AI Provider Support

Choose between Gemini, local models, or ChatGPT/Claude for timeline generation

Overview

Dayflow supports three AI provider options, each with different trade-offs for privacy, quality, and performance. You choose which provider to use based on your priorities.

All providers generate the same timeline cards — only the processing pipeline and privacy model differ.

Provider Comparison

Provider	Privacy	Quality	Speed	LLM Calls	Cost
Gemini	Cloud (BYO API key)	Excellent	Fast	2 per batch	Free tier available
Local Models	100% offline	Good	Moderate	33+ per batch	Free
ChatGPT/Claude	Cloud (paid subscription)	Best	Fast	4-6 per batch	$20/month

Gemini (Cloud)

How It Works

Gemini uses Google’s native video understanding API to process your screenshots efficiently:

1. Upload + Transcribe (1 LLM call): Batch of screenshots is uploaded and analyzed
2. Generate Cards (1 LLM call): AI generates timeline cards from the analysis

Total: 2 LLM calls per batch

Privacy Considerations

• Data sent to Google: Screenshot batches are uploaded to Google’s Gemini API
• Bring your own API key: You control the API key and billing
• Paid Services data handling: If you enable Cloud Billing on any Gemini project, Google does not use your prompts/responses to train models (see Gemini API Terms)
• Abuse monitoring: Google logs prompts for policy enforcement, even under Paid Services

EEA/UK/Switzerland users: Paid-style data handling applies by default, even without billing enabled.

Setup

1. Get a Gemini API key: https://ai.google.dev/gemini-api/docs/api-key
2. Open Dayflow → Settings → AI Provider
3. Select Gemini
4. Enter your API key

See Gemini Configuration for detailed setup instructions.

Model Selection

Dayflow supports multiple Gemini models:

• Gemini 1.5 Flash (default): Fast and cost-effective
• Gemini 1.5 Pro: Higher quality, slower, more expensive
• Gemini 2.0 Flash: Latest model with improved reasoning

Configure in Settings → AI Provider → Model.

Fallback to Gemma

If Gemini’s video API fails (e.g., rate limits, API errors), Dayflow automatically falls back to Gemma (a text-only model) to extract frame descriptions. This increases LLM call count but ensures timeline generation continues.

---

Local Models (Ollama / LM Studio)

How It Works

Local models process screenshots entirely on your Mac using a local inference server:

1. Extract 30 frames from the screenshot batch
2. Describe frames (30 LLM calls): Each frame is analyzed individually
3. Merge descriptions (1 call): Combine frame descriptions into observations
4. Generate title (1 call): Create a title for the observation
5. Merge check (1 call): Decide if observations should be combined
6. Merge cards (1 call): Finalize timeline cards

Total: 33+ LLM calls per batch

Privacy Considerations

• 100% offline: All processing happens on your Mac
• No data leaves your machine: Screenshots, prompts, and responses stay local
• Model downloads: Models are downloaded once from Ollama/LM Studio and cached locally

Local models are the most private option — no cloud services involved.

Setup

Option 1: Ollama

1. Install Ollama: https://ollama.com/
2. Pull a vision model (e.g., llama3.2-vision):
   ollama pull llama3.2-vision
   
3. Start Ollama server:
   ollama serve
   
4. Open Dayflow → Settings → AI Provider
5. Select Local (Ollama/LM Studio)
6. Enter endpoint: http://localhost:11434 (default Ollama port)

Option 2: LM Studio

1. Download LM Studio: https://lmstudio.ai/
2. Download a vision-capable model from the LM Studio UI
3. Start the local server (default: http://localhost:1234)
4. Open Dayflow → Settings → AI Provider
5. Select Local (Ollama/LM Studio)
6. Enter endpoint: http://localhost:1234

See Local Models Configuration for detailed setup.

Performance & Battery

• GPU-heavy: Local inference uses Apple Silicon GPU acceleration
• Battery drain: Expect faster battery drain during processing
• Recommendation: Keep your Mac plugged in for long capture sessions
• Quality: Open-source models may underperform cloud models on complex contexts

Local models trade quality and battery life for maximum privacy. For best results, use plugged-in and test different models.

---

ChatGPT / Claude (CLI-Based)

How It Works

ChatGPT/Claude use CLI tools to batch-process screenshots with frontier reasoning models:

1. Extract frames (every 60s)
2. Batch describe (10 frames per call): Describe frames in batches
3. Merge segments (1 call): Combine descriptions into observations
4. Generate Cards (1 call): Create timeline cards

Total: 4-6 LLM calls per batch

Privacy Considerations

• Data sent to OpenAI/Anthropic: Screenshots are processed by OpenAI (ChatGPT) or Anthropic (Claude)
• Subscription required: You must have an active paid subscription:
  • ChatGPT Plus/Pro: $20+/month
  • Claude Pro: $20/month
• No offline mode: Requires active internet connection

Setup

ChatGPT (Codex CLI)

1. Install Codex CLI: https://github.com/openai/codex
2. Sign in with your ChatGPT Plus/Pro account
3. Open Dayflow → Settings → AI Provider
4. Select ChatGPT/Claude
5. Choose Codex as the CLI tool

Claude (Claude Code)

1. Install Claude Code: https://docs.anthropic.com/en/docs/claude-code
2. Sign in with your Claude Pro account
3. Open Dayflow → Settings → AI Provider
4. Select ChatGPT/Claude
5. Choose Claude as the CLI tool

See ChatGPT/Claude Configuration for detailed setup.

Quality

ChatGPT and Claude provide the highest quality timeline summaries:

• Best-in-class narrative quality: Frontier reasoning models excel at context understanding
• Nuanced insights: Better at distinguishing similar activities
• Consistent formatting: More reliable card structure

If quality is your top priority and you already have a ChatGPT Plus or Claude Pro subscription, this is the best option.

---

Processing Pipeline Comparison

Gemini

Efficiency: 2 LLM calls per batch
Strengths: Fast, cost-effective, leverages native video understanding
Trade-offs: Data sent to Google

Local

Efficiency: 33+ LLM calls per batch
Strengths: 100% offline, maximum privacy
Trade-offs: GPU-intensive, battery drain, variable quality

ChatGPT/Claude

Efficiency: 4-6 LLM calls per batch
Strengths: Best quality, frontier reasoning models
Trade-offs: Requires paid subscription, data sent to OpenAI/Anthropic

---

Choosing the Right Provider

I want maximum privacy and don't mind slower processing

→ Local Models (Ollama/LM Studio)

Best for: Privacy advocates, offline workflows, users with powerful Macs

I want a good balance of quality, speed, and cost

→ Gemini

Best for: Most users, especially those comfortable with Google’s data handling under Paid Services

I want the absolute best quality and already have a ChatGPT/Claude subscription

→ ChatGPT/Claude

Best for: Users prioritizing quality, existing Plus/Pro subscribers

I don't trust any cloud provider with my data

→ Local Models

Only local models keep 100% of your data on your machine.

Switching Providers

You can switch providers at any time in Settings → AI Provider. Existing timeline cards remain unchanged — only future batches use the new provider.

Next Steps

Gemini Setup

Configure Gemini with your API key

Local Models Setup

Set up Ollama or LM Studio for offline processing

ChatGPT/Claude Setup

Configure CLI-based inference

Privacy Overview

Learn about data handling and privacy