Model Configuration

Rowboat Desktop works with any AI provider - from hosted APIs like OpenAI and Anthropic to fully local models via Ollama. This guide covers how to configure your model setup.

Configuration File

Model settings are stored in:

~/.rowboat/config/models.json

Structure:

{
  "provider": {
    "flavor": "openai",
    "apiKey": "sk-...",
    "baseURL": "https://api.openai.com/v1",
    "headers": {}
  },
  "model": "gpt-5.2",
  "knowledgeGraphModel": "gpt-4.1"
}

You can configure a separate model for knowledge graph operations. If knowledgeGraphModel is omitted, Rowboat uses the main model for all operations.

Configuring via UI

The easiest way to configure models is through the Settings dialog:

Open Settings

Click the settings icon in the sidebar or press Cmd+, (Mac) / Ctrl+, (Windows/Linux)

Select Provider

Choose from OpenAI, Anthropic, Google, Ollama, OpenRouter, or other providers

Enter Credentials

Cloud providers: Enter your API key
Local providers: Set the base URL (e.g., http://localhost:11434 for Ollama)

Choose Models

Assistant model: The main model for chat and tasks
Knowledge graph model: (Optional) A different model for graph operations

Test & Save

Click “Test & Save” to verify the connection and save your configuration

Supported Providers

OpenAI

Use GPT models from OpenAI.

Configuration
Example Config

Provider flavor: openaiRequired:

apiKey - Your OpenAI API key from platform.openai.com

Optional:

baseURL - Custom API endpoint (defaults to OpenAI’s official endpoint)
headers - Additional HTTP headers

Recommended models:

gpt-5.2 - Most capable (if you have access)
gpt-4.1 - Excellent performance
gpt-4o-mini - Fast and cost-effective

{
  "provider": {
    "flavor": "openai",
    "apiKey": "sk-proj-..."
  },
  "model": "gpt-5.2",
  "knowledgeGraphModel": "gpt-4.1"
}

Anthropic

Use Claude models from Anthropic.

Configuration
Example Config

Provider flavor: anthropicRequired:

apiKey - Your Anthropic API key from console.anthropic.com

Optional:

baseURL - Custom API endpoint
headers - Additional HTTP headers

Recommended models:

claude-opus-4-6-20260202 - Most capable Claude model
claude-sonnet-4-6-20260202 - Balanced performance
claude-3-5-sonnet-20241022 - Fast and cost-effective

{
  "provider": {
    "flavor": "anthropic",
    "apiKey": "sk-ant-..."
  },
  "model": "claude-opus-4-6-20260202"
}

Google AI Studio

Use Gemini models from Google.

Configuration
Example Config

Provider flavor: googleRequired:

apiKey - Your Google AI API key from aistudio.google.com

Optional:

baseURL - Custom API endpoint
headers - Additional HTTP headers

Available models:

gemini-2.0-flash-exp - Latest experimental model
gemini-1.5-pro - Production-ready
gemini-1.5-flash - Fast and efficient

{
  "provider": {
    "flavor": "google",
    "apiKey": "AIza..."
  },
  "model": "gemini-2.0-flash-exp"
}

Ollama (Local)

Run models locally on your machine with Ollama.

Setup
Configuration
Example Config

Install Ollama

Download from ollama.ai and install on your system

Pull a model

ollama pull llama3.3:70b
# or
ollama pull qwen2.5-coder:32b

Verify it's running

ollama list

Configure Rowboat

Set provider to “Ollama (Local)” and enter the model name

Provider flavor: ollamaRequired:

baseURL - Ollama API endpoint (default: http://localhost:11434)
model - The model name (e.g., llama3.3:70b)

Optional:

headers - Additional HTTP headers

Note: API key is not required for OllamaRecommended models:

llama3.3:70b - Highly capable open model
qwen2.5-coder:32b - Excellent for code and technical tasks
mistral-nemo:latest - Fast and efficient

{
  "provider": {
    "flavor": "ollama",
    "baseURL": "http://localhost:11434"
  },
  "model": "llama3.3:70b"
}

Ollama connection tests have a 60-second timeout (vs. 8 seconds for cloud providers) to accommodate model loading time.

OpenRouter

Access multiple models with one API key via OpenRouter.

Configuration
Example Config

Provider flavor: openrouterRequired:

apiKey - Your OpenRouter API key from openrouter.ai

Optional:

baseURL - Custom endpoint (defaults to OpenRouter’s API)
headers - Additional headers (e.g., for site identification)

Example models:

openai/gpt-4-turbo
anthropic/claude-3-opus
google/gemini-pro-1.5
meta-llama/llama-3.3-70b-instruct

{
  "provider": {
    "flavor": "openrouter",
    "apiKey": "sk-or-v1-..."
  },
  "model": "anthropic/claude-3-opus"
}

Vercel AI Gateway

Route requests through Vercel’s AI Gateway for observability and caching.

Configuration
Example Config

Provider flavor: aigatewayRequired:

apiKey - Your provider’s API key (OpenAI, Anthropic, etc.)
baseURL - Your AI Gateway endpoint from vercel.com/dashboard

Optional:

headers - Additional HTTP headers

{
  "provider": {
    "flavor": "aigateway",
    "apiKey": "sk-...",
    "baseURL": "https://gateway.vercel.ai/v1/..."
  },
  "model": "gpt-4"
}

OpenAI-Compatible APIs

Use any OpenAI-compatible API (LM Studio, LocalAI, etc.).

Configuration
Example Config

Provider flavor: openai-compatibleRequired:

baseURL - Your API endpoint (e.g., http://localhost:1234/v1 for LM Studio)
model - The model name to use

Optional:

apiKey - API key if required by your server
headers - Additional HTTP headers

{
  "provider": {
    "flavor": "openai-compatible",
    "baseURL": "http://localhost:1234/v1",
    "apiKey": "not-needed"
  },
  "model": "local-model"
}

Advanced Configuration

Custom Headers

Add custom HTTP headers to requests:

{
  "provider": {
    "flavor": "openai",
    "apiKey": "sk-...",
    "headers": {
      "X-Custom-Header": "value",
      "Organization": "org-..."
    }
  },
  "model": "gpt-4"
}

Separate Knowledge Graph Model

Use a different model for knowledge graph operations:

{
  "provider": {
    "flavor": "openai",
    "apiKey": "sk-..."
  },
  "model": "gpt-5.2",
  "knowledgeGraphModel": "gpt-4o-mini"
}

Why use a separate model?

Cost optimization - Use a cheaper model for graph operations
Speed - Use a faster model for background processing
Quality - Use a more capable model for chat, simpler one for extraction

If knowledgeGraphModel is omitted or empty, Rowboat uses the main model for all operations.

Connection Timeout

Rowboat tests model connections before saving:

Cloud providers: 8-second timeout
Local providers (Ollama, OpenAI-compatible): 60-second timeout

This allows time for local models to load into memory on first request.

Models Catalog

Rowboat caches a catalog of available models for OpenAI, Anthropic, and Google:

~/.rowboat/config/models.dev.json

This catalog powers the model dropdown in Settings. It’s automatically fetched and cached when you open the Settings dialog.

For local providers (Ollama, OpenAI-compatible), you’ll type the model name manually since available models vary by installation.

Troubleshooting

Connection Test Fails

Verify API key

Ensure your API key is valid and has not expired

Check base URL

For local providers, ensure the service is running:

# Ollama
ollama list

# LM Studio - check the server tab

Test manually

# OpenAI
curl https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY"

# Ollama
curl http://localhost:11434/api/tags

Check firewall

Ensure your firewall allows outbound connections (cloud) or localhost connections (local)

Model Not Found

For cloud providers:

Verify the model name matches the provider’s documentation
Check if you have access to that model (e.g., GPT-5 requires special access)

For Ollama:

ollama list  # See available models
ollama pull llama3.3:70b  # Download if needed

For OpenAI-compatible:

Check your server’s available models endpoint
Ensure the model is loaded and ready

Knowledge Graph Model Not Working

Test the model separately by setting it as the main model
Check that it supports the same capabilities (function calling, etc.)
Verify sufficient context window (knowledge graph operations can be token-heavy)

Best Practices

Start with defaults

Use recommended models first (gpt-5.2, claude-opus-4-6, etc.)

Test before saving

Always use “Test & Save” to verify your configuration works

Consider costs

Use a cheaper model for knowledge graph if you process many emails

Try local models

Ollama with llama3.3:70b is comparable to GPT-4 and completely private

Next Steps

Explore Features

Learn what you can do with your configured models

Understand Workspace

Explore the ~/.rowboat/ directory structure

Getting Started

Core Concepts

Desktop App

Integrations

Use Cases

Configuration

Model Configuration

Configuration File

Configuring via UI

Supported Providers

OpenAI

Anthropic

Google AI Studio

Ollama (Local)

OpenRouter

Vercel AI Gateway

OpenAI-Compatible APIs

Advanced Configuration

Custom Headers

Separate Knowledge Graph Model

Connection Timeout

Models Catalog

Troubleshooting

Connection Test Fails

Model Not Found

Knowledge Graph Model Not Working

Best Practices

Start with defaults

Test before saving

Consider costs

Try local models

Next Steps

Explore Features

Understand Workspace

Build docs developers (and LLMs) love

Getting Started

Core Concepts

Desktop App

Integrations

Use Cases

Configuration

​Configuration File

​Configuring via UI

​Supported Providers

​OpenAI

​Anthropic

​Google AI Studio

​Ollama (Local)

​OpenRouter

​Vercel AI Gateway

​OpenAI-Compatible APIs

​Advanced Configuration

​Custom Headers

​Separate Knowledge Graph Model

​Connection Timeout

​Models Catalog

​Troubleshooting

​Connection Test Fails

​Model Not Found

​Knowledge Graph Model Not Working

​Best Practices

Start with defaults

Test before saving

Consider costs

Try local models

​Next Steps

Explore Features

Understand Workspace

Build docs developers (and LLMs) love

Configuration File

Configuring via UI

Supported Providers

OpenAI

Anthropic

Google AI Studio

Ollama (Local)

OpenRouter

Vercel AI Gateway

OpenAI-Compatible APIs

Advanced Configuration

Custom Headers

Separate Knowledge Graph Model

Connection Timeout

Models Catalog

Troubleshooting

Connection Test Fails

Model Not Found

Knowledge Graph Model Not Working

Best Practices

Next Steps