Get started

Habakkuk is a personal AI agent in one Go binary. It talks to you over the terminal, a web dashboard, a desktop app and chat apps (WhatsApp, Telegram, Slack, Discord, Signal, Matrix, email, SMS, Microsoft Teams, Google Chat, iMessage and more). It uses tools, remembers what matters, runs work on a schedule, and can run several agents that hand each other work.

Every turn is a job with a lease in a SQLite queue. A kill -9 at any moment loses nothing: a restart picks up where it stopped.

Pick your goal#

Why Habakkuk#

• One binary. One static binary, no CGO and no runtime to install.
• Crash-proof by design. Every turn is a leased job. A dead process's work is reclaimed on restart, and interrupted tool calls are repaired, never blindly repeated.
• Storage that stays small. A hot control.db, a cold transcripts.db and content-addressed blobs, with retention on by default.
• Channels that stay connected. Every adapter runs under a supervisor with a heartbeat watchdog. A stuck socket is torn down and reconnected, not left hanging.
• Never answers twice. Redelivered messages are dropped, replies are delivered once, and a long reply resumes at the first unsent chunk after a crash.
• Any model. Claude, GPT, Gemini, DeepSeek, Grok, Mistral, Qwen, Bedrock, OpenRouter, local models and more, with a cheap model for quick messages if you want one.
• Config that catches typos. Keys Habakkuk does not know are reported, not silently accepted.

How it stays up#

• Durable turns. A message becomes a turn job. Workers claim jobs with a lease and heartbeat while running. Every transcript write checks the lease, so a worker that lost its job can never write.
• Fast crash recovery. Leases record host and PID. On start, Habakkuk reclaims jobs whose owner process is dead instead of waiting for the lease to expire.
• Transcript repair. A tool call whose result was never recorded gets an "interrupted, check state first" result, so side effects are not blindly repeated.
• Model resilience. Errors are classified. Transient ones retry with jittered backoff, rate limits rotate keys with cooldowns, and errors another model can fix fall back down the chain.
• Cache-stable prompts. The system prompt is frozen per session, tools are sorted, the time goes in the user turn, and transcripts are append-only.

Architecture explains the design in depth.

Next steps#