CompareChoosing an agent
Hermes agent vs OpenClaw
Two open-source personal agents that answer you in chat apps, remember things and run jobs on a schedule. What each project documents about itself, side by side.
Short answer: Both are free, MIT-licensed personal agents you can host yourself, and both fit the same small always-on machine: from $10.60 a month on Agent 37. The AI model adds about $14.70 a month on top (GLM 5.3, 30 requests a day at 40k tokens each, median zero-retention price; from $6.84 at the cheapest host). The documented differences: Hermes comes from Nous Research, which also sells a hosted Hermes Cloud and a model subscription; its memory is deliberately capped at a few thousand characters. OpenClaw is run by a non-profit foundation with no paid tier or hosted service, keeps memory as ordinary files, releases far more often, and leaves tool sandboxing off by default.
Side by side
Numbers in brackets point to the quoted sources below. Star counts and release counts are a snapshot from GitHub on the date checked.
| What | Hermes agent | OpenClaw |
|---|---|---|
| Maker and licence | Nous Research builds Hermes Agent; the LICENSE file is MIT, copyright Nous Research. [1][2] | OpenClaw is stewarded by the OpenClaw Foundation, an independent 501(c)(3); the LICENSE file is MIT, copyright OpenClaw Foundation. [3][4] |
| Chat apps | The platform table lists 28 entries, including Telegram, Discord, Slack, WhatsApp, Signal, SMS, email, iMessage through BlueBubbles or Photon, Microsoft Teams, Matrix, Google Chat, IRC and SimpleX. [5] | The channel catalogue also lists Feishu, IRC, LINE, Matrix, Microsoft Teams, Nextcloud Talk, Nostr, SMS, WeChat, WeCom, X, Zalo and others. [6] |
| AI models and bring-your-own key | Any provider works: Nous Portal (OAuth subscription), OpenRouter, Anthropic, OpenAI/Codex, Google Vertex, xAI, Qwen, MiniMax, Ollama and other local OpenAI-compatible servers. Nous Portal is the recommended route: one OAuth login covers 300+ models plus the Tool Gateway. Bring-your-own keys remain supported per provider. Keys are optional per tool; the README says you can bring your own keys per tool. [7][8][9] | Bundled provider plugins cover OpenAI, Anthropic, Google Gemini, Z.AI, OpenRouter, Amazon Bedrock, xAI, NVIDIA, MiniMax and others; Ollama, vLLM, LM Studio and llama.cpp cover local models. Keys are stored in the auth store or supplied through the Gateway environment; there is no OpenClaw account in the loop. [10][11] |
| Memory | Two files under ~/.hermes/memories: MEMORY.md (2,200 characters, about 800 tokens) and USER.md (1,375 characters, about 500 tokens), injected into the system prompt at session start. Memory is deliberately capped and does not grow without bound; the agent must consolidate or delete entries when full. [12][13] | Memory is plain files on disk; the docs say the limit is your storage, not the model. A pre-compaction flush reminds the model to write durable notes first. Session context is still bounded by the model window, so long conversations compact or truncate. [14][15] |
| Skills and plugins | Skills live in ~/.hermes/skills, follow the agentskills.io standard, and the agent can create, modify or delete them. Bundled, hub-installed and agent-created skills share one directory; MCP servers are the separate plugin route for external tools. [16][17] | Skills are SKILL.md folders with YAML frontmatter, loaded from the workspace or ~/.openclaw/skills, with bundled skills and ClawHub community installs. ClawHub is the community skill registry; plugins are separate packages with their own install and enable commands. [18][19] |
| Scheduled tasks | One cronjob_manage tool creates, pauses, edits and removes jobs; jobs can run headless scripts with no model call, deliver to a chat or file, and fire on webhooks. The messaging gateway process owns the scheduler, so scheduled jobs need the gateway running. [20][21] | Jobs persist, wake the agent on a schedule, and deliver to a chat channel, a webhook or nowhere; openclaw automations is the CLI, with openclaw cron as an alias. [22] |
| Browser and computer use | browser_exec drives a Hermes-packaged Chromium by default; cloud browsers (Nous, Browserbase, Browser Use, Firecrawl) and an attached personal Chrome are alternatives. Computer use runs through the cua-driver MCP server and works with any tool-capable model; it acts on the machine the bot runs on. [23][24] | The managed profile is isolated from the personal browser; an optional user profile attaches to the real signed-in Chrome through Chrome DevTools MCP. macOS uses Peekaboo by default or CUA; Windows and Linux use the experimental cua-computer plugin; a vision-capable model is required. [25][26] |
| Multi-user and teams | Access is granted by per-platform user allowlists, Discord role allowlists, a global allowlist or DM pairing codes. Access control is opt-in: an unconfigured gateway refuses everyone, so the owner must add the people who may use the bot. Several people can talk to one bot in a shared channel, each with a separate session. Bot Mode adds several named agents, each with its own memory and credentials. [27][28][29] | Sessions carry an immutable creator, an assignable owner and participant history, with live presence in the Control UI. Ownership and presence are usability features, not security boundaries; adversarial users need separate gateways. [30][31] |
| Install methods | One-line install.sh for Linux/macOS/WSL2, install.ps1 for Windows, desktop installers (DMG, MSIX), Docker image, Nix flake and modules, Termux APT package for aarch64 Android. The Docker image is stateless and upgraded by pulling a new tag; data lives in the mounted volume. [32][33] | One-line install.sh or install.ps1, npm install -g openclaw@latest, Docker, Nix, and desktop companion apps for macOS and Windows; the installer provisions Node when it is missing. The npm package itself needs Node 24.22.1+ or 26.12.1+, with Node 26 recommended. [34][35][36] |
| Supported operating systems | Tier 1: macOS on Apple Silicon, Windows 10/11 (x86_64 and aarch64), Linux/WSL2 (x86_64 and aarch64) and Docker. Tier 2: Nix and Termux on aarch64. Intel macOS is not a supported platform for the desktop installer; Intel bundles exist only in releases cut with --with macos-intel. [37][38] | Windows has a native Hub app, a PowerShell CLI installer or a WSL2 Gateway; macOS has a menu-bar app; Linux runs the Gateway on any server. The docs also have pages for iOS, Android and ChromeOS. [39] |
| Minimum and recommended machine size | Table on the same page: memory 1 GB minimum, 2-4 GB recommended; CPU 1 core minimum, 2 cores recommended; data disk 500 MB minimum, 2+ GB recommended. Browser use raises the practical memory floor to 2 GB. [40][41] | Same page: recommended 1-2 vCPU and 2 GB+ RAM for headroom; browser automation and multiple channels raise the need. The Fly.io guide uses shared-cpu-2x with 2 GB RAM. [42][43] |
| Security model | Layers: user authorization, dangerous-command approval, file-write safety, container isolation, MCP credential filtering, context-file scanning, cross-session isolation and input sanitization. Approvals run in smart, manual or off mode; unattended contexts (cron, webhooks, single-query) deny dangerous commands by default. [44][45] | When enabled, tool execution moves into Docker, Podman, SSH, OpenShell or Crabbox; the Gateway itself always stays on the host. Four modes: read-only, guarded (human review after an allowlist fast path), workspace (LLM review, human on ask) and full, which needs operator.admin. The Gateway binds to loopback, unknown DM senders get a pairing code, and group access is allowlisted behind a mention gate. One gateway is one trust domain; mixed-trust use needs separate gateways and credentials. [46][47][48][49] |
| Official or recommended hosting | Nous Research runs Hermes Cloud, managed from the /agents page of Nous Portal, alongside self-hosting on a VPS, Docker or serverless backends. [50] | The docs publish provider guides for Azure, DigitalOcean, Fly.io, GCP, Hetzner, Hostinger, Northflank, Oracle Cloud, Daytona and exe.dev. There is no first-party hosted OpenClaw; hosting is self-managed on a VPS, home server or one of the documented platforms. [51][52] |
| Release cadence and latest version | Latest GitHub release v0.22.5, published 2026-10-23. GitHub lists 8 releases in the 30 days to 2026-10-29 and 21 in 90 days, roughly one or two a week. [53] | Latest GitHub release 2026.10.9, published 2026-10-29; npm latest is also 2026.10.9. GitHub lists 29 releases in the 30 days to 2026-10-29 and 49 in 90 days. Four channels: stable, extended-stable, beta and dev. [54][55] |
| GitHub stars and repository age | about 258,474 stars and 56,761 forks on 2026-10-29, as counted by GitHub's API. Repository created 2025-07-22, so about 15 months old on 2026-10-29. [56][57] | about 393,376 stars and 82,688 forks on 2026-10-29, as counted by GitHub's API. Repository created 2025-11-24, so about 11 months old on 2026-10-29. [58][59] |
| Notable differences and known issues | One Hermes home serves one agent process; a second agent needs its own profile. You cannot bind several profiles to different chats on one WhatsApp number. [60][61] | The docs describe sandboxing as damage limitation, not a hardened boundary. The stock Docker image needs extra build flags to add a browser or system packages. [62][63] |
Hosting either one
Both need a machine that stays on, keeps its files across restarts and has about 2 GB of memory once a browser is involved. The guides price every provider for that and show how to set each agent up on it:hosting Hermes and hosting OpenClaw.
Sources
Each quote was read on the page linked, on the date shown
- Hermes: hermes-agent.nousresearch.com/docs/ (29 October 2026) docs-confirmed
“The self-improving AI agent built by Nous Research.”
- Hermes: github.com/NousResearch/hermes-agent/blob/main/LICENSE (29 October 2026) docs-confirmed
“Copyright (c) 2025 Nous Research”
- OpenClaw: github.com/openclaw/openclaw (29 October 2026) docs-confirmed
“OpenClaw is an open-source AI assistant that runs on your own computer and meets you in the channels you already use”
- OpenClaw: github.com/openclaw/openclaw/blob/main/LICENSE (29 October 2026) docs-confirmed
“Copyright (c) 2026 OpenClaw Foundation”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/messaging/ (29 October 2026) docs-confirmed
“Chat with Hermes from Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Home Assistant, Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, Weixin, BlueBubbles (iMessage), QQ, Yuanbao, Microsoft Teams, LINE, ntfy, or your browser.”
- OpenClaw: docs.openclaw.ai/help/faq/what-is-openclaw (29 October 2026) docs-confirmed
“It replies on the messaging surfaces you already use (Discord, Google Chat, iMessage, Mattermost, Signal, Slack, Telegram, WebChat, WhatsApp, and bundled channel plugins such as QQ Bot)”
- Hermes: hermes-agent.nousresearch.com/docs/integrations/providers (29 October 2026) docs-confirmed
“You need at least one way to connect to an LLM.”
- Hermes: hermes-agent.nousresearch.com/docs/integrations/providers (29 October 2026) docs-confirmed
“Nous Portal is Nous Research's unified subscription gateway and the recommended way to run Hermes Agent.”
- Hermes: github.com/NousResearch/hermes-agent (29 October 2026) docs-confirmed
“You can still bring your own keys per-tool whenever you want - the gateway is per-backend, not all-or-nothing.”
- OpenClaw: docs.openclaw.ai/help/faq/what-is-openclaw (29 October 2026) docs-confirmed
“Model-agnostic: use Anthropic, MiniMax, OpenAI, OpenRouter, etc., with per-agent routing and failover.”
- OpenClaw: docs.openclaw.ai/concepts/model-providers/control-ui-and-keys (29 October 2026) docs-confirmed
“Open Settings → Models in the Control UI to add, replace, or remove provider API keys.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/memory (29 October 2026) docs-confirmed
“Hermes Agent has bounded, curated memory that persists across sessions.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/memory (29 October 2026) docs-confirmed
“Character limits keep memory focused. Memory does not auto-compact: when a write would exceed the limit, the memory tool returns an error instead of silently dropping entries.”
- OpenClaw: docs.openclaw.ai/help/faq/sandboxing-and-memory (29 October 2026) docs-confirmed
“OpenClaw memory is Markdown files in the agent workspace: daily notes in memory/YYYY-MM-DD.md, curated long-term notes in MEMORY.md (main/private sessions only).”
- OpenClaw: docs.openclaw.ai/help/faq/sandboxing-and-memory (29 October 2026) docs-confirmed
“Memory files live on disk and persist until deleted; the limit is your storage, not the model.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/skills (29 October 2026) docs-confirmed
“Skills are on-demand knowledge documents the agent can load when needed.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/skills (29 October 2026) docs-confirmed
“All skills live in ~/.hermes/skills/ - the primary directory and source of truth.”
- OpenClaw: docs.openclaw.ai/tools/skills (29 October 2026) docs-confirmed
“Skills are markdown instruction files that teach the agent how and when to use tools.”
- OpenClaw: docs.openclaw.ai/cli/skills (29 October 2026) docs-confirmed
“Inspect local skills, search ClawHub, install skills from ClawHub/Git/local directories, verify ClawHub skills, and update ClawHub-tracked installs.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/cron (29 October 2026) docs-confirmed
“Schedule tasks to run automatically with natural language or cron expressions.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/messaging/ (29 October 2026) docs-confirmed
“The gateway also runs the cron scheduler, ticking every 60 seconds to execute any due jobs.”
- OpenClaw: docs.openclaw.ai/automation/cron-jobs (29 October 2026) docs-confirmed
“Automations are OpenClaw's built-in scheduler.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/browser (29 October 2026) docs-confirmed
“The browser toolset gives the agent a real web browser.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/computer-use (29 October 2026) docs-confirmed
“Hermes Agent can drive your desktop - clicking, typing, scrolling, dragging - in the background on macOS, Windows, and Linux.”
- OpenClaw: docs.openclaw.ai/tools/browser (29 October 2026) docs-confirmed
“OpenClaw can run a dedicated Chrome/Brave/Edge/Chromium profile that the agent controls.”
- OpenClaw: docs.openclaw.ai/nodes/computer-use (29 October 2026) docs-confirmed
“Computer use lets an agent see and control the Gateway's own desktop or a paired node's desktop through one built-in computer tool.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/security (29 October 2026) docs-confirmed
“When running the messaging gateway, Hermes controls who can interact with the bot through a layered authorization system.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/security (29 October 2026) docs-confirmed
“If no allowlists are configured and GATEWAY_ALLOW_ALL_USERS is not set, all users are denied.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/messaging/discord (29 October 2026) docs-confirmed
“By default, Hermes isolates session history per user inside the channel for safety and clarity.”
- OpenClaw: docs.openclaw.ai/concepts/multi-user (29 October 2026) docs-confirmed
“Multi-user mode lets several trusted people operate the same OpenClaw agent.”
- OpenClaw: docs.openclaw.ai/concepts/multi-user (29 October 2026) docs-confirmed
“Everyone who can operate an agent can make it do anything that agent can do.”
- Hermes: hermes-agent.nousresearch.com/docs/ (29 October 2026) docs-confirmed
“Install in 60 seconds on Linux, macOS, WSL2, native Windows, Nix & NixOS or Android”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/docker (29 October 2026) docs-confirmed
“The container stores all user data (config, API keys, sessions, skills, memories) in a single directory mounted from the host at /opt/data.”
- OpenClaw: github.com/openclaw/openclaw (29 October 2026) docs-confirmed
“The installer supports macOS, Linux, and Windows.”
- OpenClaw: github.com/openclaw/openclaw (29 October 2026) docs-confirmed
“Already manage Node.js? Install the published package instead (Node 24.22.1+ or 26.12.1+; Node 26 recommended)”
- OpenClaw: github.com/openclaw/openclaw (29 October 2026) docs-confirmed
“npm install -g openclaw@latest --allow-scripts=openclaw”
- Hermes: hermes-agent.nousresearch.com/docs/getting-started/platform-support (29 October 2026) docs-confirmed
“We test on the latest Ubuntu and WSL2.”
- Hermes: hermes-agent.nousresearch.com/docs/getting-started/installation (29 October 2026) docs-confirmed
“The macOS installer is Apple Silicon only. macOS on x86 (Intel) processors is not a supported platform.”
- OpenClaw: docs.openclaw.ai/install (29 October 2026) docs-confirmed
“macOS, Linux, or Windows - Windows users can start with the native Windows Hub app, the PowerShell CLI installer, or a WSL2 Gateway.”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/docker (29 October 2026) docs-confirmed
“The Hermes container needs moderate resources. Recommended minimums:”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/docker (29 October 2026) docs-confirmed
“Browser automation (Playwright/Chromium) is the most memory-hungry feature. If you don't need browser tools, 1 GB is sufficient. With browser tools active, allocate at least 2 GB.”
- OpenClaw: docs.openclaw.ai/help/faq-first-run/providers-and-hosting (29 October 2026) docs-confirmed
“Absolute minimum: 1 vCPU, 1 GB RAM, ~500 MB disk.”
- OpenClaw: docs.openclaw.ai/help/faq-first-run/providers-and-hosting (29 October 2026) docs-confirmed
“Recommended: 1-2 vCPU, 2 GB+ RAM for headroom (logs, media, multiple channels).”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/security (29 October 2026) docs-confirmed
“The security model has eight layers:”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/security (29 October 2026) docs-confirmed
“Before executing any command, Hermes checks it against a curated list of dangerous patterns. If a match is found, the user must explicitly approve it.”
- OpenClaw: docs.openclaw.ai/gateway/sandboxing (29 October 2026) docs-confirmed
“Sandboxing is off by default and controlled by agents.defaults.sandbox”
- OpenClaw: docs.openclaw.ai/gateway/permission-modes (29 October 2026) docs-confirmed
“Session permission modes set one session's filesystem boundary and exec escalation reviewer.”
- OpenClaw: docs.openclaw.ai/gateway/security (29 October 2026) docs-confirmed
“OpenClaw ships with conservative defaults.”
- OpenClaw: docs.openclaw.ai/gateway/security (29 October 2026) docs-confirmed
“OpenClaw is not a hostile multi-tenant security boundary for mutually adversarial users sharing one agent or gateway.”
- Hermes: hermes-agent.nousresearch.com/docs/guides/manage-hermes-cloud-with-mcp (29 October 2026) docs-confirmed
“Hermes Cloud runs hosted Hermes Agent instances for you.”
- OpenClaw: docs.openclaw.ai/vps (29 October 2026) docs-confirmed
“Run the OpenClaw Gateway on any Linux server or cloud VPS.”
- OpenClaw: github.com/openclaw/openclaw (29 October 2026) docs-confirmed
“has no paid tier, hosted service, or token”
- Hermes: api.github.com/repos/NousResearch/hermes-agent/releases/latest (29 October 2026) third-party
“"name":"Hermes Agent v0.22.5"”
- OpenClaw: api.github.com/repos/openclaw/openclaw/releases/latest (29 October 2026) third-party
“"name":"openclaw 2026.10.9"”
- OpenClaw: docs.openclaw.ai/install/development-channels (29 October 2026) docs-confirmed
“Stable builds usually ship to beta first, get vetted there, then get promoted to latest without a version bump.”
- Hermes: api.github.com/repos/NousResearch/hermes-agent (29 October 2026) third-party
“"stargazers_count":258474”
- Hermes: api.github.com/repos/NousResearch/hermes-agent (29 October 2026) third-party
“"created_at":"2025-07-22T22:22:28Z"”
- OpenClaw: api.github.com/repos/openclaw/openclaw (29 October 2026) third-party
“"stargazers_count":393376”
- OpenClaw: api.github.com/repos/openclaw/openclaw (29 October 2026) third-party
“"created_at":"2025-11-24T10:16:47Z"”
- Hermes: hermes-agent.nousresearch.com/docs/user-guide/features/memory (29 October 2026) docs-confirmed
“Don't point two agent processes at the same Hermes home directory.”
- Hermes: hermes-agent.nousresearch.com/docs/reference/faq (29 October 2026) docs-confirmed
“Current limitation: Hermes profiles each require their own WhatsApp number/session.”
- OpenClaw: docs.openclaw.ai/gateway/sandboxing (29 October 2026) docs-confirmed
“This is not a perfect security boundary, but it materially limits filesystem and process access when the model does something dumb.”
- OpenClaw: docs.openclaw.ai/help/faq/sandboxing-and-memory (29 October 2026) docs-confirmed
“The default image is security-first and runs as the node user, so it excludes system packages, Homebrew, and bundled browsers.”