Notion Inbox Agent

Status: Aggressive Building & Beta — Executive Function Prosthetic for processing raw notes into ranked, structured tasks.

The Problem

You capture ideas fast, but processing them is slow. The "Capture-Process Gap" creates a graveyard of unread notes. You need a Ruthless Funnel, not another content factory.

The Solution

An AI pipeline that decouples capture from decision:

1. Route → Classify notes into projects (metadata.py)
2. Rank → Score importance/urgency/impact (ranking.py)
3. Filter → Skip low-value notes (confidence threshold)
4. Enrich → Analyze high-impact ideas (enrichment.py)
5. Store → Create structured Notion tasks (task.py)

Core Philosophy: Every input is noise until proven otherwise. Expensive compute only for high-leverage ideas.

Quick Start

# Setup with uv
uv venv
.venv\Scripts\Activate.ps1

# Install with dev dependencies
uv pip install -e ".[dev]"

# Configure .env (see .env.example)
NOTION_TOKEN=secret_xxx
GOOGLE_API_KEY=xxx

# Run
python run.py

# Run tests
pytest

Architecture

Note → MetadataProcessor → RankingProcessor → EnrichmentProcessor → TaskManager → Notion
         (classify)          (score)            (analyze)           (create)

Key Design:

• Separation of concerns: Ranking (classification) is decoupled from Enrichment (generation)
• Configurable models: Support for Gemini/Gemma with automatic format handling
• Confidence scoring: Flags ambiguous notes for human review

Testing

pytest                  # All tests
pytest -m integration   # Slow integration tests (requires API keys)

Current Focus

• Tuning ranking prompts to match user mental models
• Confidence score calibration
• Field testing on live inbox

Dependencies

See pyproject.toml for the list.