Introduction
So you've decided to try Bitloops. Good call.
Here's what you need to know before diving in.
What is Bitloops?
Bitloops is a CLI that runs in the background alongside your AI coding agents. You install it, forget about it, and it quietly does two things:
It captures everything. Every prompt, every response, every file change, every decision your AI agent makes — recorded and linked to your git commits. Think of it as a flight recorder for AI-assisted development.
It builds intelligence. Bitloops parses your codebase into a structured knowledge graph — functions, dependencies, relationships — so your agents can query for exactly what they need instead of re-reading your entire repo every session. The longer you use it, the smarter it gets.
It's local-first (your code never leaves your machine), open source (Apache 2.0), and completely non-intrusive. Your workflow stays exactly the same.
How does it work?
The short version:
- You work with your AI agent as you normally would — ask questions, make changes, write code
- Bitloops captures Draft Commits in real time — recording the conversation, reasoning, and decisions as they happen
- You
git commitand those Draft Commits become Committed Checkpoints — permanent, immutable records tied to that commit - Next session, your agent (or a different one) can tap into all that accumulated context instead of starting from zero
The result: every AI session builds on the last one. Your codebase develops a memory.
Want the full architecture? See How Bitloops Works.
Why teams choose Bitloops
"My agent keeps forgetting everything." Every session starts from scratch. You re-explain your architecture, your conventions, your constraints. Bitloops fixes that — agents get persistent context that survives across sessions.
"I can't review AI-generated code properly." A PR diff tells you what changed, not why. With Committed Checkpoints, reviewers see the full reasoning chain: what was asked, what was considered, what was rejected, and why.
"We use three different agents and it's chaos." Claude Code, Cursor, and Copilot each build their own isolated understanding. No shared context. Bitloops gives them one unified knowledge store — same intelligence, regardless of which agent you're using.
"We're spending a fortune on tokens and can't tell what's working." Without observability, you can't optimize. Bitloops tracks which agents, which models, and which sessions produce results — so you can make informed decisions about your AI tooling.
"Our AI doesn't know our architecture." It's doing surface-level pattern matching, not structural analysis. Bitloops builds a real dependency graph from your code using Tree-sitter parsing, so agents understand blast radius, relationships, and impact before making changes.
Which agents does Bitloops work with?
All of them. Well, all the major ones:
| Agent | What Bitloops captures |
|---|---|
| Claude Code | Full transcripts — sessions, prompts, tasks, reasoning, tool use |
| Cursor | Shell commands, prompt submissions |
| GitHub Copilot | Tool use and code interactions |
| Codex | Session boundaries |
| Gemini | Multi-stage tool chains |
| OpenCode | Coming soon |
You can use one agent or all five. bitloops init lets you select which ones to connect — and you can always add more later.
The important part: it doesn't matter which agent writes the code. Every session, every agent, feeds into the same knowledge store. One source of truth.
What do I need?
- A git repo
- At least one AI coding agent
- About 2 minutes
That's it. No database to install, no cloud account to create, no config files to write. Bitloops bundles SQLite and DuckDB right in the binary.
Ready? The Quickstart is three commands.