internal links only)

• Start Here: Local AI and Agents Without Wasting a Weekend
• Hermes Agent setup v2
• How to find, download, evaluate local models
• Why every agent needs a cost cap on day one
• The 1-line observability hook that fits any agent
• The three-phase context loop
• The four tiers of agent memory
• Cost-aware model routing for agents
• Cloudflare Tunnel setup for a static site

Sources / what we ran

• Hermes claims are tested at ABS for the listed product capabilities on Linux, and documentation-verified against hermes-agent.nousresearch.com/docs for the rest. Memory tier names and char caps match the live ABS guides. ABS does not maintain Hermes; Nous Research does. License: MIT (LICENSE file retrieved 2026-08-07). Recent releases: v2026.8.3 / Hermes Agent v0.20.0 (2026-08-03).
• OpenClaw claims are documentation-verified only; ABS has not run OpenClaw. The OpenClaw canonical research dossier collects every source URL and the full evidence ledger (29+ distinct primary URLs, retrieved 2026-08-07). License: MIT (LICENSE file retrieved 2026-08-07). Latest release: v2026.7.1-2 (2026-08-04). ClawHub registry: openclaw/clawhub, MIT. Volatile repo metrics (stars, forks, contributors) are deliberately not used.
• The ## Sources frontmatter array above is the canonical link map for the article. Re-fetch the canonical pages at publication time because channel taxonomy, model rosters, and provider lists move quickly.