Best Personal AI Agents You Can Self-Host in 2026
The short answer
OpenClaw is the default self-hosted personal agent in 2026; pick Hermes Agent if you want self-improving skills, QwenPaw if you want to run free on local models, and AnythingLLM if your agent is mostly about your documents. All four are open source, under active development as of October 14, 2026, and run on your own machine:
| Agent | Maker · licence | Latest release | GitHub stars | Where you talk to it | Standout | Pick it when |
|---|---|---|---|---|---|---|
| OpenClaw | Community · MIT | v2026.10.3 (Oct 13, 2026) | ~392,000 | Discord, iMessage, Slack, Teams, Telegram, WhatsApp + 20 more, native apps | State, memory and credentials stay on your hardware; models and harnesses (Claude, Codex, local) are swappable plugins | You want one assistant across all your chats and devices |
| Hermes Agent | Nous Research · MIT | v0.22.0 (Oct 14, 2026) | ~254,000 | Telegram, Discord, Slack, WhatsApp, Signal, CLI/TUI, desktop | Built-in learning loop: creates skills from experience and improves them; cron scheduler; 7 execution backends | You want the agent to get better at your recurring tasks |
| QwenPaw | Alibaba AgentScope · Apache-2.0 | v2.2.1 (Sep 2026) | ~36,000 | DingTalk, Lark, WeChat, Discord, Telegram, iMessage, QQ, desktop | Own QwenPaw-Flash models (2B–35B-A3B) with a local runtime; kernel-level sandbox and tool guard | You want zero API cost, or you work in Chinese-market channels |
| AnythingLLM | Mintplex Labs · MIT | v1.18.0 (Oct 13, 2026) | ~67,000 | Desktop app, web (Docker), embeddable widget | Document workspaces + agents, MCP, scheduled tasks, model routing | Your “agent” is mainly chat over your files |
Stars and releases read from GitHub on October 14, 2026. For hosted agents (OpenAI dots, Meta Muse, Manus Cue), see our best personal AI agents ranking.
Picks by situation
One assistant in every chat app: OpenClaw. It is the most adopted self-hosted agent by a wide margin and installs with one command on macOS, Linux or Windows (it provisions Node.js itself). Its design point is ownership: memory and credentials live on your machine, and you can swap the model or the coding harness without rebuilding anything. It suits people who want an assistant that runs errands across email, calendars, code and chats around the clock.
An agent that learns your routines: Hermes Agent. Hermes turns completed work into reusable skills and refines them as it uses them, keeps agent-curated memory with full-text session search, and runs scheduled jobs that deliver to any connected platform. Native Windows is supported via PowerShell, and its execution backends include local, Docker, SSH, Modal and Daytona, so the agent’s environment can hibernate in the cloud when idle.
Free, local and locked down: QwenPaw. QwenPaw is the only one of the four that ships its own agent-tuned small models (QwenPaw-Flash 2B, 4B, 9B, 27B and 35B-A3B) and a built-in local runtime, so it works with no API key. Security is on by default: sandbox, tool guard, file guard, skill scanner and access policies block dangerous commands before they run.
Chat with your files, plus some automation: AnythingLLM. Less of an always-on assistant, more of a private workspace: drag in documents, chat with citations, then add agents, MCP tools and cron-style scheduled tasks. Use the desktop app for one person and Docker for a multi-user team.
What to set up before you let it run unattended
- A separate OS user, container or VM for the agent — never your main admin account.
- Secrets in the agent’s credential store, not pasted into chats or config files that sync.
- An allowlist of who can message it; a public bot that can run shell commands is a remote shell.
- Approval for destructive actions (deleting files, sending email, spending money).
- A model budget. A busy always-on agent on a frontier model can burn millions of tokens a day; route routine work to a cheap model. See current API prices.
Projects we left out: Letta (last release May 14, 2026) and Khoj (last release a March 2026 beta) have slowed; check their activity before you build on them. Compare frameworks for building your own agent in most popular AI agent frameworks.
Last verified: October 14, 2026.