Dyad vs bolt.diy: Free Local AI App Builders for Windows (2026)
Dyad vs bolt.diy: Free Local AI App Builders for Windows (2026)
Documentation reviewed November 17, 2026 · Windows setup and buyer's guide · Hands-on benchmark not performed
Quick answer: Dyad is a practical starting point if you want an installed desktop app and generated source files in an ordinary folder on your Windows computer. bolt.diy is worth exploring if you prefer a browser-based development workspace you host yourself, with configurable cloud and local AI providers. Both support local models via Ollama. Neither guarantees an entirely offline workflow unless your models, dependencies and development tools are already available locally.
This is a source-backed setup and decision guide, not a performance ranking. We compare the products' documented workflows, file storage, local-model setup and limitations. We have not run the same benchmark in both applications, measured their generation speed or verified which creates better apps.
The real question is not simply which interface looks better. It is where your project files end up, which AI model sees your prompts, how you recover from a bad edit, and whether you can continue working outside the builder.
What does “local AI app builder” mean?
A product can be local in one way and cloud-dependent in another. Four terms are worth separating:
| Term | What it means | What it does not guarantee |
|---|---|---|
| Locally installed builder | The application or development interface runs on your PC | AI inference stays on your PC |
| Locally stored project | You control the generated source files or a backup | The builder never contacts a remote service |
| Local AI inference | A downloaded model runs via Ollama or another local runtime | Every dependency and integration works offline |
| Offline development | Your chosen workflow can work without the internet after preparation | Cloud databases, deployment and package downloads remain available |
For example, a local Dyad project using a cloud model still sends model requests to the chosen provider. Similarly, running bolt.diy at localhost:5173 does not mean all generated projects are saved as normal Windows folders. Its standard browser-based flow uses browser storage, so backup and export deserve attention. The bolt.diy team has publicly discussed this design choice in its project-persistence issue.
Privacy checklist: Check the model provider, API endpoint, browser storage, optional telemetry, third-party packages, databases and deployment integrations before working on sensitive code.
Dyad vs bolt.diy: key differences
| Question | Dyad | bolt.diy |
|---|---|---|
| What do I install? | A Windows desktop app; Node.js to run generated apps | A desktop release, or Git/Node.js/pnpm to self-host its web interface |
| How do I work? | Prompt, inspect changes, approve, preview locally | Prompt and edit in a self-hosted browser workspace using an integrated runtime |
| Where is my source? | Ordinary local project files you can inspect with an editor | Browser-backed project state in the standard web flow; export/sync/backup options |
| Can I use Ollama? | Yes, via the Local Models model picker | Yes, via Settings → Providers → Local Providers |
| Can I use cloud models? | Yes, depending on provider and account | Yes, through configured providers and credentials |
| Is it free? | Free open-source core; optional paid features and possible model charges | Open-source project; provider costs and licensing considerations may still apply |
| Main operational risk | Smaller local models may struggle to follow complex app-editing instructions | Browser storage, browser runtime limitations and environment configuration |
| Better fit for… | Beginners who prioritize local files and a desktop workflow | Developers who want a browser workspace with more environment and provider controls |
How to read this table: The comparison reflects documented product behavior, not measured feature reliability. Features can change between releases. Dyad's own comparison with bolt.diy provides context, but it is a vendor-authored source; the bolt.diy repository is used separately to cross-check its side of the comparison.
Dyad: getting started on Windows
Dyad is a desktop-first, open-source AI app builder. You can describe an application, inspect generated code changes and approve them before the files are used to run your project. Dyad documents support for Ollama and LM Studio as local AI model sources.
When Dyad makes sense
Choose this route when you want to build a prototype and then open the same code in VS Code or another editor. For example, a small client tracker might need a form, editable records, filters and persistence. Dyad can help scaffold that project; you still need to inspect the generated logic and test it before treating it as reliable software.
Step-by-step: connect Dyad to Ollama
-
Download Dyad from its official site and install the Windows application. If Windows shows a reputation or publisher warning, verify the source before deciding whether to proceed. Do not dismiss warnings automatically.
-
Install Node.js if Dyad requests it. Dyad uses Node.js to run the web apps it creates. Check that PowerShell can find it:
node --version -
Install Ollama for Windows from the official downloads page and start it. Its default local API endpoint is
http://127.0.0.1:11434. -
Download a coding-model example in PowerShell:
ollama pull qwen2.5-coder:7b ollama listThis is a reproducible baseline example, not a claim that Qwen2.5-Coder is the best coding model in late 2026. If memory is tight, try
qwen2.5-coder:3b; smaller models may be less reliable with multi-file editing instructions. -
Open Dyad's model picker, choose Local Models, and select an available Ollama model. If it is missing, first check that Ollama is running and that its hostname is correct. Dyad documents use of
OLLAMA_HOSTorhttp://localhost:11434by default. -
Ask for one small feature rather than a complete SaaS application. Approve the changes only after reviewing them, start the local preview and check whether the workflow works after a refresh.
The authoritative setup references are Dyad's quickstart and local-model guide. If a new Dyad version changes the interface, follow its current guide instead of guessing which buttons moved.
What to check before continuing
Find the generated project folder and open it outside Dyad. Can you identify the entry point, package manifest and relevant source files? Can you run the project independently? Create a Git commit or copy the project to a backup location before attempting larger edits. These checks matter more for long-term ownership than how polished the first preview looks.
bolt.diy: run a local, browser-based AI app workspace
bolt.diy is an open-source development workspace for prompting, generating and editing web apps. It supports multiple providers, including Ollama and LM Studio, and its standard app-building experience operates in a browser-based runtime. The project also publishes desktop binaries.
A self-hosted browser UI can feel convenient, especially if you already manage development tooling. But self-hosted is not a synonym for “all files are safely on disk”, and a browser-based runtime is not identical to running every possible backend on Windows directly.
Install bolt.diy from source on Windows
According to the project's official installation documentation, the current minimum requirements are Node.js 22.12.0 or later and pnpm 9.15.9 or later, plus Git. Check versions before installing:
node --version
pnpm --version
If pnpm is missing, install it using npm install -g pnpm, then check the version again. For a source installation, the repository's quickstart currently shows:
git clone https://github.com/stackblitz-labs/bolt.diy.git
cd bolt.diy
pnpm install
pnpm run dev
Open the URL printed by the terminal, typically http://localhost:5173. Some repository instructions mention a separate stable branch: if you intentionally select that branch, follow its matching setup and requirements rather than mixing commands from different revisions.
If you would rather avoid building from source, check the official releases for a Windows desktop binary.
Connect Ollama to bolt.diy
- Start the Ollama application and confirm a model is installed using
ollama list. - In bolt.diy, open Settings → Providers → Local Providers.
- Enable Ollama and check its endpoint. For Ollama running directly on the same Windows machine, the documented default is
http://127.0.0.1:11434. - Select the Ollama provider and an installed model in the builder. Test with a small request before creating a large application.
If you run bolt.diy in Docker, on a different computer or behind a reverse proxy, localhost may refer to the container or remote machine—not the Windows host. See the troubleshooting section below before changing any network settings.
Two trade-offs not to overlook
Project recovery: In the standard browser workflow, project state has relied on browser storage such as IndexedDB/localStorage. Don't count on a browser profile as your only backup. Export a ZIP, sync to a folder if supported by your version, or push to a Git repository and verify the files are usable elsewhere. A maintainer response explains the browser-storage design.
Commercial use: Although the bolt.diy source is distributed under MIT, its README licensing section explains that its WebContainers API dependency may require separate licensing for production use in a commercial, for-profit setting. Prototypes and proofs of concept are treated differently. Review the current terms if a company plans to deploy a production service using that API; a repository license alone does not settle dependency licensing.
Can an 8GB GPU run a local coding model?
Yes, it can be a reasonable starting point for experimentation, but “the model fits” and “the model reliably builds an app” are different claims. The total memory footprint depends on quantization, context length, the runtime, GPU memory and system RAM. Longer contexts can consume substantially more memory.
The following figures are download sizes listed in Ollama's model registry, not measured VRAM usage or benchmark results:
| Example Ollama tag | Listed download size | Practical interpretation |
|---|---|---|
qwen2.5-coder:3b |
1.9 GB | Smaller starting experiment; check whether it follows multi-step instructions |
qwen2.5-coder:7b |
4.7 GB | Reasonable baseline to try with an 8GB card, subject to context and overhead |
qwen2.5-coder:14b |
9.0 GB | Larger than 8GB before runtime overhead; may need partial CPU/RAM execution |
Source: Ollama Qwen2.5-Coder model tags. The registry's listed context capability should not be confused with the context a particular PC can afford to run.
Start with the 7B tag for a controlled test, a modest context setting and a simple UI task. After sending a prompt to load the model, run:
ollama ps
The PROCESSOR column can indicate whether the loaded model is executing entirely on GPU, CPU or across both. See Ollama's FAQ. Record the model tag or digest, actual context setting, Windows version, GPU and system RAM if you publish any performance observations. This guide has not independently measured speed or code quality on an 8GB machine.
What does “free” really cost?
The software license is only one part of the calculation:
| Cost area | Potential expense |
|---|---|
| Builder | Free core or source; optional premium features and certain commercial licensing contexts |
| Local AI inference | Electricity, hardware, memory and time; generally no per-call charge for a locally downloaded model |
| Cloud AI inference | Paid API usage or plan limits if you choose a hosted model |
| App infrastructure | Hosting, domain, authentication, managed database, email or storage |
| Verification | Time spent reviewing security, testing edge cases and fixing errors |
A local setup can reduce model API spending, especially while you experiment, but it may not minimize the total cost of shipping and maintaining a real product. Compare the cost per completed, working feature rather than only the subscription price.
A repeatable test you can run in both builders
A fair comparison should use the same prompt, model and hardware in each tool, as far as their integrations allow. Don't compare a polished cloud model in one tool with a small local model in the other and attribute the difference entirely to the builder.
Copy this app-building prompt
Build a responsive personal expense tracker with a form for date, category, amount and notes. Show a monthly total and category filter. Allow users to add, edit and remove expenses. Persist records locally across refreshes without a paid backend. Validate missing fields and amounts that are zero or negative. Include an accessible empty state, form labels and keyboard-visible focus. First summarize the planned file structure, then generate the application. Do not add authentication or external services.
Record the results, not just a screenshot
| Test | Pass condition | Evidence to save |
|---|---|---|
| Initial run | App launches without hand-editing code | Console/terminal output and running preview |
| Add & edit | Create and update three expenses | Short screen recording or genuine screenshots |
| Persistence | Saved entries survive refresh or restart | Before/after evidence |
| Validation | Empty fields and invalid amounts are blocked | Visible messages and test notes |
| Accessibility | Form labels and keyboard focus work | Manual keyboard test notes |
| Recovery | You can undo a bad change and export the code | Exported project and revision notes |
| Repeatability | Model, runtime, settings and prompt are recorded | Test log with version details |
Testing disclosure: These are instructions for a future Mazikbox evaluation. No pass/fail scores, speed comparisons or generated-app screenshots are claimed in this article. A useful test report would identify the Dyad version, bolt.diy commit or release, Windows build, exact model tag or digest, context setting, VRAM/RAM, number of retries and whether a cloud provider was used.
Troubleshooting Ollama connections on Windows
Step 1: confirm Ollama has a model
Run in PowerShell:
ollama list
Invoke-RestMethod -Uri http://127.0.0.1:11434/api/tags
The first command lists installed models; the second checks whether the local Ollama API responds with model data. If it fails, ensure the Ollama Windows app is running before making any other configuration changes.
Step 2: check which machine the endpoint refers to
If Dyad and Ollama are both installed on Windows, try the default localhost:11434. For bolt.diy running directly on the same Windows computer, its documented local endpoint is http://127.0.0.1:11434. If your builder runs in Docker, WSL, another computer or a remote server, the hostname must refer to the machine where Ollama is actually reachable from that runtime. Do not assume localhost works across these boundaries.
Avoid exposing Ollama to the public internet simply to fix a local-connection issue. If a network bind address must change, follow Ollama's configuration FAQ and restrict access appropriately.
Step 3: diagnose browser-origin errors separately
A browser-to-Ollama request may be blocked because the requesting origin is not allowed. Ollama documents the OLLAMA_ORIGINS environment variable for additional trusted web origins. If the browser console explicitly reports a CORS error, configure only the specific origin you control and restart Ollama. Avoid * or broad third-party origin patterns as a generic fix. On Windows, Ollama reads user/system environment variables when the app starts, so quit and relaunch the app after changing them.
A server-side request failing with a connection-refused error is not the same problem as a browser CORS block; fix the address or service first.
Other common problems
Dyad says Node.js is missing: Open a new PowerShell window, run node --version, then check Dyad's Node.js help. Restart the app after installing Node.js.
bolt.diy cannot install dependencies: Recheck node --version and pnpm --version against the repository's current requirements. Don't assume any version labeled LTS is recent enough.
The model produces broken edits: Make the task smaller, ask for one feature at a time, inspect changed files and test after each revision. A larger model or a cloud model may be more effective for difficult tasks; verify that independently for your setup.
The bolt.diy project is missing after clearing browser data: Restore from a prior ZIP export, synced folder or Git repository. If you never created a backup, recovery may not be possible.
Which one should you choose?
Choose Dyad if you want a desktop-first workflow, easily inspected source files and a straightforward path from AI-assisted prototype to a conventional codebase.
Choose bolt.diy if you prefer a self-hosted, browser-based app-building workspace, need control over supported model providers and are prepared to manage project backups and environment quirks.
Choose neither as your default if you're primarily editing an existing large repository. In that case, tools such as OpenCode or OpenHands target coding-agent workflows. They are related options, not equivalent prompt-to-app builders.
You can also browse the larger Mazikbox AI App Builders directory or the AI Coding Agents category to shortlist tools by task. Product availability, features and pricing should be checked with the vendors before purchase or deployment.
Frequently asked questions
Is Dyad free, or do I need a paid plan?
Dyad provides an open-source core with optional paid functionality. Running a downloaded model in Ollama can avoid paid model API calls, but using a hosted AI provider may incur that provider's charges or limits.
Is bolt.diy the same product as Bolt.new?
No. bolt.diy is an open-source project inspired by the Bolt-style development experience. A self-hosted bolt.diy installation and the hosted Bolt.new service have different setup, provider and billing arrangements.
Can Dyad or bolt.diy work completely offline?
Some development tasks can work without internet access after you've installed the builder, model, dependencies and project files. Cloud models, external integrations, new packages and deployment require a connection. Test the exact workflow before relying on offline operation.
Why doesn't bolt.diy detect Ollama on Windows?
Confirm the Ollama service is running, use ollama list, test http://127.0.0.1:11434/api/tags, then check bolt.diy's configured endpoint. With Docker or remote runtimes, verify host networking. Investigate OLLAMA_ORIGINS only when the browser specifically reports an origin/CORS error.
Does an 8GB GPU guarantee fast AI app generation?
No. It can run some downloaded coding models, but performance and reliability depend on context, quantization, available VRAM/RAM, runtime and task complexity. The 3B/7B/14B figures above are download sizes, not benchmarked performance.
Can I move the source code to another editor?
Dyad saves normal project files locally. bolt.diy provides export/sync options; verify your specific installation's method and whether all essential files were included. For either tool, open the project independently and try a clean run before relying on an export.
Can I use bolt.diy commercially?
Review both the source license and the WebContainers API terms. The repository says production use in some for-profit settings requires a separate WebContainers license, while prototypes and proofs of concept are treated differently. Check the latest terms for your actual use case.
Final recommendation
For someone beginning local AI-assisted app building on Windows, Dyad is the more direct option when ordinary source files are the priority. bolt.diy is a useful alternative for a configurable browser-based development workspace. That is a workflow recommendation—not a claim that either generates more accurate code.
Whichever you choose, start with a small app that saves data, validate negative and empty inputs, export your code and rerun it independently. The ability to inspect, recover and maintain an application is a more useful buying criterion than a spectacular first prompt.
Sources and review method
Documentation checked November 17, 2026. This article is a secondary-source comparison based primarily on official documentation and repository-maintained information. No controlled hands-on benchmark, commercial legal review or independent price audit has been performed. A named author/reviewer and actual screenshots should be added by the editorial team before publication.
- Dyad — official website
- Dyad — Windows quickstart
- Dyad — connecting local models
- Dyad — vendor-authored bolt.diy comparison (treated as an interested source)
- bolt.diy — source code and current setup
- bolt.diy — detailed setup requirements
- bolt.diy — project persistence discussion with maintainer reply
- Ollama — Windows download
- Ollama — FAQ, network configuration and GPU placement
- Ollama — Qwen2.5-Coder model tags
Browse related categories