Self-coaching · AI-assisted work
Fertilizer Loop
You finished another day stitching together prompts, agents, tools, and skills — and most of what you learned evaporated by morning. The fruit of every working day should become the fertilizer for tomorrow's systems, but only if you can see the day plainly. Fertilizer Loop captures the junctions between projects, prompts, agents, tools, skills, artifacts, friction, and follow-ups, then writes you a reviewable daily report — on your machine, without piping raw logs anywhere.
- License
- MIT open-source
- Runtime
- Python 3.11+
- Privacy
- Local-first opt-in providers
- Price
- $0 free, forever
What it is
A reflection layer for AI-assisted work — on your machine.
A local-first desktop app and CLI that records normalized events at the junctions where workflow quality is determined — prompts issued, agents invoked, tools and skills used, artifacts produced, friction encountered, decisions deferred — and turns them into a structured daily improvement report. Two editions ship together: Light runs a mock provider that never touches the network (safe for demos and screenshots), and Full adds local HTTP or OpenAI-compatible providers, opt-in only.
01
Private by construction
Raw events stay on your machine. Only minimized, sanitized summaries reach a provider, and only if you turn one on. The GUI stores the name of a secret's environment variable, never the secret itself.
02
Provider-pluggable
Default mode is mock and never calls the network. Point it at a local OpenAI-compatible endpoint (Ollama, llama.cpp, vLLM), or at a cloud endpoint — your call, configured per profile.
03
A real artifact every day
The report is not a dashboard you forget to open. It is a versioned Markdown file with executive summary, workstream map, repeated friction, high-leverage wins, and tomorrow's operating notes — reviewable, editable, exportable.
How it works
Capture, generate, review.
A three-step loop you can run from the GUI or the CLI — whichever fits the way you already work.
Capture
Events at the junctions
Ingest a JSONL of session events, drop in a quick note from the CLI, or let the GUI log what you're working on. Each event is normalized against a documented schema and stored locally.
Generate
A daily improvement report
Run fertilizer-loop report --date YYYY-MM-DD or click Generate in the Reports tab. The configured provider summarizes the day's junctions into an eleven-section Markdown report.
Review
Accept, edit, queue
Improvement proposals land in a review queue with approval gates — accept, edit, or reject. History is searchable; exports are JSON or Markdown. The system never changes your setup on its own.
Who it's for
If this sounds like you, it will fit.
- You are an AI-assisted developer who finishes most days knowing you learned something, but cannot say exactly what — and you want a written daily artifact that survives the night.
- You are an indie engineer running multi-agent workflows across several tools and want one place that ties prompts, agents, skills, and artifacts back to outcomes.
- You want a self-coaching habit — "did today's work make tomorrow easier?" — without piping raw logs to a third-party service to get it.
Plainly
What it does, and what it doesn't.
What it does
- Captures normalized events from your AI-assisted work into a local store
- Generates a structured daily improvement report from those events
- Runs a mock provider by default — no network calls, safe for demos
- Plugs into a local HTTP or OpenAI-compatible provider when you opt in
- Keeps an improvement backlog and review queue with human approval gates
- Ships both a desktop GUI and a scriptable CLI from the same package
What it doesn't
- It does not send raw private logs anywhere — only minimized summaries, and only if a provider is on
- It does not silently change your tools, scripts, or system settings — every action is gated
- It is not hidden keystroke monitoring — capture is explicit and inspectable
- It is not a project manager — it observes work, it does not plan it
- It does not assume a cloud account — you can run it forever in Light mode
It's free and open-source. Clone it.
Fertilizer Loop is MIT-licensed — no tiers, no account, no paid version held back. Clone the repo, install the package, and run fertilizer-loop gui or fertilizer-loop doctor to start the loop on your own machine.