footprint

Your Claude, learned locally.

Footprint watches your Claude Code sessions, fine-tunes a small local model (LoRA) on how you and Claude work in your projects, and serves it as an OpenAI-compatible API. When your Claude quota runs out, OpenCode (or Cursor, or Codex CLI) keeps working in the same style — fully offline, on your machine.

trace → chat with claude → collect → train → install → /footprint in OpenCode

How it works

1. Claude Code already logs every session (prompts, replies, tool calls) to ~/.claude/projects.
2. footprint trace arms a marker so only sessions from that point on become training data.
3. footprint collect parses those transcripts into chat-format training examples.
4. footprint train LoRA fine-tunes a small model (default: Qwen2.5-Coder-1.5B, 4-bit MLX) on them.
5. footprint install registers a launchd agent — the server runs itself at http://127.0.0.1:8399/v1, starts at login, restarts on crash. You never run it by hand.
6. Any OpenAI-compatible tool points at that URL and behaves like your Claude.

Requirements

• Claude Code installed and used at least once
• Python 3.9+
• Node.js 18+ (only if installing via npm)

footprint picks the backend automatically: MLX on Apple Silicon, PyTorch + PEFT everywhere else (uses CUDA when available; CPU works but trains slowly — prefer a GPU or drop FOOTPRINT_ITERS).

Install

macOS

npm install -g footprint-trace

or from source:

git clone https://github.com/Amanlabh/footprint.git
cd footprint
python3 footprint.py setup

Linux

npm install -g footprint-trace      # or the git clone above

First run installs the torch backend (one-time, ~2 GB). CUDA GPU picked up automatically.

Windows

Use PowerShell (Python 3 from python.org or the Microsoft Store):

npm install -g footprint-trace

Same as Linux: full trace/collect/train/serve via the torch backend. WSL2 also works.

Getting started

1. Arm tracing — before you start Claude.

footprint trace

Claude Code already logs every session; the trace marker makes footprint train only on sessions from this point on. (Skip this step — or delete ~/.claude/footprint-trace — to train on your entire history instead.)

2. Use Claude Code normally.

claude

Every session (your prompts, Claude's replies, tool calls) is captured automatically.

3. Collect and train.

footprint collect     # transcripts -> data/train.jsonl + data/valid.jsonl
footprint train       # LoRA fine-tune (~10 min, downloads ~1 GB base model first time)

4. Install the always-on server.

footprint install

This registers the server with your OS — launchd (macOS), systemd (Linux) or Task Scheduler (Windows) — so it starts at login, restarts if it crashes, and you never run it by hand. It also wires up OpenCode:

• provider footprint at http://127.0.0.1:8399/v1
• a /footprint <task> command inside OpenCode

5. Claude quota over? Keep going.

Open OpenCode and type:

/footprint fix the failing test in auth.py

Or point any OpenAI-compatible tool at the server:

Commands

Configuration

Troubleshooting

• no transcripts found — either you haven't used Claude Code in this project, or the trace marker is newer than all sessions. Chat first, or rm ~/.claude/footprint-trace.
• server not responding — macOS: /tmp/footprint-serve.log, launchctl kickstart -k gui/$UID/com.footprint.serve; Linux: journalctl --user -u footprint, systemctl --user restart footprint; Windows: schtasks /Run /TN footprint-serve.
• backend import fails (mlx_lm / torch) — re-run footprint setup.
• quality is rough — more data beats more iterations: keep tracing, re-collect, re-train. 1.5B is small; try FOOTPRINT_MODEL=mlx-community/Qwen2.5-Coder-7B-Instruct-4bit if you have ≥16 GB RAM.

Privacy

Everything stays on your machine. data/ (your transcripts) and adapters/ (weights trained on them) are gitignored — never commit or publish them.

License

MIT — see LICENSE.