Two open-source always-on agents matter in October 2026: OpenClaw (MIT, roughly 393,000 GitHub stars, a gateway daemon with 20-plus chat channels) and Hermes Agent (MIT, roughly 256,000 stars, a gateway with a built-in scheduler and a learning loop). The rest are small, useful or archived. Huginn and n8n were always-on before the word existed. None of them solve identity and permissions for you.
Part of the Always-on agents series, post 13 of 33.
I have a rule for this kind of comparison: read everything on one day, with one list of questions, and write down the star count with the date, because by the time you read this the numbers will have moved. Everything below was read on 20 October 2026. What keeps it alive? Can it run a free local model? What machine does it want? What licence? And the question that matters most to me after a year of running jobs unattended: what happens to the secrets.
One more thing before the table. I have not run every project here. This post is the reading; the OpenClaw and Hermes setup guides later in the series are where I install them on fresh machines and report what broke. Where I say I would or would not run something, that is a judgement from the code and the documentation, and I say so.
The table, as read on 20 October 2026
| Project | What keeps it alive | Free local models | Wants to run on | Licence | Stars |
|---|---|---|---|---|---|
| OpenClaw | A gateway daemon: sessions, tools, events, channels. Installed with openclaw onboard --install-daemon | Yes, Ollama through its native API, plus LM Studio | macOS (the Mac mini favourite), Linux, Windows via WSL2 or PowerShell, Docker | MIT | 392.8k |
| Hermes Agent | A single gateway process plus a built-in cron scheduler | Yes, through its own-endpoint provider, so Ollama or anything OpenAI-compatible | Linux, macOS, WSL2, native Windows, Docker, a $5 VPS | MIT | 255.5k |
| n8n | A workflow server with triggers, schedules and webhooks | Yes, Ollama nodes | Docker, any Linux host | Fair-code, Sustainable Use License | 208.1k |
| Huginn | A Rails app plus a background worker; agents pass events along a graph | Not a model framework; call one over HTTP | A Linux server with MySQL or Postgres | MIT | 50.1k |
| agent-compose | A long-lived daemon that builds, runs and schedules isolated runtimes; cron, interval, event and webhook triggers | Through OpenAI-compatible endpoints such as Ollama or vLLM | Docker by default, Kubernetes via Helm, microVMs with KVM | AGPL-3.0 | 272 |
| ClawCode | A Claude Code plugin with heartbeat and a nightly memory consolidation; /agent:service install for 24/7 | No, it is Claude Code | launchd, systemd, WSL2 | MIT | 64 |
| Clank | A local-first gateway daemon with cron; archived and read-only | Yes, Ollama, LM Studio, llama.cpp, vLLM auto-detected | Windows, macOS, Linux (Node 20+) | Apache-2.0 | 8 |
| CLU | A daemon loop over a SQLite task queue with heartbeat checks and cron | Yes, LM Studio, Ollama, vLLM | Windows, Linux, macOS | MIT | 6 |
| OpenClaw Scheduler | A durable SQLite scheduler that runs beside the OpenClaw gateway, with retries and approval gates | Whatever OpenClaw uses | launchd, systemd, PM2, WSL2 | MIT | 0 |
Stars are what GitHub showed me that day, rounded the way GitHub rounds them. They measure attention, not quality, and I include them because the gap between the top two and everyone else is the most important fact in the table.
The two that matter
OpenClaw
OpenClaw is the project that made always-on agents a consumer phenomenon this year, and the reason so many Mac minis were bought as agent hosts. Its own description is the AI that really does things, and the architecture is a gateway: one long-lived process that owns your sessions, tools, events and the connections to Discord, iMessage, Slack, Teams, Telegram, WhatsApp, Signal, Google Chat and twenty-odd more. The CLI, the terminal UI and the control UI all talk to that gateway. It wants Node.js 24.16 or newer.
Two things stand out to me. First, it treats local models as first class: the documentation is explicit that you should point it at Ollama's native API, not the OpenAI-compatible endpoint, because the compatible one breaks tool calling and models start printing raw tool JSON as text. That is the kind of detail you only write down after being burned by it, and it made me trust the docs more. Second, the project had a public security reckoning in early 2026, which I go through in the permissions post. The advice that came out of it is the advice I would give anyway: run it on a machine that is not your main one, and never expose the gateway to the internet.
Hermes Agent
Hermes Agent, from Nous Research, calls itself the agent that grows with you, and the growing is the point: it writes skills from past work, refines them as it uses them, and searches its own earlier conversations. Architecturally it is also a single gateway process, reachable from Telegram, Discord, Slack, WhatsApp, Signal, email and Home Assistant, with a cron scheduler built in so unattended jobs do not need an external timer.
What draws me to it for business use is how little it asks for. The install is one line, it runs on Linux, macOS, WSL2 or native Windows, and the documentation is written for a small VPS rather than a workstation. Model providers are Nous's own portal, OpenRouter, OpenAI or your own endpoint, and that last option is how you run it against Ollama for nothing. The learning loop is the thing I most want to test and most want to watch: an agent that rewrites its own skills is an agent whose behaviour can drift, and I want to see where the review step is before I let it near anything that acts.
The small ones worth knowing about
- agent-compose is the one I keep thinking about. It is docker-compose for agents: a daemon that gives each agent an isolated runtime and runs it on cron, interval, event or webhook triggers. It drives the coding CLIs (Codex, Claude Code, OpenCode, Pi) rather than being an agent itself. AGPL, which matters if you build a product on it.
- ClawCode turns Claude Code into a persistent agent with a heartbeat and a nightly memory consolidation it calls dreaming, at 3am. Crons only run while Claude Code is open unless you install its service. It is a plugin, not a runtime, so it inherits Claude Code's cost model.
- CLU is a coding daemon: a loop over a SQLite queue, heartbeat checks for compile errors and new files, cron schedules, and a local dashboard. Six stars and honest about it.
- Clank billed itself as the open-source OpenClaw alternative and is now archived and read-only. I list it because people still recommend it without checking.
- OpenClaw Scheduler is a sidecar that keeps OpenClaw's scheduled jobs in SQLite, separate from the gateway, with retries and approval gates. Zero stars and a good idea: your scheduler should survive your agent being unhealthy, which is exactly how I split my own publishing job from the thing it publishes.
The workflow tools that were always-on before the word
Huginn was created in 2013 and its tagline has not needed to change: create agents that monitor and act on your behalf. Your agents are standing by. It is a Rails app with a background worker; agents emit and consume events along a graph. It has no model in it, and that is a feature. n8n is the modern equivalent, fair-code rather than open source in the strict sense, with a Docker one-liner and native nodes for calling models, including Ollama. Both have run unattended on thousands of servers for years, which no project in the first table can say.
I made this point in the definition post and I will keep making it: if the decision can be written as a rule, write the rule. Huginn and n8n are where rules live. Add a model node only where there is a judgement, and the rest of the system stays debuggable.
Which one I would hand to a non-engineer
None of the agent frameworks, not yet. Every one of them is installed from a terminal, configured in files, and secured by things like firewall rules and reverse proxies. OpenClaw has native apps and a control UI, and it is still a project for someone who is comfortable when a daemon stops. If a friend who does not code asked me, I would say: n8n in Docker on a small server, with someone technical setting it up once, and a vendor product for anything that touches your email or money. That is not a criticism of the projects. It is where the category is.
Which one I would run for a business
Hermes Agent on a small VPS, under a dedicated user, with a systemd service, a narrow set of tokens and a nightly backup of its memory directory. It is the one whose documentation assumes the deployment I would actually do. OpenClaw if the team lives in chat and wants the agent in twenty channels, on its own machine, gateway off the public internet. agent-compose if the job is a team of coding agents with schedules and isolation. For all three I would write the identity rule before the first job, which brings me to the gap.
What none of them solve
Identity and permissions. Every project in the table will happily run with your personal tokens in a config file, and most tutorials tell you to do exactly that. The vendors' pitch this autumn, from OpenAI's dots to Microsoft's Autopilot, is precisely that the agent gets its own identity, its own credentials and its own access scope. Open source gives you the loop and leaves the identity to you.
My rule, learned the hard way when a scheduled job of mine failed because my own login had expired: the agent never logs in as me. It gets a service account with the narrowest permissions that still let it work, its secrets live in a file only its user can read, and anything that spends money has a cap I set outside the agent. None of the frameworks stop you doing this. None of them do it for you either.
How I read a repository before trusting it
- Does the gateway listen on the network by default? If yes, the first task is turning that off, and the project should say so on page one.
- Where do secrets go? A config file in the home directory is normal. A config file that gets committed, synced or printed in logs is how keys leak.
- How did it handle its last security report? A fixed release with a date and a CVE is a good sign. Silence is a bad one.
- Is it still moving? Clank is archived. Check before you build on anything.
- Can it run a free model? Not because free is the goal, but because an agent that can only talk to one paid API has a single point of failure and a bill you do not control.
The setup guides for OpenClaw and Hermes Agent follow in this series, each on a fresh machine, each with what actually went wrong. If you want to see the always-on agent I run for social video, which is none of the above, it is here.
Sources
- GitHub, openclaw/openclaw (accessed October 20, 2026)
- OpenClaw docs, Ollama provider (accessed October 20, 2026)
- GitHub, NousResearch/hermes-agent (accessed October 20, 2026)
- GitHub, n8n-io/n8n (accessed October 20, 2026)
- GitHub, huginn/huginn (accessed October 20, 2026)
- GitHub, chaitin/agent-compose (accessed October 20, 2026)
- GitHub, crisandrews/ClawCode (accessed October 20, 2026)
- GitHub, ClankLabs/Clank (archived) (accessed October 20, 2026)
- GitHub, Continuous-Learning-Utility/clu (accessed October 20, 2026)
- GitHub, amittell/openclaw-scheduler (accessed October 20, 2026)
Written by Lera Leonteva. Star counts and licences are as GitHub displayed them on the access date; they will have moved by the time you read this.