Most AI app builders require a constant internet connection and send your code through someone else's servers. If you want local AI coding, where the model runs on your hardware and nothing leaves your machine, options are limited. Dyad is the only open-source AI app builder that supports local models as a first-class feature, meaning you can build full applications offline, for free, with no API keys required.
Why Run AI Models Locally?
There are practical reasons to keep the AI on your machine:
- Privacy. Your code and prompts never leave your device. No third-party data policies to evaluate.
- Cost. Local models have zero per-token cost. Once downloaded, you can generate as many responses as your hardware allows.
- Offline access. You can build apps on a plane, in a coffee shop with bad Wi-Fi, or on an air-gapped network.
- Control. You pick the exact model and version. No surprise behavior changes from a provider updating their API.
The trade-off is quality. The best cloud models (Claude Sonnet 4, GPT-4.1, Gemini 2.5 Pro) are still stronger at complex coding tasks than most models you can run locally. For simpler apps, prototyping, or privacy-sensitive work, local models are a solid option.
Setting Up Ollama with Dyad
Ollama is a free, open-source tool for running language models on your machine. It handles downloading, managing, and serving models with a single command.
- Install Ollama from ollama.com. It runs on Mac, Windows, and Linux.
- Pull a model. Open a terminal and run:
This downloads the Qwen 2.5 Coder model (14 billion parameters). Smaller models like qwen2.5-coder:7b work on machines with less RAM. Larger models need more memory but produce better results.
- Start Ollama. It typically runs as a background service after installation. Confirm it is running by visiting http://localhost:11434 in your browser.
- Open Dyad. Click the model picker, select Local Models, and your Ollama models will appear in the list.
That is the entire setup. No API key, no account, no sign-up. Dyad detects Ollama automatically on localhost:11434. If you have set a custom OLLAMA_HOST environment variable, Dyad uses that instead.
Which Models Work Best?
Not every model handles Dyad's code generation instructions well. Smaller models (under 7B parameters) often struggle to follow structured output formats. I recommend starting with one of these:
- Qwen 2.5 Coder 14B or 32B for the best local AI coding results. Requires 16GB+ RAM for 14B, 32GB+ for 32B.
- DeepSeek Coder V2 if you want a strong alternative.
- Qwen 2.5 Coder 7B if you are on a machine with limited RAM (8GB minimum).
Keep in mind that generation speed depends on your hardware. A machine with a dedicated GPU (NVIDIA with CUDA, or Apple Silicon with Metal) will be significantly faster than CPU-only inference.
Using LM Studio
LM Studio is another option for running local models. It provides a graphical interface for downloading and managing models, which can be more approachable if you prefer not to use the terminal.
- Download LM Studio from lmstudio.ai.
- Search for and download a coding model (the same ones listed above work here).
- Load the model and start the local server. By default, LM Studio serves on
http://localhost:1234. - In Dyad, open the model picker and select Local Models. Your LM Studio models will appear.
If you run LM Studio on a different port, you can add it as a custom AI provider in Dyad's settings instead.
Custom AI Providers with OpenAI-Compatible APIs
Dyad supports any AI provider that offers an OpenAI-compatible API. This covers a wide range of self-hosted and third-party options beyond Ollama and LM Studio.
To add a custom provider:
- Go to Settings and find the AI Providers section.
- Click Add Custom Provider.
- Enter an ID (internal, cannot change later), a display name, and the API base URL from your provider's documentation.
- After adding the provider, click on it and select Add Custom Model. Fill in the model ID (must match the provider's API exactly), a display name, and optionally the context window size and max output tokens.
This works with services like vLLM, text-generation-inference, or any OpenAI-compatible server. If you are running a model on a different machine on your network, point the base URL to that machine's address.
Enterprise Providers: AWS Bedrock, Google Vertex AI, Azure
For teams that need to use models through their organization's cloud accounts, Dyad supports enterprise AI providers. These were contributed by the open-source community:
- AWS Bedrock: Access Anthropic, Meta, and other models through your AWS account. Useful when your organization already has AWS billing and governance set up.
- Google Vertex AI: Use Google's models through your GCP project with enterprise-grade access controls.
- xAI (Grok): Use xAI's models with your own API key.
These providers are configured in Dyad's Settings under AI Providers, the same way you set up any other provider. The key difference is that billing and access go through your organization's existing cloud accounts rather than direct API keys.
Local vs. Cloud: Choosing the Right Setup
There is no single right answer. The best setup depends on what you are building.
Use local models when:
- Privacy is a hard requirement
- You want zero ongoing cost
- You need offline access
- You are prototyping or building simpler apps
Use cloud models when:
- You need the best code generation quality for complex tasks
- You want faster iteration on large codebases
- You are working on production apps where accuracy matters more than cost
Use both. Dyad does not lock you into one provider. You can use Ollama for initial prototyping, switch to Gemini or Claude for tricky features, then go back to local models for refinement. The model picker in Dyad lets you swap providers mid-conversation.
What Makes Dyad Different Here
Other AI app builders (Bolt.new, Lovable, v0, Replit) are cloud-only. They require an internet connection and run on the vendor's servers. You cannot plug in a local model.
Dyad runs on your desktop. It is an Electron app for Mac, Windows, and Linux. Your projects are standard codebases stored on your machine. The AI generates real code using common frameworks, and you can open that code in any editor, commit it to Git, and deploy it anywhere.
For local AI coding, this architecture is the differentiator. Because Dyad runs locally, connecting to a local model is a direct localhost call with no proxy, no relay, and no internet hop.
Dyad is open-source under the MIT license (with FSL 1.1 for pro features). You can inspect exactly what the application sends to any model provider. The code is on GitHub and the community actively contributes new providers, templates, and integrations.
Getting Started
If you want to try local AI coding today:
- Download Dyad for your platform.
- Install Ollama and pull a coding model.
- Open Dyad, select a local model, and start building.
No account needed. No API key needed. Everything runs on your machine.
