Configuration

Supported LLM Providers

Gravity uses a generic OpenAI-compatible client, so any provider exposing an OpenAI-compatible API works out of the box.

Provider	Base URL	Auth	Notes
Ollama	http://localhost:11434	None	Local models, auto-detects running instance
LM Studio	http://localhost:1234	None	Local server with model management UI
OpenRouter	https://openrouter.ai/api/v1	API Key	Access to 100+ models via single key
OpenAI	https://api.openai.com/v1	API Key	GPT-4o, GPT-4o-mini, etc.
Groq	https://api.groq.com/openai/v1	API Key	Ultra-fast inference, free tier available
Google AI Studio	https://generativelanguage.googleapis.com/v1beta/openai/	API Key	Gemini models
TogetherAI	https://api.together.xyz/v1	API Key	Open-source models, competitive pricing
GitHub Models	https://models.inference.ai.azure.com	GitHub Token	Free for GitHub users, GPT-4o access
Custom	Any URL	API Key	Any OpenAI-compatible endpoint

Model Selection

Each provider supports different models. Recommended coding models by category:

Local (Ollama / LM Studio)

• qwen2.5-coder:1.5b - Fast, good for simple tasks (1.5B params)
• qwen2.5-coder:7b - Balanced performance (7B params)
• deepseek-coder:6.7b - Strong code generation
• codellama:7b - Meta's code-specialized model
• deepseek-coder:33b - Best local quality (needs 24GB+ RAM)

Cloud (OpenRouter / OpenAI / Others)

• anthropic/claude-3.5-sonnet - Best overall coding agent
• anthropic/claude-3.5-haiku - Faster, cheaper, still excellent
• google/gemini-pro-1.5 - Large context window (1M tokens)
• openai/gpt-4o - Strong general purpose
• openai/gpt-4o-mini - Cost-effective for simple tasks
• meta-llama/llama-3.1-405b - Largest open model on OpenRouter

Development Modes

Controls how much autonomy agents have when executing tasks:

Mode	Behavior	Use Case
Autopilot (default)	Agents execute without confirmation. Files are written, commands run automatically.	Trusted environments, clear tasks, greenfield projects
Review	Approval required before destructive operations (file writes, shell commands). Shows diff preview.	Production codebases, sensitive changes, learning the system
Assisted	Full human-in-the-loop. Every tool call requires explicit approval.	High-risk changes, teaching/mentoring, audit requirements

Configure in Settings → Development Mode. The mode persists across sessions.

Advanced Settings

• Max Steps - Maximum reasoning iterations per task (default: 25)
• Context Budget - Max tokens injected per step (default: 700 chars × 2 chunks)
• RAG Re-index Interval - Full workspace re-index (default: 5 minutes)
• Agent Pool Size - Max concurrent AppEngine instances (default: 5)
• Theme - Light/Dark/System (affects AvalonEdit, WebView2, Quill editor)
• Font Size - Editor and chat font scaling
• Telemetry - Opt-in anonymous usage statistics

All settings stored in %APPDATA%/Gravity/settings.json.