Skip to content
Habakkuk
Install
Newhabakkuk migrate brings your old agent across

Your own AI agent, keeping watch

An open-source agent that runs on your own machine. Message it on WhatsApp, Telegram or Slack. It remembers you, asks before anything risky, and finishes what it started, even after a crash.

Source on GitHubApache 2.0macOS, Linux, Windows, Docker

WhatsAppTelegramSlackClaudeGPTOllama

27 channels, and the model you choose

  • WhatsApp
  • Telegram
  • Slack
  • Discord
  • Signal
  • iMessage
  • Google Chat
  • Matrix
  • Mattermost
  • Claude
  • GPT
  • Gemini
  • Grok
  • Mistral
  • Ollama
  • OpenRouter

How it works

Running in about a minute

No YAML to write and no runtime to install. Not sure what to run next? Type hk on its own and it tells you.

Step 1

Install

One command for macOS, Linux or Windows, or a container image. A single binary, with hk as its short name.

Step 2

Run hk setup

Pick a model, pick where to talk (a WhatsApp agent is the easy one), and name your agent. It ends with a summary and three things to try.

Step 3

Say hi from your phone

It introduces itself, asks a few questions, and remembers the answers from then on.

curl -fsSL https://habakkuk.ai/install.sh | sh

Apple silicon and Intel. Verified against SHA256SUMS, installed to ~/.local/bin.

then hk setup

Coming from another agent? habakkuk migrate brings your agents, memories, skills and keys.

Other ways to install

Where you talk

It lives in the chats you have

No new app to learn. Your agent sends the morning brief, asks before it runs a command, and tells you when the work is done, in WhatsApp, Telegram, Slack, Discord, Signal, iMessage, email and more.

It keeps the thread. Reply to one message on WhatsApp, Telegram, Signal or Discord and it sees the quote. It also sees what else landed in the chat, such as a schedule’s result, so a short “yes” is read against the right message.

Habakkuk

in Telegram

  1. Today
  2. Your Tuesday

    Passport renewal is on today’s list. The venue still hasn’t replied; I’ll nudge them at noon. Rain after 4.

    08:00
  3. The server’s almost out of disk. Can you clean it up?08:02
  4. Old Docker images are using 18 GB. I’d like to run:

    docker image prune -a08:03
  5. You approved · 08:03
  6. Done. 18 GB freed; the disk is at 61%.08:04
An illustration. Your agent writes its own words.

Durable by design

Pull the plug. It carries on.

Every message becomes a job saved to disk before any work starts. If the process dies mid-task, from a power cut, a bad deploy or kill -9, the next start takes the job back, repairs the conversation and finishes. Nothing is lost, and nothing is done twice.

what happens when it crashes
  1. 09:14:02queuejob 4821 saved: “pull the repo and fix the test”
  2. 09:14:02workerjob 4821 leased by mini, pid 5120
  3. 09:14:05toolexec: go test ./... (running)
  4. 09:14:06hostprocess killed: kill -9
  5. 09:14:11startjob 4821: owner is gone, reclaimed
  6. 09:14:11repaircut-off call marked interrupted, not re-run blindly
  7. 09:14:19replysent once; duplicate dropped
An illustration of the recovery steps, not real log output.

Dashboard

See everything it does

Every Habakkuk serves a dashboard: chat, sessions, the task board, schedules, memory, usage and approvals. The same screens ship as a desktop app.

About the dashboard
127.0.0.1:8484/app/
The Habakkuk dashboard, Schedules pageThe Habakkuk dashboard, Schedules page

Recurring work, each slot run at most once, results sent to the chat that asked.

FAQ

Questions

Anything else is in the docs.

What does it cost?
Nothing. Habakkuk is free and open source. You pay your model provider for what the agent uses, or nothing at all with a local model. Quick messages go to a cheaper model automatically, every call’s cost is tracked, and spending caps stop a runaway bill.
Which models does it work with?
Claude, GPT, Gemini, DeepSeek, Grok, Mistral, Qwen, Kimi, GLM, Bedrock, OpenRouter, any OpenAI-compatible API, and local servers such as Ollama, LM Studio, vLLM and llama.cpp. You can also sign in with a ChatGPT or GitHub Copilot plan, and mix them: a cheap model for quick messages, a strong one for real work.
Where does my data live?
On the machine you run it on, in one folder of SQLite databases and Markdown files, readable only by you. No telemetry. It only talks to the model providers, chat apps and services you configure.
How is it different from a chatbot?
It acts. It runs commands, browses, reads your files and uses MCP tools, asking first for anything risky. It remembers you, messages you first on a schedule or when work finishes, and every turn is a durable job, so a crash never loses work in progress. Turn on retros and it also reviews its own week and proposes fixes.
Where should I run it?
Anywhere that stays on: a laptop, a spare machine, a small VPS or a container. It is one binary with no runtime to install. If it lives on a server, hk -r NAME runs any command there from your laptop. Rather not run a server? Habakkuk Cloud is planned; the waitlist is on the pricing page.
Is it really open source?
Yes, under the Apache License 2.0. The Habakkuk name and logo are trademarks and are not covered by the licence, so forks use their own name.

Set a watchman over your work

Install it, run hk setup, and say hi from your phone.

curl -fsSL https://habakkuk.ai/install.sh | sh

Free to run yourself. Hosted plans are on the way: see pricing.