leanpipe

CLI output filter for AI agents — strip noise, keep signal, save tokens.

kubectl get pods -o yaml | leanpipe filter kubectl

AI coding agents waste tokens on verbose CLI output. A single kubectl get -o yaml or git diff can burn 10k+ tokens. leanpipe sits in between, stripping noise and passing only what matters.

Install

pip install git+https://github.com/yunaremaia/leanpipe.git
# or
uv tool install git+https://github.com/yunaremaia/leanpipe.git

Quick Start

Pipe mode

kubectl get pods -o yaml | leanpipe filter kubectl
git diff | leanpipe filter git
pytest -v 2>&1 | leanpipe filter pytest

Run mode

leanpipe run -- git status
leanpipe run --preset kubectl -- kubectl describe pod my-pod

Aggressiveness levels

| leanpipe filter kubectl -l 1  # conservative
| leanpipe filter kubectl -l 2  # balanced (default)
| leanpipe filter kubectl -l 3  # aggressive

Built-in Presets

Preset	What it strips	Use case
kubectl	managedFields, annotations, timestamps, generation, uid	Agent needs status, not K8s internals
docker	layer hashes, build step noise	Agent sees container state, not build noise
git	diff headers, index lines	Agent focuses on actual changes
pytest	fixture setup dashes, PASSED lines	Agent sees failures only
npm	deprecation warnings, funding notices	Agent sees install results
find, ls	permissions, sizes (unless kept)	Agent sees paths
grep	extra context beyond match	Agent sees matching lines

Custom Plugins

# ~/.config/leanpipe/plugins/terraform.yaml
name: terraform
description: "Strip drift details"
patterns:
  - name: "Drop changed lines"
    match: "^\\s*~"
    action: drop_line

leanpipe run --preset terraform -- terraform plan

JSON & Streaming Output (CI/agent consumption)

| leanpipe filter kubectl --json
# {"output": "...", "saved_chars": 12000, "saved_pct": 87.5}

| leanpipe filter kubectl --format jsonl
# {"line": 1, "content": "items:"}
# {"line": 2, "content": "- name: my-pod"}

Validation

Inspired by lowfat (Rust), which saved 91.8% of tokens in personal use. leanpipe brings the same pattern to Python — where AI agents (Claude Code, Codex, OpenCode, Hermes) most commonly run.

License

MIT

leanpipe