Every AI coding tool remembers what you built. None of them remember why. Until now.
The code is always saved. The thinking behind it wasn't. Lodestar commits your reasoning to your repo — decisions, dead ends, and direction — so every session starts where the last one left off.
Works with Claude Code, Cursor, Windsurf, or whatever you use next.
Lodestar by Kylex · Free forever · kylex LLC
lodestar start # Load context from your last session
lodestar save # Mid-session checkpoint
lodestar end # Done for the day — save and commitThat's it. No paths needed if you're in your project directory. Works from the terminal or just tell your AI — "lodestar start", "lodestar save", "lodestar end".
Lodestar doesn't produce a git log summary. It captures the thinking behind your code:
- Decisions & rationale — what was decided and why, not just what changed
- Dead ends — what was tried and abandoned, so the next session doesn't retry it
- Patterns — conventions established in the codebase
- Open questions — blockers and unresolved issues
- Direction — where to pick up next, written at peak clarity
This context is committed to your repo and portable across tools and machines.
When Claude Code starts, Lodestar prints a 5-line briefing automatically — no commands needed:
═══════════════════════════════════════════════
Lodestar · my-app · yesterday
═══════════════════════════════════════════════
Where you left off:
→ Wire up auth middleware in src/server/api/trpc.ts
→ Add users table to Drizzle schema
→ Run tRPC + Drizzle integration test
Last rejected: Prisma ORM — 2-3s cold starts on Vercel
═══════════════════════════════════════════════
Full session context → lodestar review
═══════════════════════════════════════════════
10 seconds to reorient. Zero interaction required.
Run lodestar review for the full picture — build status, architecture diagrams, decisions, and more:
lodestar review # Open dashboard in browser
lodestar review --diff # Show changes vs last session (Pro)- Node.js 20+
- Git — your project must be a git repository
- An AI provider account (one of):
git clone https://github.com/Kgard/LodeStar.git
cd LodeStar
npm install
npm run build
npm linklodestar helplodestar initThe wizard will:
- Ask which AI provider you use (Anthropic, OpenAI, or Ollama)
- Walk you through API key setup
- Validate your key
- Auto-configure Claude Code, Cursor, and Windsurf
- Offer to install git hooks (auto-updates on commit)
- Offer to enable the terminal summary (SessionStart hook for Claude Code)
- Bootstrap your existing project if it has code
That's it. No .env files, no manual config.
| Command | What it does | Cost |
|---|---|---|
lodestar start |
Load previous session context | Free |
lodestar save |
Mid-session checkpoint (synthesis) | ~$0.003 |
lodestar end |
Synthesize + commit — end session | ~$0.01 |
lodestar review |
Open project dashboard in browser | Free |
lodestar summary |
Print 5-line terminal briefing | Free |
lodestar bootstrap |
Capture existing project structure | Free |
lodestar hooks |
Install git hooks | Free |
lodestar init |
First-time setup | Free |
[path] is not required if you are in the project directory.
Aliases: lodestar load = start, lodestar synthesize / sync = save.
- Captures uncommitted changes and committed changes since last synthesis
- Detects package.json changes and project brief (CLAUDE.md) changes separately
- Sends to your configured AI provider with a synthesis prompt
- AI extracts decisions, patterns, dead ends, and open questions
- Writes context to your project root (with history rotation)
- (
lodestar endonly) Commits the context file
- Reads saved context from your project
- If no context exists, auto-bootstraps from your project structure (free, no LLM)
- Returns decisions, patterns, and where to pick up
lodestar hooks # Install
lodestar hooks --remove # Remove- post-commit — lightweight feature status update (no LLM, free)
- pre-push — commits context file if it has changes (no LLM, free)
LLM is only called when you explicitly run save or end.
| Provider | Default model | API key | Cost per synthesis |
|---|---|---|---|
| Anthropic | claude-sonnet-4-6 |
Required | ~$0.003 (save) / ~$0.01 (end) |
| OpenAI | gpt-4o |
Required | ~$0.003 (save) / ~$0.01 (end) |
| Ollama | llama3.2 |
Not needed | Free (runs locally) |
Model routing: save uses a faster model (Haiku/gpt-4o-mini), end uses a stronger model (Sonnet/gpt-4o) for deeper rationale extraction.
| Tool | Setup | Session start |
|---|---|---|
| Claude Code | Auto-configured by lodestar init |
Terminal summary fires automatically |
| Cursor | Auto-configured + .cursorrules generated |
Say "lodestar start" in chat |
| Windsurf | Auto-configured by lodestar init |
Say "lodestar start" in chat |
| Any MCP-compatible tool | Add server entry to MCP config | Call lodestar_load tool |
Everything in Free, plus:
- 30-day session history timeline
- Session diff — compare any two sessions
- Team sharing — shareable review URL
- AI summary across sessions
- Hosted synthesis — no API key needed
- Mid-session checkpoints
$9.99/month · Join early access at kylex.io
Do I need to install Lodestar in every project? No. Install once, use on any git project.
What does it cost? Lodestar is free. The AI API call costs ~$0.003-0.01 per synthesis using your own key. Ollama is completely free.
Does it work with my AI coding tool?
If your tool supports MCP servers, yes. lodestar init auto-configures Claude Code, Cursor, and Windsurf.
Do I need to pass a path every time? No. Run from your project directory — no arguments needed.
What if I don't have an API key? Use Ollama — free, local, no key needed.
What about large diffs? Lodestar prioritizes architecture and config files over templates and CSS. High-priority files are always included. Low-priority files are dropped first if the budget is exceeded.
Proprietary — kylex LLC
Keep the flow state. Even when the session ends. kylex.io