Local-First Architecture

Rowboat is built with a local-first philosophy: your data stays on your machine. All processing, storage, and AI inference happen locally, giving you complete control and privacy.

Core Principles

Local Storage

All your emails, meetings, and notes are stored in ~/.rowboat/ on your machine

No Cloud Lock-in

Your knowledge graph uses standard markdown files - readable and portable

Privacy First

Data never leaves your machine unless you explicitly share it

You Own Your Data

Plain text files you can read, edit, search, and back up however you want

Data Storage

Directory Structure

Rowboat stores everything in a well-organized directory:

~/.rowboat/
├── agents/              # Custom agent definitions
├── gmail_sync/          # Synced email threads
│   ├── attachments/     # Email attachments
│   └── *.md             # Thread markdown files
├── fireflies_transcripts/ # Meeting transcripts from Fireflies
├── granola_notes/       # Meeting notes from Granola
├── knowledge/           # Your knowledge graph
│   ├── People/
│   ├── Organizations/
│   ├── Projects/
│   ├── Topics/
│   ├── Voice Memos/
│   └── .git/            # Version history
├── config/              # Application configuration
│   ├── models.json      # LLM provider settings
│   ├── note_creation.json # Strictness settings
│   └── google_oauth.json  # OAuth tokens (encrypted)
├── logs/
│   └── services.jsonl   # Service activity logs
└── runs/                # Agent execution logs

All files are human-readable markdown or JSON. You can browse, search, and edit them with any text editor.

Markdown Format

Notes use Obsidian-compatible markdown with frontmatter:

# John Doe

**Email:** [email protected]
**Organization:** [[Acme Corp]]
**Role:** Engineering Manager
**Aliases:** John, JD

## Context

- First met: 2026-01-15 (email thread about API integration)
- Working on: [[Project Atlas]]
- Interested in: distributed systems, Kubernetes

## Recent Interactions

### 2026-02-15: Meeting - Architecture Review

Discussed scaling challenges for the new API gateway...

### 2026-02-01: Email - Q1 Planning

Proposed timeline for Atlas launch...

Benefits:

Open format (works with Obsidian, VS Code, etc.)
Human-readable
Grep-able
Version controllable with git
Easy to backup

Version Control

Rowboat automatically tracks all changes to your knowledge base using Git:

cd ~/.rowboat/knowledge
git log --oneline

# Example output:
a3f5e9d Knowledge update (Rowboat)
b2c4d8e Knowledge update (Rowboat)
c1e3f7g Knowledge update (Rowboat)

How Versioning Works

export async function commitAll(
  message: string,
  author: string
): Promise<void> {
  const git = simpleGit(KNOWLEDGE_DIR);
  
  // Initialize repo if needed
  if (!fs.existsSync(path.join(KNOWLEDGE_DIR, '.git'))) {
    await git.init();
  }
  
  // Stage all changes
  await git.add('.');
  
  // Commit with timestamp
  await git.commit(message, {
    '--author': `${author} <${author}@rowboat.local>`
  });
}

Every batch of entity updates creates a commit, providing:

Complete audit trail
Ability to revert changes
Debugging information
Export capability

AI Model Configuration

Rowboat supports multiple LLM providers - you choose where to send requests:

Cloud Providers
Local Models
Self-Hosted

{
  "provider": {
    "flavor": "anthropic",
    "apiKey": "sk-ant-..."
  },
  "model": "claude-3-5-sonnet-20241022"
}

Supported providers:

Anthropic (Claude)
OpenAI (GPT-4, etc.)
Google (Gemini)
OpenRouter (access to many models)

When using cloud providers, only the prompt and conversation context are sent. Your stored files and notes never leave your machine.

{
  "provider": {
    "flavor": "ollama",
    "baseURL": "http://localhost:11434"
  },
  "model": "llama3.1:70b"
}

Benefits of local models:

100% private - nothing leaves your machine
No API costs
Works offline
Full control over model choice

Popular local options:

Ollama (easiest setup)
LM Studio
LocalAI
vLLM

Local models require significant compute resources. For knowledge graph building, we recommend at least 70B parameter models for good results.

{
  "provider": {
    "flavor": "openai-compatible",
    "baseURL": "https://your-inference-server.com/v1",
    "apiKey": "your-key"
  },
  "model": "custom-model-name"
}

Self-hosted options:

Deploy models on your own infrastructure
Use OpenAI-compatible APIs (vLLM, Ollama, etc.)
Full control over data flow
Can be on-premises or in your VPC

Configuration file: ~/.rowboat/config/models.json

OAuth & Credentials

Rowboat needs OAuth access for Gmail and meeting tools:

How OAuth Works

Initial authorization

You authorize Rowboat via OAuth flow (opens browser)

Token storage

Access and refresh tokens are stored in ~/.rowboat/config/google_oauth.json

Token refresh

Rowboat automatically refreshes expired tokens without re-authorization

Scope limiting

Only requests minimal required scopes (e.g., gmail.readonly)

OAuth Token Storage

{
  "access_token": "ya29.a0AfH6...",
  "refresh_token": "1//0gX7...",
  "scope": "https://www.googleapis.com/auth/gmail.readonly",
  "token_type": "Bearer",
  "expiry_date": 1709136000000
}

Security:

Tokens are stored with file permissions 0600 (owner read/write only)
Never logged or transmitted except to authorized APIs
Can be revoked at any time through Google account settings

Scopes Requested

Service	Scope	Purpose
Gmail	`gmail.readonly`	Read email threads for knowledge graph
Fireflies	API Key	Fetch meeting transcripts
Granola	API Key	Fetch meeting notes

Rowboat only requests read-only access. It never sends, modifies, or deletes your emails or meetings.

Data Portability

Export Your Data

Since everything is markdown and JSON:

# Backup entire knowledge base
tar -czf rowboat-backup.tar.gz ~/.rowboat/

# Export knowledge graph only
tar -czf knowledge-backup.tar.gz ~/.rowboat/knowledge/

# Copy to cloud storage (optional)
rsync -av ~/.rowboat/ /path/to/cloud/backup/

Use with Other Tools

Obsidian

Open ~/.rowboat/knowledge/ as an Obsidian vault

VS Code

Browse and edit notes with full markdown support

Command Line

Use grep, ripgrep, or any text tool to search

Example searches:

# Find all mentions of a project
grep -r "Project Atlas" ~/.rowboat/knowledge/

# List all people from a specific organization  
grep "Organization: Acme" ~/.rowboat/knowledge/People/*.md

# Search email threads
rg "API integration" ~/.rowboat/gmail_sync/

Privacy Considerations

What Stays Local

✅ Always local:

All synced emails and meetings
Your entire knowledge graph
Service logs
Configuration files
OAuth refresh tokens

What Gets Sent to LLM APIs

⚠️ When using cloud LLMs:

Agent prompts and instructions
Specific conversation context
Entities being processed in current batch

Batch processing limits exposure: Only the current batch (10 files) is sent to the LLM at a time, not your entire knowledge base.

❌ Never sent:

Your full email archive
Your complete knowledge graph
Unrelated notes or conversations

Using 100% Local Setup

For complete privacy:

Use local LLM (Ollama, LM Studio)
Disable telemetry (Rowboat has none by default)
Air-gapped option: Sync data externally, then disconnect

Open Source Benefits

Audit the Code

Review exactly how your data is processed

Self-Host Everything

Run Rowboat on your own infrastructure

Customize Freely

Modify agents, storage, or processing to your needs

No Vendor Lock-in

Standard formats mean you can migrate anytime

Source code: github.com/rowboat-ai/rowboat

Source Code Reference

Key implementation files:

apps/x/packages/core/src/config/config.ts - WorkDir and configuration paths
apps/x/packages/core/src/knowledge/version_history.ts - Git versioning
apps/x/packages/core/src/models/models.ts - LLM provider configuration
apps/x/packages/core/src/knowledge/google-client-factory.ts - OAuth handling

Getting Started

Core Concepts

Desktop App

Integrations

Use Cases

Configuration

Local-First Architecture

Core Principles

Local Storage

No Cloud Lock-in

Privacy First

You Own Your Data

Data Storage

Directory Structure

Markdown Format

Version Control

AI Model Configuration

OAuth & Credentials

How OAuth Works

Scopes Requested

Data Portability

Export Your Data

Use with Other Tools

Obsidian

VS Code

Command Line

Privacy Considerations

What Stays Local

What Gets Sent to LLM APIs

Using 100% Local Setup

Open Source Benefits

Audit the Code

Self-Host Everything

Customize Freely

No Vendor Lock-in

Source Code Reference

Build docs developers (and LLMs) love

Getting Started

Core Concepts

Desktop App

Integrations

Use Cases

Configuration

​Core Principles

Local Storage

No Cloud Lock-in

Privacy First

You Own Your Data

​Data Storage

​Directory Structure

​Markdown Format

​Version Control

​AI Model Configuration

​OAuth & Credentials

​How OAuth Works

​Scopes Requested

​Data Portability

​Export Your Data

​Use with Other Tools

Obsidian

VS Code

Command Line

​Privacy Considerations

​What Stays Local

​What Gets Sent to LLM APIs

​Using 100% Local Setup

​Open Source Benefits

Audit the Code

Self-Host Everything

Customize Freely

No Vendor Lock-in

​Source Code Reference

Build docs developers (and LLMs) love

Core Principles

Data Storage

Directory Structure

Markdown Format

Version Control

AI Model Configuration

OAuth & Credentials

How OAuth Works

Scopes Requested

Data Portability

Export Your Data

Use with Other Tools

Privacy Considerations

What Stays Local

What Gets Sent to LLM APIs

Using 100% Local Setup

Open Source Benefits

Source Code Reference