OpenClaw vs Hermes Agent: which self-hosted AI agent?
By AppDuo Editorial Team · Updated
Quick verdict
OpenClaw and Hermes Agent are both MIT-licensed, open-source AI agents that run on your own machine and take orders through chat apps. The software is free; the real cost is the AI model.
OpenClaw: suits people who want to message it through iMessage or Zalo, or want iOS and Android apps and a multilingual interface. The trade-offs are a confirmed history of malicious skills on its ClawHub marketplace and tools that run directly on the host unless you configure sandboxing.
Hermes Agent: suits people who want one agent with one memory across channels, scheduled jobs, and the option of a cloud agent included with Nous Portal plans from $20 a month. The trade-offs are an independent audit, reported by the Cloud Security Alliance, listing 4 Critical findings in v0.8.0 with no stated patch status, and a Nous Portal policy that lets Nous use data to improve services and share it with AI-training providers unless you turn on Privacy Mode.
Both have publicly reported weaknesses in their default configuration, so you have to lock down access before letting the agent run commands.
Head-to-head
OpenClaw leads on 2, Hermes Agent on 3, 5 even
| Feature | OpenClaw | Hermes Agent |
|---|---|---|
| Setup and configuration effort[1],[3],[5],[8],[13],[17] | One-line installer, npm, Docker, Nix or desktop apps for macOS 15+, Windows 10/11 and Linux. It still needs Node, API keys and your own security setup; a maintainer warns it is unsafe for people who cannot use a command line. | A one-line install on macOS, Linux or Windows, or a desktop app for macOS 12+ and Windows 10/11 (the Windows app is on the Microsoft Store). You still pick a model provider, chat channels and a runtime; the README warns Windows Defender or Bitdefender may quarantine the bundled uv.exe, which it calls a false positive. |
| → Both are aimed at technical users. Hermes's desktop app runs on macOS 12 and later, while OpenClaw needs macOS 15+; OpenClaw, on the other hand, also installs through npm, Docker or Nix for people who already use them. | ||
| Where the agent runs[1],[8],[9],[10] | You run the Gateway process on your own computer or server; the site states there is no hosted version. | Local machine, Docker, SSH, Modal and other backends. Paid Nous Portal plans include one cloud agent (Plus: up to 4 GB RAM and 25 GB disk). |
| → If you would rather not keep a machine running all day, a paid Nous Portal plan gives Hermes a cloud agent; with OpenClaw you always provide the server yourself. | ||
| Software and model costs (personal)[1],[9] | Free under the MIT license, with no paid tier. You pay your chosen model provider's API fees and for the machine running the Gateway. | Free software (MIT). Nous Portal is optional: Free at $0 (free models only), Plus at $20/month, Super at $100/month, Ultra at $200/month, no yearly billing. API keys draw on prepaid credits topped up in $10, $20, $50, $100 or $200 amounts. |
| → With both, total cost depends on model usage. Nous Portal gives a fixed monthly fee, but the pricing page does not state included credits and pricier models use the allowance faster. | ||
| Memory and automation[1],[8],[10] | Per the site: remembers context across conversations; handles email and calendar, browses the web, fills forms, reads and writes files and runs shell commands. The agent can write new skills itself. | One agent with one memory whether you message it from the CLI, Telegram or email; conversations, memory and skills survive restarts. It adds natural-language scheduling, isolated subagents, browser automation, image generation and text-to-speech. |
| → If you want recurring jobs or work split across subagents, Hermes documents both; the OpenClaw sources on record mention neither. | ||
| Chat channels and devices[1],[3],[4],[10] | The site cites 29 channels, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Microsoft Teams, Google Chat and Zalo (as a plugin). iOS and Android apps act as nodes. | CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, email, SMS, DingTalk, Feishu, WeCom, Microsoft Teams, Google Chat and more. Desktop apps for macOS and Windows are on record; no iOS or Android app was found. |
| → Of the two, only OpenClaw is recorded with iMessage, Zalo and phone apps. Hermes explicitly lists email, SMS and work chat apps such as DingTalk, Feishu and WeCom. | ||
| Interface and documentation languages[4],[10] | The iOS app interface comes in 22 languages (English, Vietnamese, Japanese, Korean, …). | Documentation in English and Simplified Chinese; no other interface languages were verified. |
| → Readers less comfortable in English gain from OpenClaw's multilingual iOS app. How well either agent replies in a given language depends on the model you choose. | ||
| Default security and published flaws[2],[5],[14],[15] | The README says tools run on the host unless sandboxing is configured. Wikipedia, citing The Register, says 2.0 neither encrypts secrets at rest nor enables the sandbox by default; this is a criticism with no project response found. | Per the Cloud Security Alliance, an independent audit (11 Apr 2026) reported 4 Critical and 9 High findings in v0.8.0, with no patch status given. CVE-2026-53870 (conversation history readable by other local users) is fixed in 0.16.0. |
| → With both, you must enable sandboxing and command approval and keep updating. The available sources record no developer response to the audit findings or the criticism above. | ||
| Third-party skill marketplace[6],[10],[14] | ClawHub has hosted malicious skills (confirmed): per Unit 42, Koi Security disclosed 341, and Unit 42 still found 5 that slipped past screening between February and May 2026. The project banned the accounts, deleted the skills, screens with VirusTotal and ClawScan, and on 1 June 2026 announced a partnership with NVIDIA on screening. | Has a Skills Hub; no malicious-skill incident is on record. The independent audit notes the agent can write skills that persist across sessions and that plugins run unsandboxed. |
| → Anyone planning to install many community skills should be more careful with ClawHub given its record. No similar incident is recorded for Hermes, but that does not prove its hub is safe. | ||
| Where your data goes[1],[2],[8],[11] | Runs on your machine and by default sends only a daily version check, which can be turned off. Messages still go to your chosen model provider, and the agent needs access to email, calendar, files and the shell. | A local install keeps conversations, memory and skills in a Hermes directory on your machine. With Nous Portal or a cloud agent, Nous may use prompts, outputs and uploads to improve its services and share them with AI-training providers unless you turn on Privacy Mode (effective only from then on). |
| → Self-hosting either with your own model provider keeps data on your machine and sends prompts to that provider. If you use Nous Portal, turn on Privacy Mode from the start. | ||
| Community, ratings and backing[2],[4],[5],[13],[16],[17] | About 394k GitHub stars and about 4.8k open issues; 3.5/5 from 76 ratings on the US App Store. Stewarded by the OpenClaw Foundation, a US 501(c)(3) nonprofit; OpenAI is a donor, not an owner. | About 262.6k GitHub stars and 58.8k forks; 4.7/5 from 16 ratings on the US Microsoft Store. Built by Nous Research, which per Fortune raised a $50 million Series A in 2025. |
| → Both offer docs and community help; neither is recorded with staffed support or a response-time commitment. Store rating counts are too small to judge quality. | ||
Choose OpenClaw if…
- People who want to give orders through iMessage or Zalo, or use iOS and Android apps as nodes
- People who prefer a project stewarded by a nonprofit (the OpenClaw Foundation) with no paid tier from the developer
- People who need a multilingual interface (22 languages in the iOS app)
Choose Hermes Agent if…
- People who want one agent with one memory across channels, scheduled jobs and subagents
- People who would rather not run a server and accept a Nous Portal plan (from $20/month) for a cloud agent
- People who need channels Hermes explicitly lists: email, SMS, Matrix, Mattermost, DingTalk, Feishu or WeCom
Same idea, different approach
Both give you an AI agent that runs on your own hardware, takes orders through chat apps and can read files, run commands and browse on your behalf. OpenClaw, started by Peter Steinberger in late 2025 and now stewarded by the OpenClaw Foundation, focuses on reach: 29 chat channels per its site and apps for desktop and phone. Nous Research's Hermes Agent is built around one agent with one memory: whether you message it from the CLI, Telegram or email, it remembers projects, writes its own skills, runs scheduled jobs and hands work to subagents. For models, OpenClaw lists Claude, GPT and local models; Hermes lists Nous Portal, OpenRouter, OpenAI or any endpoint.
Costs: free software, you pay for the model
Both are free under the MIT license. OpenClaw has no paid or hosted tier, so you pay only the model provider's API fees and for the machine running the Gateway. Hermes is the same when self-hosted, but also offers Nous Portal: Free at $0 (free models only), Plus at $20, Super at $100 and Ultra at $200 a month, with no yearly billing. The paid plans include Hermes usage capped by daily, weekly and monthly limits plus one cloud agent (Plus: 4 GB RAM, 25 GB disk; Super: 4 GB RAM, 100 GB; Ultra: 8 GB RAM, 100 GB), but the pricing page does not state the included credit amounts, bigger sizes or extra agents draw on credits, and the page states no refund terms. With either agent, the monthly bill depends mostly on how much you call the model.
Security: each project's record
This is the part to read most carefully. For OpenClaw, the confirmed incident is malicious skills on ClawHub: per Unit 42, Koi Security disclosed 341 malicious skills, Bitdefender estimated about 17% of the skills it analysed in the first weeks carried malicious payloads, and five skills still slipped past screening between February and May 2026. The project banned the accounts, deleted the skills and partnered with NVIDIA on screening. Wikipedia, citing The Register, says 2.0 neither encrypts secrets at rest nor enables the sandbox by default; no project response was found. For Hermes, per the Cloud Security Alliance, an independent audit reported 4 Critical and 9 High findings in v0.8.0, such as a terminal tool guarded only by bypassable regex checks and a file tool that can reach SSH keys; these have no CVE IDs, CSA gives no patch status and records no response from Nous Research. There are also three Medium CVEs (two fixed; CSA cites no patch for CVE-2026-7396) and CVE-2026-53870, fixed in 0.16.0. The mitigations are the same for both: sandbox or containerise, restrict who can send commands, enable command approval and keep updating.
Privacy and outside reports
When self-hosted, both keep data on your machine, though prompts still reach the model provider you choose. OpenClaw by default sends only a daily version check, which can be turned off. For Hermes, if you use Nous Portal or a cloud agent, the policy updated on 10 Nov 2026 lets Nous use data to improve its services and share it with AI-training providers unless you turn on Privacy Mode, which applies only from then on. On outside reports: Bloomberg reported that Chinese state agencies and state-owned banks were warned against installing OpenClaw on office devices; the report relies on anonymous sources and concerns work devices, not individual users. The AI Incident Database lists two July 2026 intrusions in which attackers reportedly used Hermes, one involving a framework built on both Hermes and OpenClaw; these are allegations of misuse, not a flaw in Hermes. Wikipedia also summarises a student's account that his OpenClaw agent created a profile on the third-party dating platform MoltMatch without his explicit direction; this is an unverified account.
Which one to pick
There is no overall winner. Pick OpenClaw if you need to message it through iMessage or Zalo, want phone apps or a multilingual interface, and are prepared to run your own server; then be sparing with ClawHub skills. Pick Hermes Agent if you value one shared memory, scheduled jobs and subagents, or want a fixed monthly fee for a cloud agent; then track the patch status of the audit findings and turn on Privacy Mode if you use Nous Portal. Neither suits people who just want a simple AI chat app without configuration, or organisations that need strict security boundaries by default.
Frequently asked questions
Do OpenClaw or Hermes Agent cost money?
Both are free under the MIT license. You still pay your chosen AI model provider for API usage. Hermes also offers optional Nous Portal plans: a Free tier limited to free models, and Plus at $20, Super at $100 and Ultra at $200 a month.
Which one is safer?
There is no clear answer. OpenClaw has a confirmed malicious-skill incident on ClawHub, and The Register, as cited by Wikipedia, criticised 2.0 for not enabling its sandbox by default. Hermes has an independent audit reporting Critical findings in v0.8.0 with no stated patch status, plus four published CVEs, three of them fixed. With either, enable sandboxing, restrict who can send commands and update regularly.
Can I use Vietnamese or Zalo?
OpenClaw lists Zalo as a chat channel (as a plugin) and its iOS app has a Vietnamese interface. Zalo is not among the Hermes channels on record, and only English and Simplified Chinese documentation could be verified; how well it replies in Vietnamese depends on the model you choose.
Who is behind each project?
OpenClaw was started by Peter Steinberger; after he joined OpenAI in February 2026, the OpenClaw Foundation took over stewardship, and the README says OpenAI is only a donor. Hermes Agent is built by Nous Research, which per Fortune raised a $50 million Series A in 2025.
Sources
Numbers in the table refer to these sources, read on the dates shown.
- 1.Official website · Vendor-provided · Nov 11, 2026
- 2.GitHub repository · Vendor-provided · Nov 11, 2026
- 3.Official docs · Vendor-provided · Nov 11, 2026
- 4.App Store listing (US) · store · Nov 11, 2026
- 5.Wikipedia: OpenClaw · Community-reported · Nov 11, 2026
- 6.Palo Alto Networks Unit 42 (2026-06-23) · press · Nov 11, 2026
- 7.The Star / Bloomberg (2026-03-11) · press · Nov 11, 2026
- 8.Official Hermes Agent homepage · Vendor-provided · Nov 11, 2026
- 9.Nous Portal pricing · Vendor-provided · Nov 11, 2026
- 10.Hermes Agent documentation · Vendor-provided · Nov 11, 2026
- 11.Nous Portal privacy policy · Vendor-provided · Nov 11, 2026
- 12.Nous Research homepage · Vendor-provided · Nov 11, 2026
- 13.GitHub: NousResearch/hermes-agent · Vendor-provided · Nov 11, 2026
- 14.Cloud Security Alliance research note on Hermes Agent CVEs (2026-05-04) · press · Nov 11, 2026
- 15.OffSeq Threat Radar: CVE-2026-53870 · Community-reported · Nov 11, 2026
- 16.Fortune: Paradigm backs Nous Research (2025-04-25) · press · Nov 11, 2026
- 17.Microsoft Store: Hermes Agent · store · Nov 11, 2026
- 18.AI Incident Database: Hermes Agent entity · Community-reported · Nov 11, 2026
- 19.Borncity: Hermes misuse report (2026-08-03) · press · Nov 11, 2026
Prepared by the AppDuo Editorial Team from the sources cited in this article. AI assists with organising and cross-checking data; editorial responsibility remains with AppDuo.
Comments
No comments yet. Be the first to share your experience.
Comments are readers’ opinions, not content from the AppDuo Editorial Team.

