All three are open-source agents you run yourself. They overlap, and they are good at different things. This page is written by the HQ maintainers, so weigh it accordingly.
HQ facts come from HQ's own repository and docs. That is not an outside audit.
Facts about OpenClaw and Hermes Agent come from their official repositories and docs, checked on 2026-10-05, and each cell links its source. Both projects ship often, so details will age.
"Reported" means we saw it from third parties and did not verify it. "Not reviewed" means we did not look.
Found a mistake? Open an issue and we will fix it.
| Topic | Agent HQ | OpenClaw | Hermes Agent |
|---|---|---|---|
| What it is | A local-first agent hub. One Rust binary, a markdown vault, and a focus on running coding agents. Source: README | A self-hosted personal assistant. One local Gateway process holds sessions, routing and channel connections. Source: docs | A self-improving agent with a learning loop that creates skills from experience and refines them. Source: repo |
| Language and license | Rust (edition 2024). MIT. Source: LICENSE | TypeScript on Node. MIT. Source: repo | Python, with some Node. MIT. Source: repo |
| Installing | An install script for Linux (x86_64 and ARM) and macOS (Apple Silicon and Intel), Homebrew, npx agent-hq-cli, and a Docker image, all from signed releases. Windows gets a PowerShell installer for Full HQ in WSL2 or HQ Lite. There is also a one-form Hetzner server, servers that update themselves, and building from source with cargo. Source: install guide | Install script, PowerShell installer, npm, Docker and Nix. Source: docs | Install script, PowerShell installer and an APT route for Termux. Source: repo |
| Chat channels | Discord, Telegram, a web PWA and the terminal. No Slack or WhatsApp. Source: README | Discord, iMessage, Slack, Teams, Telegram, WhatsApp, Google Chat, Signal and more. Plugins add others. Source: docs | Telegram, Discord, Slack, WhatsApp, Signal, Email and a CLI, through one gateway process. The docs list more platforms. Source: repo |
| Memory | Markdown vault plus SQLite with FTS5 and embeddings. A consolidator and a forgetter tidy it over time. Source: architecture | Not reviewed for this page. | Persistent memory, FTS5 cross-session search with LLM summarization, and Honcho user modeling. Source: repo |
| Extending it | About 100 internal tools behind a two-tool MCP gateway, markdown skills, remote MCP servers bridged in. HQ is also an MCP server other agents can call, and can be added to claude.ai as a custom connector (off by default). Source: README | Skills, a plugin SDK for channels, providers and tools, the ClawHub marketplace, and companion apps for voice, Canvas, camera and screen. Source: docs | Skills compatible with agentskills.io and a Skills Hub, subagents, built-in cron and an MCP client. Source: repo |
| Supervising coding agents | What HQ is built around. It starts, watches and steers Claude Code, Codex, OpenCode, Cursor, Copilot CLI and others as sessions in its built-in host, locally or on a paired machine, and alerts you when one blocks. It never answers a trust prompt for you. Source: sessions docs | We have not verified how OpenClaw handles this. | We have not verified how Hermes Agent handles this. |
| Running the others | HQ can run Hermes Agent and OpenClaw as personal agents in its sandbox, for research, review and monitoring, and hand them tasks from its board. They never get code work, an HQ token or your credentials, and a person reviews what they produce. Source: personal agents guide | Runs under HQ in local mode, with its own provider key and without its gateway daemon. Source: personal agents guide | Runs under HQ, interactively or as one-shot jobs, with its own provider key. Source: personal agents guide |
| Isolating what the agent runs | Agent shell commands run in a bubblewrap or sandbox-exec sandbox, required by default after 0.9.0, with an environment allowlist and taint tracking after untrusted reads. Source: bash sandbox | Sandboxing is off by default. When enabled, only tool execution moves into Docker, Podman, SSH, OpenShell or Crabbox while the Gateway stays on the host. The docs call it "not a perfect security boundary". Source: sandboxing docs | Command approval and container isolation. Execution backends include local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox. Source: docs |
| Updates | Servers pull minisign-signed releases on a timer and roll back on a failed health check. Source: update system | Not reviewed for this page. | Not reviewed for this page. |
| Where it runs | A Mac or Linux machine, a VPS, Docker, or Windows (WSL2 or HQ Lite). An Android phone build is a developer preview: you build the APK yourself, and whether a sleeping, unplugged phone answers in time is not verified yet. To reach it, Tailscale is documented; hq net is a self-hosted WireGuard network (its root helper is not in release builds yet); hq edge is a public forwarder you run yourself for a device behind NAT. There is no hosted service. Source: mobile docs | Self-hosted. Source: docs | Local, or on serverless backends such as Modal and Daytona. Source: docs |
| Size and maturity | Small and early: version 0.9.x, no outside audit. Source: changelog | A very large community and frequent releases. At 2026-10-05 the latest was v2026.9.8, from 2026-10-03. Source: releases | A large community and frequent releases. At 2026-10-05 the latest was v0.21.5, from 2026-09-24. Reported to have first shipped in February 2026. Source: releases |
| Security record | Too young to have one. The model and its known gaps are written down. Source: security page | Reported by third parties, and not verified by us: past vulnerabilities and malicious skills on its marketplace. Read the project advisories yourself. Source: repo | Reported by third-party aggregators, and not verified by us: a number of mostly moderate CVEs in 2026. Read the project advisories yourself. Source: repo |
Choosing
These are not exclusive. Hermes Agent can import from OpenClaw (its repository documents
hermes claw migrate), and HQ can run either one for you as a personal agent: sandboxed, with
its own provider key, kept away from code, and handed research or review tasks from your board. The
personal agents guide covers setup and removal. If you are unsure,
start with the one whose channels you already use.
