OpenClaw and Hermes Agent are the two open-source AI agents most people end up choosing between in 2026. Both are MIT-licensed, both run on hardware you control, both talk to you over Telegram, Slack, Discord, or WhatsApp, and both can now write their own skills. That overlap is why "Hermes vs OpenClaw" became one of the most searched agent comparisons of the year. It is also why so many of the answers are already out of date.
This guide compares OpenClaw v2026.10.5 and Hermes Agent v0.22.2, the latest stable releases as of October 18, 2026. Claims link to each project's documentation, release notes, or public security records, so you can check them yourself.
TL;DR: OpenClaw or Hermes Agent?
- Pick OpenClaw if you want a resident assistant on one always-on machine, a large channel catalog (including niche options like Twitch, Nostr, and Nextcloud Talk), and tool policy enforced in code. It is written in TypeScript, stewarded by the nonprofit OpenClaw Foundation, and has no paid tier.
- Pick Hermes Agent if you want a personal agent that works inside your real project directories, keeps small and predictable memory, and can push execution out to Docker, SSH, Modal, Daytona, or Vercel Sandbox. It is written in Python by Nous Research, with optional paid inference and hosting.
- Self-learning no longer separates them. Hermes launched with agent-written skills, and OpenClaw has learned skills automatically by default since v2026.8.1.
- Neither one is "the safe one." OpenClaw has the longer public vulnerability record. Hermes has fewer published advisories, but its own security policy says the operating system is the only boundary against an adversarial model.
- You are not locked in. Each project ships an importer for the other's configuration, memory, and skills.
OpenClaw vs Hermes Agent at a glance
| OpenClaw | Hermes Agent | |
|---|---|---|
| Maker | OpenClaw Foundation, a 501(c)(3) nonprofit | Nous Research, venture-backed |
| License | MIT | MIT |
| Language and runtime | TypeScript on Node 24.16+ or 26 | Python 3.14 in managed installs (3.11+ from source) |
| Latest stable release | v2026.10.5 (calendar versioning) | v0.22.2 (SemVer) |
| GitHub stars (Oct 18, 2026) | About 392,000 | About 255,000 |
| Process model | Resident Gateway daemon that owns sessions, tools, and channels | CLI, TUI, desktop app, and one gateway per machine serving every profile |
| Where tools run | On the host by default; optional Docker, Podman, SSH, OpenShell, or Crabbox sandbox | Local, Docker, SSH, Singularity, Modal, Daytona, or Vercel Sandbox |
| Built-in memory | Markdown files plus SQLite keyword, vector, and hybrid search | Capped MEMORY.md and USER.md plus full-text search over every past session |
| Self-learning | Skill Workshop, automatic by default since August 2026 | Agent-written skills and a curator, core since launch |
| Skill sources | ClawHub | Skills Hub (which includes ClawHub) plus a reviewed plugin catalog of about 600 entries |
| Web search out of the box | Hidden until you configure a provider | Works immediately through a keyless free-tier fallback |
| Paid offering | None | Nous Portal plans, Hermes Cloud, and a Tool Gateway (all optional) |
| Import from the other | openclaw migrate hermes | hermes claw migrate |
Channel and provider counts appear later in this guide, but treat them loosely. OpenClaw's own Hermes comparison page points out that the catalogs mix vendors, auth routes, and plugin types, so the totals are not directly comparable.
Five outdated claims in other Hermes vs OpenClaw comparisons
Most of the comparisons ranking today were written between April and June 2026. OpenClaw alone has published more than 130 stable releases since November 2025, and Hermes shipped a major release in October. Check any comparison's date against these changes before trusting it:
- "Only Hermes writes its own skills." OpenClaw introduced Skill Workshop in v2026.4.22 and made automatic self-learning the default in v2026.8.1 on August 31, 2026.
- "Hermes has zero CVEs." CVE-2026-14625, published July 4, 2026, covers a protection-mechanism failure in hermes-agent up to 0.15.2. What is still true is that Hermes has no public GitHub security advisories.
- "Hermes is CLI-only" or "Hermes supports six channels." Hermes now ships a desktop app, a web dashboard, and roughly 29 messaging adapters.
- "Hermes requires Python 3.11." Managed installs moved to Python 3.14 in v0.22.0. Source installs still accept 3.11 through 3.14.
- "Hermes tracks the main branch, so expect breakage." Since v0.22.0, Hermes ships tested SemVer releases and source installs follow a stable channel by default.
What is OpenClaw?
OpenClaw is an open-source, self-hosted gateway for personal AI agents. One Gateway process is the control plane for sessions, tools, events, and channels. The web Control UI, CLI, TUI, macOS menu-bar app, Windows Hub, and the iOS and Android companion "nodes" all connect to it.
Peter Steinberger started it as a weekend project in November 2025. It was called Clawd (often written Clawdbot) and then Moltbot before it settled on OpenClaw on January 29, 2026. Steinberger joined OpenAI in February 2026, and in July the project's stewardship was formalized under the OpenClaw Foundation, a donor-funded 501(c)(3) whose backers include OpenAI and the University of Michigan. The project states that it has no paid tier, hosted service, or token.
Installation on macOS, Linux, or WSL2 is one line, followed by onboarding:
curl -fsSL https://openclaw.ai/install.sh | bashWindows has a native PowerShell installer and a Windows Hub app, and there is an official Docker image at ghcr.io/openclaw/openclaw. See the install guide for the npm route and release channels.
What is Hermes Agent?
Hermes Agent is Nous Research's open-source agent, pitched as "the agent that grows with you." Nous is best known for its Hermes family of open-weight models and raised a Paradigm-led $50 million Series A in April 2025. The agent's repository dates to July 2025, the public launch came in February 2026, and the first tagged release, v0.2.0, shipped on March 12, 2026. Interest grew fast: US searches for "hermes agent" went from about 480 in February to 27,100 in March, according to DataForSEO.
The October v0.22.0 "Threshold" release reworked the product around one gateway and one backend per machine, a signed desktop app, a reviewed plugin catalog, and stable SemVer releases. Installation is also one line:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashA PowerShell installer and a nousresearch/hermes-agent Docker image cover Windows and containers. Hermes requires a model with at least 64,000 tokens of context and rejects smaller windows at startup, which matters if you plan to run small local models.
Architecture: a resident gateway vs a portable runtime
The most practical difference between OpenClaw and Hermes is where the agent lives.
OpenClaw is the Gateway. You install it on a Mac mini, a home server, or a VPS, and it stays running. Named agents are bound to channels, and tool calls execute on the Gateway host unless you enable a sandbox backend (Docker, Podman, SSH, OpenShell, or Crabbox) or route work to a paired node or cloud worker. The Gateway process itself always stays on the host. OpenClaw can also run Codex, the Copilot SDK, or the Claude Code CLI as native agent runtimes, and its Code Mode runs JavaScript in Node or QuickJS.
Hermes is a runtime you invoke. You can start it from a terminal inside a project, open the desktop app, or message it through the gateway. Seven terminal backends decide where commands run: local, docker, ssh, singularity, modal, daytona, and vercel_sandbox. The agent can sit on a small VPS while heavy commands run in a Modal or Daytona sandbox that hibernates when idle. Subagents are spawned through the delegate_task tool.
That difference shapes daily use. OpenClaw agents work from a workspace directory (by default ~/.openclaw/workspace) that holds their persona, rules, and memory files. Hermes works in whatever directory you launch it from, which feels more natural when you want an agent to edit a real repository in place.
Memory: unbounded files vs capped notes
Memory is where the two designs diverge most, and it drives both answer quality and token cost.
OpenClaw keeps memory as plain Markdown in the workspace. MEMORY.md holds long-term facts and loads at session start. Daily notes in memory/YYYY-MM-DD.md are indexed for search rather than injected, and a "dreaming" sweep distills them back into MEMORY.md. The default memory engine is SQLite with keyword, vector, and hybrid search, and plugins add Honcho, LanceDB, or a memory wiki. OpenClaw also tracks provenance: every indexed chunk carries an origin class (owner, agent, untrusted, or system) stored outside the text, so recalled content cannot promote its own trust level, and openclaw memory forget records deletions so participating paths do not re-ingest them.
Hermes keeps two deliberately small files in ~/.hermes/memories/. MEMORY.md is capped at 2,200 characters (about 800 tokens) and USER.md at 1,375 characters (about 500 tokens). Both are injected as a frozen snapshot when a session starts, and the agent edits them with its memory tool. Everything else is searchable through session_search, which runs SQLite FTS5 queries across every stored session. External memory providers such as Honcho, Mem0, and Supermemory install from the plugin catalog.
The trade-off is straightforward. Hermes keeps the base prompt small and predictable and leans on search for everything else. OpenClaw retains more and adds stronger lineage and deletion controls. Memory writes are autonomous by default in Hermes, with an optional memory.write_approval gate if you want to review them.
Skills and self-learning
Both agents use the same SKILL.md format from the Agent Skills standard, so a skill written for one usually loads in the other.
OpenClaw loads skills from the workspace, shared agent directories, and bundled sets. Skill Workshop runs a background review after every 10 model iterations in a session and captures reusable lessons as skills. In automatic mode it applies scanner-approved new skills and Workshop-owned edits on its own, but leaves changes to skills you wrote pending for your review. Unused learned skills are archived after 30 days. The public registry is ClawHub, which shows VirusTotal, ClawScan, and static-analysis results per listing.
Hermes creates and patches skills through its skill_manage tool, nudged by a background review, and /learn builds a skill from material you point it at. A curator marks agent-created skills stale after 14 days without use and archives them after 30. The Skills Hub pulls from the official set, skills.sh, GitHub taps, LobeHub, and ClawHub itself. Separately, the v0.22.0 plugin catalog lists about 600 plugins, each pinned to a reviewed commit.
The practical difference is less about whether an agent learns and more about how much you trust it to rewrite its own instructions. OpenClaw's default protects hand-written skills. Hermes leans further toward autonomy, with the curator keeping the skill library from growing without bound.
Channels, interfaces, and model providers
Channels. OpenClaw's channel catalog covers Telegram, WhatsApp, Slack, Discord, Signal, iMessage, Microsoft Teams, Google Chat, Matrix, Mattermost, LINE, IRC, Twitch, Nostr, Nextcloud Talk, SMS, WeChat, and more, plus a voice-call plugin. Email is handled as automation (an IMAP trigger or Gmail Pub/Sub) rather than as a chat adapter. Hermes covers most of the same mainstream apps and adds conversational email over IMAP and SMTP, SimpleX, ntfy, Home Assistant, and iMessage through BlueBubbles or Photon. It also exposes an OpenAI-compatible API server.
Interfaces. Does Hermes have a GUI like OpenClaw? It does now. OpenClaw offers the web Control UI, a macOS menu-bar app, a Windows Hub, and mobile nodes. Hermes offers Hermes Desktop (signed builds, including a Microsoft Store package and a macOS Apple Silicon app), a web dashboard, and a TUI.
Model providers. Both projects are model-agnostic. OpenClaw's providers index lists about 77 routes, including Anthropic, OpenAI with Codex OAuth, Gemini, OpenRouter, Bedrock, and local runtimes such as Ollama, LM Studio, vLLM, SGLang, and llama.cpp. Hermes supports a similar spread plus Nous Portal, which puts more than 300 models behind a single login, and a one-click llama.cpp setup for local models since v0.22.0.
Security: what the public record actually shows
Both agents can run shell commands, read your files, and act on messages from the outside world. Security depends far more on how you configure them than on which one you choose. The records still differ.
OpenClaw has the larger and more public history. CVE-2026-25253 allowed one-click token exfiltration through a gatewayUrl query parameter and was fixed in 2026.1.29. CVE-2026-32922, a device-token privilege escalation scored 9.9, was fixed in 2026.3.11. In February 2026, Koi Security audited 2,857 ClawHub skills and found 341 malicious ones, most from a single campaign that installed the Atomic Stealer malware on macOS. OpenClaw responded by adding VirusTotal scanning to ClawHub. The repository lists hundreds of published GitHub advisories, and the project's own docs note that "counts are not a safety score."
OpenClaw's security defaults include pairing codes for unknown DM senders, allowlisted and mention-gated groups, and an openclaw security audit command. Sandboxing is off by default, so tools run on the host until you configure a backend. The docs recommend an auto permission mode for host commands:
openclaw config set tools.exec.mode autoThe project also states that each Gateway is one trust boundary, not a hostile multi-tenant system.
Hermes has a shorter record: one third-party CVE (CVE-2026-14625, affecting versions up to 0.15.2) and no public repository advisories. Its security model has eight layers, including DM pairing, dangerous-command approval, MCP credential filtering, and scanning context files for injected instructions. The default approval mode, smart, asks an auxiliary model to assess flagged commands, and cron or one-shot sessions deny them by default. Two caveats matter. Dangerous-command checks are skipped on container backends, where the container is treated as the boundary. And the project's SECURITY.md says directly: "The only security boundary against an adversarial LLM is the operating system."
In both projects, plugins run in-process and unsandboxed. Whichever you pick, run it as a dedicated user, in a container, or on a separate machine. Never expose the gateway publicly without authentication, review skills before installing them, and treat every scraped page as untrusted input that may contain prompt injection.
Cost: is Hermes cheaper than OpenClaw?
Both are free software. Your real costs are hosting, model tokens, and any paid tool APIs such as search.
OpenClaw has no first-party paid product. You pay your host, your model provider, and whichever search provider you configure.
Hermes costs the same to self-host and adds optional paid services from Nous. Nous Portal has a free plan, Plus at $20 per month (with $22 in monthly credits), Super at $100, and Ultra at $200. Hermes Cloud runs managed instances billed per second of uptime plus storage, and stopped instances pay only for storage. The Tool Gateway bundles web search, extraction, image generation, text-to-speech, and a cloud browser on paid plans.
The "$5 per month for Hermes vs $20 to $32 for OpenClaw" figures that circulate come from hosting sellers' estimates, not measurements. Model usage dominates the bill, and it depends on your workload more than on the agent. Hermes' capped memory snapshot keeps its base prompt small. Token totals can still swing the other way: in one widely shared September 2026 run on GLM 5.3, published by the agent host Atomic Bot, Hermes used about 2.9 million tokens to OpenClaw's 2.1 million, yet cost slightly less (about $4.00 vs $4.50), and much of Hermes' extra usage went to checking its own work. That is a single vendor run, not a benchmark. Measure your own tasks before you switch to save money.
Web access: search, fetch, and scraping
An always-on agent is only as current as what it can read. Both agents ship web tools, and they make different default choices.
OpenClaw's web tools
OpenClaw provides web_search, web_fetch, and x_search. web_search stays hidden from the agent until a provider is configured. Bundled options include Brave, DuckDuckGo, Exa, Firecrawl, Gemini, Grok, Kimi, MiniMax, Ollama, Parallel, Perplexity, SearXNG, and Tavily, and auto-detection prefers keyed providers, starting with Brave. Results are cached for 15 minutes. web_fetch is a plain HTTP GET with Readability extraction and no JavaScript execution. It blocks private hosts and can fall back to a configured fetch provider. For interactive pages, there is a separate browser tool.
Hermes Agent's web tools
Hermes provides web_search and web_extract, with Firecrawl as the default backend. Other options include SearXNG, Brave, DDGS, Exa, Parallel, Tavily, Perplexity, Keenable, xAI, and OpenAI's native search. With nothing configured, requests rotate through a keyless free-tier ring with failover, so search works on first run. You can also split search and extraction across providers:
web:
search_backend: "searxng"
extract_backend: "firecrawl"For pages that need real interaction, Hermes supports browser backends such as Browser Use, Browserbase, Camofox, Lightpanda, and a local Chrome over CDP.
Where web access breaks down
Defaults get an agent started, but production web tasks hit the same walls in both projects. Pages render their content with JavaScript, bot protection blocks plain fetches, raw HTML floods the context window, and recurring checks re-read the same pages every run. If your agent briefs you on competitor pricing, tracks documentation changes, or researches leads overnight, the web layer decides whether the answer is accurate. For the trade-offs between tool styles, see our guide to web search APIs and MCP tools for AI agents.
Adding Context.dev to OpenClaw or Hermes
Context.dev plugs into both agents. In OpenClaw, the official ClawHub plugin registers Context.dev as the provider behind the native web_search and web_fetch tools and adds dedicated context_search and context_scrape tools:
openclaw plugins install clawhub:@contextdev/openclaw-context
openclaw gateway restartSet CONTEXT_API_KEY in the Gateway environment, then add the hosted MCP server with openclaw mcp add if you want the full tool catalog.
In Hermes, add the same MCP server to ~/.hermes/config.yaml. The auth: oauth line is required; without it, Hermes asks for a bearer token:
mcp_servers:
context:
url: "https://mcp.context.dev/mcp"
auth: oauthhermes mcp login context
hermes mcp test contextRun /reload-mcp in your session afterward. Either agent can then search the web, scrape any URL into clean Markdown with JavaScript rendering, crawl a site, extract JSON that matches a schema, capture screenshots, monitor pages for changes, and run large URL sets as asynchronous batches. The agent gets compact, structured results instead of raw HTML, and you skip running a browser fleet next to your gateway.
Migrating between OpenClaw and Hermes
Each project imports the other, which makes a side-by-side trial cheap.
OpenClaw to Hermes. The migration guide reads ~/.openclaw/ (and legacy ~/.clawdbot/ or ~/.moltbot/ directories) and imports your SOUL.md, MEMORY.md, USER.md, daily memory, skills, model providers, MCP servers, and tokens and allowlists for Telegram, Discord, Slack, Signal, Matrix, and Mattermost:
hermes claw migrate --dry-run
hermes claw migrateCron jobs, plugins, hooks, and a few workspace files are archived rather than imported, and WhatsApp needs to be paired again.
Hermes to OpenClaw. OpenClaw's importer brings over models, providers, MCP servers, SOUL.md, AGENTS.md, memory files, and skills. Sessions, cron jobs, plugins, and the session database are archived only:
openclaw migrate hermes --dry-run
openclaw migrate apply hermes --yesAdd --include-secrets to carry credentials across. Because the two use separate state directories, you can run both on one machine while you compare them on real tasks.
Which should you choose?
Match the agent to the job rather than to its star count:
- A household or team assistant reachable from many chat apps: OpenClaw. It pairs a large channel catalog with person-level role ceilings and an always-on Gateway built to be the single home for your agents.
- A solo developer's agent that edits real repositories: Hermes. It runs in your working directory, and its remote terminal backends keep risky or heavy commands off your laptop.
- Strict, auditable control over what tools may do: OpenClaw. Permission modes, code-enforced tool policy, exec approvals bound to exact commands, and memory provenance give you more levers.
- Managed inference or hosting without running a server: Hermes. Nous Portal and Hermes Cloud are first-party options. OpenClaw has none, although third-party hosts exist.
- A tight token budget: lean toward Hermes for its capped memory, then measure, because self-verification can erase the savings.
- Already happy on one: stay. Switch only for a specific missing capability, and use the importer when you do.
FAQs
Is Hermes better than OpenClaw?
Neither is better overall. Hermes is the stronger fit for a personal agent that works in project directories and runs commands on remote backends such as Modal or Daytona. OpenClaw is stronger for an always-on, multi-channel assistant with code-enforced tool policy. Since August 2026 both learn skills automatically, so self-learning alone is not a reason to switch.
Is Hermes Agent a fork of OpenClaw?
No. Hermes Agent is an independent Python project from Nous Research whose repository predates OpenClaw's. It reads OpenClaw's configuration through its migration command, and both use the same SKILL.md format, which is why the two are often confused.
Does Hermes have a GUI like OpenClaw?
Yes. Hermes ships Hermes Desktop for Windows and macOS, a web dashboard, and a terminal UI. OpenClaw offers a web Control UI, a macOS menu-bar app, a Windows Hub, and iOS and Android companion apps.
Is Hermes cheaper than OpenClaw?
Both are free and open source, so cost comes down to hosting and model tokens. Hermes keeps its injected memory capped at about 1,300 tokens, which helps, but total token use depends on the task. Hermes also sells optional Nous Portal plans from $20 per month and usage-billed Hermes Cloud instances, while OpenClaw has no paid tier.
Can Hermes work with OpenClaw?
Yes, in two ways. Each project imports the other's memory, skills, providers, and MCP servers, and Hermes' Skills Hub can install skills directly from ClawHub. Because they keep separate state directories, you can also run both on the same machine.
Is Hermes Agent safe?
Hermes has layered defaults, including DM pairing, model-assisted approval of dangerous commands, and credential filtering for MCP servers. Its own security policy still says the operating system is the only real boundary against an adversarial model. Run it in a container or under a dedicated user, and review skills and plugins before installing them. The same advice applies to OpenClaw.
Who makes Hermes Agent, and is it open source?
Nous Research, the lab behind the Hermes open-weight models, builds Hermes Agent. It is open source under the MIT license, and you can self-host it without a Nous account.
How do I give OpenClaw or Hermes better web access?
Configure a dedicated search and scraping provider rather than relying on plain HTTP fetches. For OpenClaw, install the Context.dev plugin from ClawHub so it powers the native web tools. For Hermes, add the Context.dev MCP server to config.yaml with OAuth. Both then get search, JavaScript-rendered scraping, crawling, structured extraction, and change monitoring.
The bottom line
OpenClaw and Hermes Agent have converged on features while diverging on philosophy. OpenClaw is a governed, always-on gateway that wants to be the one place your agents live. Hermes is a portable, autonomy-first agent that goes where your work is. Pick the shape that fits your setup, lock down where it can run commands, and give it a web layer it can trust. If you want that last part handled, Context.dev works with both.