Skip to content
Habakkuk
Install

Compare

How Habakkuk compares, including where it loses

There are good agents out there already, and you may be running one. This page puts Habakkuk next to them as plainly as we can, so you can decide whether it is worth your evening.

See the tableMove from another agent

Why you might pick Habakkuk

Next to the other open-source agents

Hermes Agent and OpenClaw are the two you are most likely weighing us against. Both are free and open source, like Habakkuk, and both are good. The table repeats what each project’s own documentation says.

HabakkukHermes AgentOpenClaw
What you installOne program, written in Go. The install script downloads it and nothing else needs to be on the machine.A Python program. An install script sets it up, and the machine needs Git, curl and tar.A Node.js program. Install it with a script, npm or Docker; the machine needs Node 24 or newer.
Chat apps27, with WhatsApp three ways: an agent inside WhatsApp, a linked number, or the Business API.More than 20. WhatsApp by linked number or the Business API.More than 30, the most of the three. WhatsApp by linked number.
If it stops in the middle of a jobThe job is saved to disk before work starts. The next start picks it up from the saved conversation, and you get the reply once.A finished reply is saved and delivered after a crash. An interrupted session resumes once after a restart.After a restart the agent gets one message telling it to check what was in progress and carry on. Its docs say effects are not guaranteed exactly once.
Before a risky stepOn a new install it asks you before running any command, with Approve and Deny buttons in your chat.On a new install a second model judges each dangerous command. It lets the safe ones through, refuses the worst and asks you about the rest.Approvals are there, but on a standard install commands run without asking until you turn them on.
After it reads a web page or an emailThe job is marked as having read outside content. From then on only a person can approve its commands, and it cannot change what the agent remembers about you.Scans project files and memory writes for known injection patterns.Wraps outside content in markers that tell the model not to trust it.
Spending limitsDaily, monthly and per-job limits in dollars, for the whole install and for each agent, with a warning at 80%.Shows what you spent, and can limit the number of steps and the running time. We found no limit in dollars.Shows usage and cost. Its docs say its token budgets are not a billing cap.
Several agentsYes. Each has its own model, tools and memory. They hand work to each other and share a task board.Yes. Separate profiles that pass work to each other through a task board.Yes. Separate agents that can message each other, with a work board.
MemoryText files you can open and edit. Every change goes into a history, and you can undo any of them.Two text files you can read and edit. We found no history or undo.Text files you can read and edit. We found no history or undo for memory.
Learning from its own workWeekly retros. It goes through its own records, proposes changes, and checks the next week whether they helped.Writes and improves its own skills, and tidies away the stale ones.Tidies its memory every night, and writes and improves its own skills, keeping the old versions.
Skills other people madeInstalls skills in the common SKILL.md format from GitHub or a link. Each one is scanned and stays off until you approve it. There is no public directory yet.A hub with well over a thousand skills, drawn from several sources. Installs are scanned.ClawHub, a public directory anyone can publish to. Uploads are virus-scanned.
UpdatesOne command. The download is checked against a signed list, and one command puts the previous version back.One command pulls the newest release.One command. It checks the files and goes back by itself if the new version fails the check.
Apps and screensA web dashboard, a desktop app and a terminal app.A web dashboard and a desktop app.A web control panel, Mac and Windows apps, and iPhone and Android apps.
Price and licenceFree. Apache-2.0.Free. MIT.Free. MIT.

Where the others are ahead

Habakkuk is the newest of the three. If any of these matter more to you than what is in the table, one of the others is the better choice today.

  • Far more people use them

    On the day we checked, OpenClaw had about 394,000 stars on GitHub and Hermes Agent about 262,000. That means more guides, more answers when you search for an error, and more people who hit your problem before you did. Habakkuk is new.

  • Bigger skill directories

    OpenClaw has ClawHub, and Hermes Agent has a hub with well over a thousand skills. Habakkuk installs the same kind of skill, but it has no directory of its own yet.

  • Phone apps

    OpenClaw has iPhone and Android apps. With Habakkuk, your phone app is WhatsApp, Telegram or whichever chat you connected.

  • A few more chat apps

    OpenClaw lists more than 30 chat apps against our 27. If yours is on their list and not on ours, that settles it.

Next to ChatGPT, Claude, Grok and Viktor

These are a different kind of product. A company runs them for you, on its computers, with its model. That is easier to start with, and for plenty of people it is the right answer. Habakkuk is for when you want the agent on your own machine, on your own chat apps, with the model you choose.

Where this comes from

We checked every line about another product against that product’s own website, documentation or source code on 10 November 2026. These projects move quickly, so some of it will go out of date. If you see something wrong, open an issue and we will fix the page.

Already running one of them? habakkuk migrate copies your agents, memories, skills and keys across. It shows you the plan first and changes nothing until you say yes.