How to Run Bolt.diy with Ollama on Windows: Step-by-Step Setup Guide (Tested on 8GB RAM)
> Author's Note: This hands-on setup guide was originally tested and published with full high-resolution screenshots, latency logs, and reproducible troubleshooting steps on AISparkHub.
---
Quick Answer: How to Run Bolt.diy with Ollama on Windows
> To run Bolt.diy with Ollama locally on Windows:
> 1. Install Git and Ollama via PowerShell: winget install -e --id Git.Git; winget install -e --id Ollama.Ollama
> 2. Allow script execution: Set-ExecutionPolicy RemoteSigned -Scope CurrentUser -Force
> 3. Pull a lightweight coding model: ollama pull qwen2.5-coder:1.5b
> 4. Clone Bolt.diy and add OLLAMA_API_BASE_URL=http://127.0.0.1:11434 to a new .env.local file.
> 5. Run pnpm install and pnpm run dev, then select Ollama in the web interface.
Performance & Hardware Compatibility Matrix
| Hardware Profile | Recommended Model | Generation Latency | RAM Consumption | Suitability |
| :--- | :--- | :--- | :--- | :--- |
| 8 GB RAM / Integrated GPU (Tested) | qwen2.5-coder:1.5b | 7m 42s (Full scaffold) | ~2.45 GB combined | Working (Local CPU) |
| 16 GB RAM / No GPU | qwen2.5-coder:7b | ~3–5 mins | ~5.8 GB combined | Moderate |
| 16 GB+ / NVIDIA GPU (8GB+ VRAM) | qwen2.5-coder:7b or deepseek-r1:8b | ~20–45 seconds | GPU offloaded | Optimal |
| Any PC / Low Spec ($0 Cloud Fallback) | Groq / Gemini 2.0 Flash API | 1–3 seconds | < 500 MB | Fastest / 100% Free |
---
Test Machine Specifications
To provide realistic benchmarks for developers without expensive hardware, this guide was tested on a standard budget-tier laptop:
- Operating System: Windows 10/11 Home (64-bit)
- Processor: Intel Core i5 (CPU inference only)
- Graphics: Intel(R) HD Graphics 620 (No dedicated NVIDIA VRAM)
- Memory: 8 GB DDR4 (1.5 GB available physical memory at start)
- Runtime Stack: Node.js v24.20, Ollama v0.40.12, pnpm v9.15.9, Bolt.diy v1.0.0
---
Step 1: Installing Git, Ollama, and PNPM
Open PowerShell and install Git and Ollama in a single step using the native Windows Package Manager:
winget install -e --id Git.Git; winget install -e --id Ollama.Ollama
!Installing Git and Ollama via winget on Windows PowerShell
When the Ollama setup window appears, click "Continue", then select "No thanks, I'll use Ollama locally". Ollama will minimize to your Windows system tray.
Next, install pnpm globally:
npm install -g pnpm
---
Step 2: Resolving Common Windows Configuration Errors
Windows users routinely encounter three specific environment errors that prevent Bolt.diy from running. Apply these configurations before proceeding:
1. Fix the PowerShell Execution Policy Error
If running npm or pnpm produces this red error:
File npm.ps1 cannot be loaded because running scripts is disabled on this system
!PowerShell Security Error regarding Execution Policy
Run this command in PowerShell to grant execution rights to your user account without requiring Administrator privileges:
Set-ExecutionPolicy RemoteSigned -Scope CurrentUser -Force
!Execution Policy resolved and pnpm installed cleanly
2. Enable Ollama Cross-Origin Requests (CORS)
Because Bolt.diy runs in your browser via a local WebContainer, Windows security blocks requests to http://localhost:11434 unless CORS is explicitly permitted. Run:
[System.Environment]::SetEnvironmentVariable('OLLAMA_ORIGINS', '*', 'User')
3. Resolving the PowerShell "QuickEdit" Console Freeze
> [!WARNING]
> If you click inside the blue PowerShell console, Windows enters text-selection mode. The console title will change to Select Windows PowerShell, and Windows will pause the Node.js process.
> Resolution: Click inside the terminal and press the Esc key. The word Select will disappear, and the server will resume immediately.
---
Step 3: Downloading the 8GB-Friendly Model
Heavier models (such as 7B or 14B parameter models) consume 5GB–9GB of memory and will cause an 8GB Windows machine to freeze or crash.
For smooth local CPU execution, pull Alibaba's specialized 1.5B coding model (~986 MB download):
ollama pull qwen2.5-coder:1.5b
To verify that Ollama is serving the model correctly, run:
curl http://localhost:11434/api/tags
---
Step 4: Cloning Bolt.diy & Environment Setup
Clone the official repository:
git clone https://github.com/stackblitz-labs/bolt.diy.git
cd bolt.diy
Configuring the Required .env.local File
> [!IMPORTANT]
> By default, Bolt.diy will hide Ollama from the model dropdown unless it detects a base URL configured in an environment file.
Create a file named .env.local in the root bolt.diy directory and add the following lines:
OLLAMA_API_BASE_URL=http://127.0.0.1:11434
DEFAULT_NUM_CTX=32768
Now install dependencies and start the development server:
pnpm install
pnpm run dev
!Bolt.diy server running with secrets loaded from env.local
---
Step 5: Connecting the Engine & The 7m 42s Benchmark
Open your browser and navigate to:
http://localhost:5173
With .env.local loaded, click the provider dropdown. You will see a green status indicator: • Ollama, with qwen2.5-coder:1.5b (1.5B) selected as the default model:
!Bolt.diy interface showing Ollama connected with qwen2.5-coder 1.5b
The Benchmark Test
To benchmark real-world performance, we submitted a standard application prompt:
Create a clean and responsive Pomodoro Timer app with Start, Pause, and Reset buttons, a 25-minute countdown, and a simple task checklist using React and Tailwind CSS.
The Workbench Scaffolding Phase
The local model immediately scaffolded the React template, configured package.json, and set up the directory tree:
!Bolt Workbench creating initial project files on local CPU
- Total Scaffolding Latency: 7 minutes 42 seconds (462s) on CPU.
- Context Consumption: 4,421 tokens processed (3,996 prompt tokens, 425 completion tokens).
- Financial Cost: $0.00 (Zero API fees; no subscription caps).
---
Step 6: Overcoming the 1.5B Model Limitation ("No Preview Available")
When you switch to the Preview tab after the initial prompt, you may see a blank screen or the placeholder message: "Start prompting (or editing) to see magic happen".
Why This Happens (Technical Explanation)
Sub-3B parameter models lack the instruction-following precision needed to parse Bolt's custom XML action tags (<boltAction type="file">). Instead of injecting code into src/App.tsx, the 1.5B model emits raw JSON tool calls:
{"name": "project_create_directory", "arguments": {"path": "./src/components"}}
The model creates the project structure, but leaves App.tsx with starter placeholder code.
Resolving the Issue: Manual Code Insertion
1. Click the Code tab in Bolt.diy.
2. In the left file tree, click src $\rightarrow$ App.tsx.
3. Replace the placeholder code with the complete React timer component:
import React, { useState, useEffect } from 'react';
export default function App() {
const [seconds, setSeconds] = useState(1500);
const [isActive, setIsActive] = useState(false);
useEffect(() => {
let interval: any = null;
if (isActive && seconds > 0) {
interval = setInterval(() => setSeconds((s) => s - 1), 1000);
} else if (seconds === 0) {
setIsActive(false);
alert('Time is up!');
}
return () => clearInterval(interval);
}, [isActive, seconds]);
const formatTime = (sec: number) => {
const mins = Math.floor(sec / 60);
const remainingSecs = sec % 60;
return `${mins.toString().padStart(2, '0')}:${remainingSecs.toString().padStart(2, '0')}`;
};
return (
<div className="min-h-screen bg-slate-900 text-white flex flex-col items-center justify-center p-4">
<div className="bg-slate-800 p-8 rounded-2xl shadow-xl text-center max-w-sm w-full border border-slate-700">
<h1 className="text-2xl font-bold mb-4 text-purple-400">Pomodoro Timer</h1>
<div className="text-6xl font-mono font-bold mb-8 tracking-wider">{formatTime(seconds)}</div>
<div className="flex gap-4 justify-center">
<button
onClick={() => setIsActive(!isActive)}
className={`px-6 py-2 rounded-lg font-semibold ${isActive ? 'bg-amber-500 hover:bg-amber-600' : 'bg-emerald-500 hover:bg-emerald-600'}`}
>
{isActive ? 'Pause' : 'Start'}
</button>
<button
onClick={() => { setIsActive(false); setSeconds(1500); }}
className="px-6 py-2 rounded-lg font-semibold bg-rose-500 hover:bg-rose-600"
>
Reset
</button>
</div>
</div>
</div>
);
}
4. If the preview does not refresh automatically, click into the Bolt Terminal at the bottom and enter:
npm run dev
5. Click back to the Preview tab. Hot Module Replacement (HMR) triggers immediately:
!Live interactive Pomodoro timer rendered inside Bolt.diy preview
The application is fully interactive, running inside your browser's local sandbox at zero cost.
---
Architectural Recommendation: The $0 Hybrid Setup
Running models locally via Ollama offers complete privacy and cost immunity. However, waiting 7+ minutes per generation on CPU hardware introduces friction during rapid prototyping.
The Recommended Workflow for Budget Hardware:
1. Use Bolt.diy as Your Base: Keep bolt.diy running locally to retain unrestricted code exports and bypass commercial token paywalls.
2. Plug in Free Cloud API Tiers: In Bolt.diy's Settings, add a free API key from Google Gemini 2.0 Flash or Groq Cloud. Both providers offer generous free tiers that deliver sub-second generation speeds without straining your laptop CPU.
3. Keep Ollama for Offline Development: Use your local qwen2.5-coder:1.5b setup whenever you are traveling or coding without internet access.
---
Frequently Asked Questions (FAQ)
Can I run Bolt.diy without a dedicated GPU on Windows?
Yes. You can run lightweight models like qwen2.5-coder:1.5b entirely on a CPU with 8GB of RAM. While project scaffolding takes around 7 minutes on entry-level hardware, it runs completely offline for $0. For instant generation without a GPU, you can also plug free cloud API keys (like Groq or Gemini Flash) directly into Bolt.diy.
Why doesn't Ollama appear in the Bolt.diy provider dropdown?
Bolt.diy hides Ollama by default if it cannot detect an active endpoint URL. You must create a file named .env.local in your root bolt.diy directory containing OLLAMA_API_BASE_URL=http://127.0.0.1:11434, then restart your local server with pnpm run dev.
Is Bolt.diy completely free to use?
Yes. Bolt.diy is an open-source, self-hosted alternative to Bolt.new. When paired with local models via Ollama or free developer API keys, you pay $0 in subscription fees and have zero daily token limits.
Why does the preview window show "No preview available"?
Lightweight models (under 3B parameters) sometimes struggle to trigger Bolt's automated XML code-injection tags. If the preview is blank, simply open the Code tab, paste your component logic into src/App.tsx, and run npm run dev inside the embedded Bolt Terminal.
What is the best Ollama model for Bolt.diy?
- For 8GB RAM / No GPU: qwen2.5-coder:1.5b (safely runs on CPU without freezing Windows).
- For 16GB RAM / Dedicated GPU: qwen2.5-coder:7b or deepseek-r1:8b (delivers full automated code injection and faster generation).
How to Bypass Bolt.new Token Limits: 5 Workarounds That Actually Work (2026)
> Author's Note: This benchmark was originally tested and published with full interactive comparison tables, memory profiles, and terminal execution logs on AISparkHub.
---
That Sinking Feeling at 2:00 AM
Picture this: You are deep in the zone building a brand-new SaaS idea on Bolt.new. The layout looks stunning, the components are clicking into place, and you feel like a 10x engineer who just discovered the future of vibe coding.
You type a simple prompt: "Change the pricing card border to emerald-500 and center the checkout button."
You hit Enter.
Instead of a smooth preview reload, your screen freezes. A bright red toast slams across your dashboard:
> "Token limit exceeded. You have reached your daily allowance of 300,000 tokens."
Your project grinds to a dead halt. StackBlitz politely points you to a button: "Upgrade to Pro for $25/month."
If you're an indie hacker, solo founder, or hobbyist building a prototype, spending $25/month before validating if anyone even wants your app hurts. Worse, if you don't understand why you ran out of tokens in the first place, upgrading won't even save you—you'll burn through that paid allowance in three days flat.
Let’s get one thing out of the way immediately: There is no secret backend hack or illegal cheat code to bypass StackBlitz’s cloud billing.
However, there are 5 battle-tested engineering workarounds that professional developers use to stretch their tokens, edit code for $0, stop burning millions of tokens in dumb error loops, and cleanly eject their projects to finish building for free.
Here is the unfiltered, practitioner-grade guide to mastering the Bolt.new token limit in 2026.
---
The Math: Why Does Bolt.new Burn Tokens Like Jet Fuel?
Before we fix the problem, you need to understand why your tokens vanished after just two or three prompts.
Most beginners assume Bolt works like ChatGPT. You send a 20-word prompt, you get a 100-word response, and you burn maybe 300 tokens.
That is fundamentally NOT how Bolt.new works.
Bolt runs an entire Node.js operating system inside your browser called a WebContainer. When you type a prompt, Bolt doesn’t just send your message to the AI. Behind the scenes, it bundles three massive payloads into every single execution cycle:
1. Your Entire Conversational History: Every prompt, every response, and every console error message since you opened the tab.
2. The Full Project File Tree: All directory paths, package configs (package.json, tsconfig.json, tailwind.config.js), and environment states.
3. The Raw Content of Every Touched Component: If your prompt touches the navbar, Bolt often reads the parent layout, the router, and related components into the model's context window.
We call this the Context-Sync Tax.
[Your 15-Word Prompt]
+
[Entire Accumulated Chat Log: ~35,000 tokens]
+
[Project File Tree & Configs: ~25,000 tokens]
+
[5 Source Code Files Re-Read: ~40,000 tokens]
════════════════════════════════════════════
Total Burn per Single Prompt: ~100,000+ Tokens!
The Real Data From Our Benchmark Lab
We didn't just guess this math; we measured it with a stopwatch and token counters in our systematic test (Lovable vs. Bolt.new Real Build Benchmark):
- Prompt 1 (Initial App Scaffolding): Consumed 101,000 tokens (~33.6% of daily cap).
- Prompt 2 (Adding Authentication & Persistence): Consumed 95,000 tokens (~31.6% of daily cap).
- Total 2-Prompt Burn: 196,000 tokens.
!Bolt.new Token Consumption Dashboard
Figure 1: Our empirical benchmark dashboard showing 196,000 tokens burned across just two prompts on Bolt.new's free tier.
Notice what happened: On Prompt 2, we only asked for a basic auth modal and local storage, yet it burned nearly 100,000 tokens.
Why? Because the project had expanded to 27 files.
Once your project reaches 20+ files, a one-line CSS color tweak can easily burn 120,000 to 150,000 tokens. On a free plan with a 300,000 daily cap, you hit an immovable brick wall by prompt #3.
---
The 5 Workarounds to Keep Building for $0
---
Workaround 1: The "Code View" Free-Editing Hack (0 Tokens)
The single biggest mistake developers make on Bolt.new is asking the AI chat box to make trivial adjustments.
If you type: "Fix the typo in the header on line 14" or "Change the background from slate-900 to zinc-950", Bolt re-indexes your entire project context. You just burned 110,000 tokens for a five-second manual edit!
The Fix: Bolt is not just a chatbot—it embeds a full Monaco code editor (the exact same engine powering VS Code) directly in your browser.
!Bolt.new Code View Manual Editing
Figure 2: Clicking the </> Code toggle at the top bar opens the full file tree and code editor, allowing manual edits with 0 token consumption.
How to use it:
1. In the top navigation bar, toggle from the Preview eye icon to the </> Code button.
2. In the left-hand sidebar, click into any file (e.g., src/components/PromptCard.tsx, index.html, or .env).
3. Make your styling changes, fix typos, or paste logic directly into the editor.
4. Save the file (Ctrl + S on Windows or Cmd + S on Mac).
5. Look at the bottom terminal: Bolt's built-in Vite development server immediately hot-reloads your changes in real-time.
Token Cost: Exactly 0. You can spend six hours polishing UI, tweaking colors, and refactoring logic without burning a single token from your daily allowance.
---
Workaround 2: Use a .bolt/ignore File to Slash Context Burn by 50%
By default, Bolt’s indexing crawler reads almost everything in your workspace when packaging context for the LLM. If your project contains large documentation files, sample data, mock JSON arrays, or heavy SVG icons, Bolt blindly packages all of that raw text into every prompt.
You can blindfold the AI to non-essential files using a .bolt/ignore file—similar to how .gitignore works for git.
How to set it up:
1. Open the Code View tab.
2. In your root directory, create a new file named .bolt/ignore (or place it inside the .bolt folder).
3. Add paths that the AI does not need to analyze on every prompt:
# Ignore documentation and markdown guides
README.md
docs/
*.md
# Ignore static mock data and heavy JSON files
src/data/mockData.json
public/assets/
# Ignore compiled build outputs and locks
dist/
package-lock.json
# Ignore test suites during UI prototyping
**/*.test.ts
**/*.spec.tsx
The Impact:
Shrinking your active context tree prevents Bolt from sending unnecessary kilobytes of text to the LLM. In our testing, properly excluding heavy assets and lockfiles reduced prompt context weight by 35% to 50%, effectively doubling your daily prompt capacity.
---
Workaround 3: Use Version History Instead of "Error-Loop Prompting"
Here is a horror story that drains millions of tokens every day:
1. Bolt generates a component that introduces a TypeScript error.
2. The user types: "Fix the error." (Cost: 110,000 tokens)
3. Bolt tries to fix it, but breaks another file.
4. The user types: "You broke the layout, undo that and fix it." (Cost: 130,000 tokens)
5. Result: 240,000 tokens burned in 3 minutes, your daily limit is gone, and the code is worse than when you started.
Never use AI chat prompts to undo mistakes.
Bolt.new includes a native Version History checkpoint system built directly into the UI.
!Bolt.new Version History Rollback
Figure 3: Bolt's Version History drawer lets you review and revert to any prior working checkpoint instantly for 0 tokens.
How to use it:
1. On the left chat panel, click the clock icon or look at your prompt history checkpoints.
2. Click Version History.
3. You will see a chronological list of every stable checkpoint Bolt saved (e.g., “Create PromptVault prompt library - Oct 01”).
4. Click on the last working checkpoint and select Revert to this version.
Bolt instantly rolls your entire WebContainer back to the exact working state. Total token cost: 0.
---
Workaround 4: The External LLM Debugging Trick
When your app throws a complex runtime error or a broken React hook, do not feed the error back into Bolt’s chat.
Instead, let free external models do the heavy lifting:
1. Copy the error message and the broken file’s code from Bolt’s Code View.
2. Open a separate tab with ChatGPT (Free GPT-4o), Claude.ai, or Google Gemini.
3. Prompt the external model:
> "I have a Vite + React component throwing this TypeScript error: [Paste error]. Here is the file: [Paste code]. Give me ONLY the corrected code block to paste back in."
4. Copy the fixed code from ChatGPT, switch back to Bolt's Code View, paste it in, and save.
Why this is brilliant: You get the debugging power of frontier models without burning a single token on your Bolt account.
---
Workaround 5: The Local Eject Pipeline (Continue in Cursor / VS Code for $0)
When your application exceeds 20 files, continuing to build inside a browser chat window is an anti-pattern. The Context-Sync Tax becomes too heavy, and browser memory limits start causing latency.
This is the exact point where professional developers eject from Bolt.new and transition to their local machine.
!Bolt.new GitHub Sync and Eject
Figure 4: The GitHub sync modal in Bolt.new. One click exports the entire repository to GitHub so you can clone it locally.
The Step-by-Step Ejection Workflow:
Step 1: Sync to GitHub
In the top right corner of Bolt, click the GitHub icon. Authorize your account and click Create Repository. Bolt will automatically push all source files, configurations, and assets into a fresh, clean GitHub repository.
Step 2: Clone to Your Laptop
Open your terminal and clone the repository:
git clone https://github.com/your-username/your-bolt-app.git
cd your-bolt-app
Step 3: Install and Run Locally
Because Bolt scaffolds standard Vite + React applications, there is zero proprietary lock-in. Run:
npm install
npm run dev
In our local eject audit of Bolt.new's code, the entire application installed in under 60 seconds and compiled cleanly with zero terminal errors into a featherlight 185 KB bundle:
!Bolt.new Local Terminal Build
Figure 5: Bolt.new local build terminal: Clean exit code 0, 4 dependencies, compiled in 6.17 seconds.
Step 4: Continue Building with Free AI in Cursor or VS Code
Once the project is on your laptop, you no longer depend on Bolt's 300,000 token limit:
- Cursor Free Tier: Download Cursor IDE. It includes generous free AI completions and chat requests.
- VS Code + Roo Code / Cline: Install the free Roo Code or Cline extension in VS Code. Connect it to cheap model endpoints (like DeepSeek-V3 via OpenRouter) where 100,000 tokens cost less than $0.02—a fraction of the cost of SaaS subscriptions.
- Local LLMs (Ollama): If you have an Apple Silicon Mac or Nvidia GPU, connect VS Code to a local open-weights model (like Qwen 2.5 Coder) for completely free, infinite offline coding.
---
Bonus Alternative: Self-Host "Bolt.diy" (Zero Daily Limits)
If you love the Bolt.new in-browser WebContainer interface but refuse to pay $25/month for arbitrary token markups, the developer community created the ultimate solution: Bolt.diy.
Bolt.diy is an open-source, community-maintained clone of Bolt.new that runs locally on your machine or private server.
Why Bolt.diy Eliminates Token Walls:
1. Bring Your Own Key (BYOK): Instead of paying StackBlitz a flat monthly fee, you enter your own API keys from OpenRouter, Groq, Anthropic, or OpenAI.
2. True Pay-As-You-Go Economics: On Bolt.new, a heavy project can burn a $25/month plan in a week. On Bolt.diy using DeepSeek-V3 or Gemini Flash via API, building that exact same app typically costs between $0.30 and $1.50 in total API compute.
3. Local Offline Models: Bolt.diy natively connects to Ollama and LM Studio. You can run models like qwen2.5-coder:14b or deepseek-r1 directly on your hardware with zero token limits and zero cost.
How to Launch Bolt.diy in 2 Minutes:
git clone https://github.com/stackblitz-labs/bolt.diy.git
cd bolt.diy
pnpm install
pnpm run dev
Open http://localhost:5173 in your browser. You get the exact same multi-file AI coding workspace, WebContainer execution, and live preview—with zero daily caps.
---
The Strategic Cheat Sheet: What to Do at Every Stage
Bookmark this table. The next time you build on Bolt, follow this roadmap to stretch your tokens 5x:
| Project Stage | Token Danger Level | Recommended Action |
| :--- | :--- | :--- |
| 0 – 5 Files (Initial MVP) | 🟢 Safe (Low token burn) | Use AI for broad architecture & full-page scaffolding. |
| 5 – 15 Files (Iterating) | 🟡 Medium (Context rising) | Add .bolt/ignore. Use Code View for styling & text tweaks. |
| Code Breaks / Syntax Bug | 🔴 DANGER (Token trap) | Use Version History rollback. Do NOT re-prompt the AI. |
| 15 – 25+ Files (Advanced) | 🔴 Critical (Context-Sync Tax) | Eject to GitHub. Open in Cursor or VS Code. |
| Daily Limit Hit (0 Left) | ⛔ Blocked (Daily cap hit) | Edit manually in Code View, or clone locally with npm run dev. |
| Long-Term Scaling | 🟢 Infinite Runway | Transition to Bolt.diy with your own OpenRouter/DeepSeek keys. |
---
Frequently Asked Questions (FAQ)
When does the Bolt.new daily token limit reset?
The Bolt.new daily free token limit (300,000 tokens) resets every 24 hours on a rolling window from the time your account registered its first prompt, or at 00:00 UTC depending on your regional datacenter.
Can I buy more tokens without subscribing to Pro?
Currently, StackBlitz does not sell standalone one-off token top-ups for free accounts. To increase your hosted token pool on Bolt.new, you must upgrade to the Pro tier ($25/month), which grants 10 million monthly tokens and removes the 300,000 daily cap.
Does editing code manually in Bolt.new use tokens?
No. Toggling the Code View and typing directly in the Monaco editor uses zero AI compute. You can write hundreds of lines, add dependencies, and tweak CSS for free.
Why did my token limit burn on a tiny prompt?
Because Bolt.new charges for the entire context window on every execution. Even if your prompt was only 10 words, Bolt re-read your conversation history, file tree, and related components into the model, burning between 80,000 and 150,000 tokens behind the scenes.
---
Final Takeaway: Build Smarter, Not More Expensive
AI web builders are remarkable for rapid zero-to-one prototyping. Bolt.new proved in our benchmark lab that it can scaffold a production-ready, 4-dependency React MVP in under 9 minutes.
However, treating an in-browser AI tool like a full-lifecycle development environment is a recipe for depleted wallets and frozen previews.
The winning workflow in 2026 is simple:
1. Scaffold the core MVP on Bolt.new within your 300,000 daily free allowance.
2. Make micro-adjustments in Code View for $0.
3. The moment your project stabilizes, click the GitHub icon and eject.
Take your clean 185 KB codebase into Cursor or VS Code, and finish your application with complete developer freedom.
---
Authored by the AISparkHub Technical Research Team. Codebases, terminal logs, and live benchmarks verified across Bolt.new and StackBlitz WebContainers. Related benchmarks: Lovable vs. Bolt.new Head-to-Head and v0 vs. Lovable Code Quality Audit.
Mentions
9/ 🌐 bolt.diy prompt, run, edit and deploy full-stack web apps with any LLM you want. the open-source take on Bolt, on your own keys. 20k ★ stackblitz-labs/bolt.diy
How to Run Bolt.diy with Ollama on Windows: Step-by-Step Setup Guide (Tested on 8GB RAM) > Author's Note: This hands-on setup guide was originally tested and published with full high-resolution screenshots, latency logs, and reproducible troubleshooting steps on AISparkHub. --- Quick Answer: How to Run Bolt.diy with Ollama on Windows > To run Bolt.diy with Ollama locally on Windows: > 1. Install Git and Ollama via PowerShell: winget install -e --id Git.Git; winget install -e --id Ollama.Ollama > 2. Allow script execution: Set-ExecutionPolicy RemoteSigned -Scope CurrentUser -Force > 3. Pull a lightweight coding model: ollama pull qwen2.5-coder:1.5b > 4. Clone Bolt.diy and add OLLAMA_API_BASE_URL=http://127.0.0.1:11434 to a new .env.local file. > 5. Run pnpm install and pnpm run dev, then select Ollama in the web interface. Performance & Hardware Compatibility Matrix | Hardware Profile | Recommended Model | Generation Latency | RAM Consumption | Suitability | | :--- | :--- | :--- | :--- | :--- | | 8 GB RAM / Integrated GPU (Tested) | qwen2.5-coder:1.5b | 7m 42s (Full scaffold) | ~2.45 GB combined | Working (Local CPU) | | 16 GB RAM / No GPU | qwen2.5-coder:7b | ~3–5 mins | ~5.8 GB combined | Moderate | | 16 GB+ / NVIDIA GPU (8GB+ VRAM) | qwen2.5-coder:7b or deepseek-r1:8b | ~20–45 seconds | GPU offloaded | Optimal | | Any PC / Low Spec ($0 Cloud Fallback) | Groq / Gemini 2.0 Flash API | 1–3 seconds | < 500 MB | Fastest / 100% Free | --- Test Machine Specifications To provide realistic benchmarks for developers without expensive hardware, this guide was tested on a standard budget-tier laptop: - Operating System: Windows 10/11 Home (64-bit) - Processor: Intel Core i5 (CPU inference only) - Graphics: Intel(R) HD Graphics 620 (No dedicated NVIDIA VRAM) - Memory: 8 GB DDR4 (1.5 GB available physical memory at start) - Runtime Stack: Node.js v24.20, Ollama v0.40.12, pnpm v9.15.9, Bolt.diy v1.0.0 --- Step 1: Installing Git, Ollama, and PNPM Open PowerShell and install Git and Ollama in a single step using the native Windows Package Manager: winget install -e --id Git.Git; winget install -e --id Ollama.Ollama !Installing Git and Ollama via winget on Windows PowerShell When the Ollama setup window appears, click "Continue", then select "No thanks, I'll use Ollama locally". Ollama will minimize to your Windows system tray. Next, install pnpm globally: npm install -g pnpm --- Step 2: Resolving Common Windows Configuration Errors Windows users routinely encounter three specific environment errors that prevent Bolt.diy from running. Apply these configurations before proceeding: 1. Fix the PowerShell Execution Policy Error If running npm or pnpm produces this red error: File npm.ps1 cannot be loaded because running scripts is disabled on this system !PowerShell Security Error regarding Execution Policy Run this command in PowerShell to grant execution rights to your user account without requiring Administrator privileges: Set-ExecutionPolicy RemoteSigned -Scope CurrentUser -Force !Execution Policy resolved and pnpm installed cleanly 2. Enable Ollama Cross-Origin Requests (CORS) Because Bolt.diy runs in your browser via a local WebContainer, Windows security blocks requests to http://localhost:11434 unless CORS is explicitly permitted. Run: [System.Environment]::SetEnvironmentVariable('OLLAMA_ORIGINS', '*', 'User') 3. Resolving the PowerShell "QuickEdit" Console Freeze > [!WARNING] > If you click inside the blue PowerShell console, Windows enters text-selection mode. The console title will change to Select Windows PowerShell, and Windows will pause the Node.js process. > Resolution: Click inside the terminal and press the Esc key. The word Select will disappear, and the server will resume immediately. --- Step 3: Downloading the 8GB-Friendly Model Heavier models (such as 7B or 14B parameter models) consume 5GB–9GB of memory and will cause an 8GB Windows machine to freeze or crash. For smooth local CPU execution, pull Alibaba's specialized 1.5B coding model (~986 MB download): ollama pull qwen2.5-coder:1.5b To verify that Ollama is serving the model correctly, run: curl http://localhost:11434/api/tags --- Step 4: Cloning Bolt.diy & Environment Setup Clone the official repository: git clone https://github.com/stackblitz-labs/bolt.diy.git cd bolt.diy Configuring the Required .env.local File > [!IMPORTANT] > By default, Bolt.diy will hide Ollama from the model dropdown unless it detects a base URL configured in an environment file. Create a file named .env.local in the root bolt.diy directory and add the following lines: OLLAMA_API_BASE_URL=http://127.0.0.1:11434 DEFAULT_NUM_CTX=32768 Now install dependencies and start the development server: pnpm install pnpm run dev !Bolt.diy server running with secrets loaded from env.local --- Step 5: Connecting the Engine & The 7m 42s Benchmark Open your browser and navigate to: http://localhost:5173 With .env.local loaded, click the provider dropdown. You will see a green status indicator: • Ollama, with qwen2.5-coder:1.5b (1.5B) selected as the default model: !Bolt.diy interface showing Ollama connected with qwen2.5-coder 1.5b The Benchmark Test To benchmark real-world performance, we submitted a standard application prompt: Create a clean and responsive Pomodoro Timer app with Start, Pause, and Reset buttons, a 25-minute countdown, and a simple task checklist using React and Tailwind CSS. The Workbench Scaffolding Phase The local model immediately scaffolded the React template, configured package.json, and set up the directory tree: !Bolt Workbench creating initial project files on local CPU - Total Scaffolding Latency: 7 minutes 42 seconds (462s) on CPU. - Context Consumption: 4,421 tokens processed (3,996 prompt tokens, 425 completion tokens). - Financial Cost: $0.00 (Zero API fees; no subscription caps). --- Step 6: Overcoming the 1.5B Model Limitation ("No Preview Available") When you switch to the Preview tab after the initial prompt, you may see a blank screen or the placeholder message: "Start prompting (or editing) to see magic happen". Why This Happens (Technical Explanation) Sub-3B parameter models lack the instruction-following precision needed to parse Bolt's custom XML action tags (<boltAction type="file">). Instead of injecting code into src/App.tsx, the 1.5B model emits raw JSON tool calls: {"name": "project_create_directory", "arguments": {"path": "./src/components"}} The model creates the project structure, but leaves App.tsx with starter placeholder code. Resolving the Issue: Manual Code Insertion 1. Click the Code tab in Bolt.diy. 2. In the left file tree, click src $\rightarrow$ App.tsx. 3. Replace the placeholder code with the complete React timer component: import React, { useState, useEffect } from 'react'; export default function App() { const [seconds, setSeconds] = useState(1500); const [isActive, setIsActive] = useState(false); useEffect(() => { let interval: any = null; if (isActive && seconds > 0) { interval = setInterval(() => setSeconds((s) => s - 1), 1000); } else if (seconds === 0) { setIsActive(false); alert('Time is up!'); } return () => clearInterval(interval); }, [isActive, seconds]); const formatTime = (sec: number) => { const mins = Math.floor(sec / 60); const remainingSecs = sec % 60; return `${mins.toString().padStart(2, '0')}:${remainingSecs.toString().padStart(2, '0')}`; }; return ( <div className="min-h-screen bg-slate-900 text-white flex flex-col items-center justify-center p-4"> <div className="bg-slate-800 p-8 rounded-2xl shadow-xl text-center max-w-sm w-full border border-slate-700"> <h1 className="text-2xl font-bold mb-4 text-purple-400">Pomodoro Timer</h1> <div className="text-6xl font-mono font-bold mb-8 tracking-wider">{formatTime(seconds)}</div> <div className="flex gap-4 justify-center"> <button onClick={() => setIsActive(!isActive)} className={`px-6 py-2 rounded-lg font-semibold ${isActive ? 'bg-amber-500 hover:bg-amber-600' : 'bg-emerald-500 hover:bg-emerald-600'}`} > {isActive ? 'Pause' : 'Start'} </button> <button onClick={() => { setIsActive(false); setSeconds(1500); }} className="px-6 py-2 rounded-lg font-semibold bg-rose-500 hover:bg-rose-600" > Reset </button> </div> </div> </div> ); } 4. If the preview does not refresh automatically, click into the Bolt Terminal at the bottom and enter: npm run dev 5. Click back to the Preview tab. Hot Module Replacement (HMR) triggers immediately: !Live interactive Pomodoro timer rendered inside Bolt.diy preview The application is fully interactive, running inside your browser's local sandbox at zero cost. --- Architectural Recommendation: The $0 Hybrid Setup Running models locally via Ollama offers complete privacy and cost immunity. However, waiting 7+ minutes per generation on CPU hardware introduces friction during rapid prototyping. The Recommended Workflow for Budget Hardware: 1. Use Bolt.diy as Your Base: Keep bolt.diy running locally to retain unrestricted code exports and bypass commercial token paywalls. 2. Plug in Free Cloud API Tiers: In Bolt.diy's Settings, add a free API key from Google Gemini 2.0 Flash or Groq Cloud. Both providers offer generous free tiers that deliver sub-second generation speeds without straining your laptop CPU. 3. Keep Ollama for Offline Development: Use your local qwen2.5-coder:1.5b setup whenever you are traveling or coding without internet access. --- Frequently Asked Questions (FAQ) Can I run Bolt.diy without a dedicated GPU on Windows? Yes. You can run lightweight models like qwen2.5-coder:1.5b entirely on a CPU with 8GB of RAM. While project scaffolding takes around 7 minutes on entry-level hardware, it runs completely offline for $0. For instant generation without a GPU, you can also plug free cloud API keys (like Groq or Gemini Flash) directly into Bolt.diy. Why doesn't Ollama appear in the Bolt.diy provider dropdown? Bolt.diy hides Ollama by default if it cannot detect an active endpoint URL. You must create a file named .env.local in your root bolt.diy directory containing OLLAMA_API_BASE_URL=http://127.0.0.1:11434, then restart your local server with pnpm run dev. Is Bolt.diy completely free to use? Yes. Bolt.diy is an open-source, self-hosted alternative to Bolt.new. When paired with local models via Ollama or free developer API keys, you pay $0 in subscription fees and have zero daily token limits. Why does the preview window show "No preview available"? Lightweight models (under 3B parameters) sometimes struggle to trigger Bolt's automated XML code-injection tags. If the preview is blank, simply open the Code tab, paste your component logic into src/App.tsx, and run npm run dev inside the embedded Bolt Terminal. What is the best Ollama model for Bolt.diy? - For 8GB RAM / No GPU: qwen2.5-coder:1.5b (safely runs on CPU without freezing Windows). - For 16GB RAM / Dedicated GPU: qwen2.5-coder:7b or deepseek-r1:8b (delivers full automated code injection and faster generation).
Bolt.new Costs $25/Month. This Open-Source AI App Builder Is Free - bolt.diy Bolt.new Pro: $25/month (billed monthly). bolt.diy: free and open source (MIT). ⚡ ~20K GitHub stars. 1️⃣ Prompt to app: describe a full-stack web app and watch it build in your browser 2️⃣ Real workbench: code editor, terminal and live preview in one place 3️⃣ One-click deploys to Netlify, Vercel and GitHub Pages 🧠 Pick your AI: 20+ providers (OpenAI, Anthropic, Gemini, DeepSeek, Groq...) or local models with Ollama / LM Studio 💻 Run it on your own machine or self-host with Docker Note: bolt.diy is free; cloud AI providers charge for their own API usage. Use local models (Ollama) for $0. 🔗 stackblitz-labs/bolt.diy 📖 stackblitz-labs.github.io Follow for more free AI tools. #shorts #aitools #boltnew #opensource #vibecoding
How to Bypass Bolt.new Token Limits: 5 Workarounds That Actually Work (2026) > Author's Note: This benchmark was originally tested and published with full interactive comparison tables, memory profiles, and terminal execution logs on AISparkHub. --- That Sinking Feeling at 2:00 AM Picture this: You are deep in the zone building a brand-new SaaS idea on Bolt.new. The layout looks stunning, the components are clicking into place, and you feel like a 10x engineer who just discovered the future of vibe coding. You type a simple prompt: "Change the pricing card border to emerald-500 and center the checkout button." You hit Enter. Instead of a smooth preview reload, your screen freezes. A bright red toast slams across your dashboard: > "Token limit exceeded. You have reached your daily allowance of 300,000 tokens." Your project grinds to a dead halt. StackBlitz politely points you to a button: "Upgrade to Pro for $25/month." If you're an indie hacker, solo founder, or hobbyist building a prototype, spending $25/month before validating if anyone even wants your app hurts. Worse, if you don't understand why you ran out of tokens in the first place, upgrading won't even save you—you'll burn through that paid allowance in three days flat. Let’s get one thing out of the way immediately: There is no secret backend hack or illegal cheat code to bypass StackBlitz’s cloud billing. However, there are 5 battle-tested engineering workarounds that professional developers use to stretch their tokens, edit code for $0, stop burning millions of tokens in dumb error loops, and cleanly eject their projects to finish building for free. Here is the unfiltered, practitioner-grade guide to mastering the Bolt.new token limit in 2026. --- The Math: Why Does Bolt.new Burn Tokens Like Jet Fuel? Before we fix the problem, you need to understand why your tokens vanished after just two or three prompts. Most beginners assume Bolt works like ChatGPT. You send a 20-word prompt, you get a 100-word response, and you burn maybe 300 tokens. That is fundamentally NOT how Bolt.new works. Bolt runs an entire Node.js operating system inside your browser called a WebContainer. When you type a prompt, Bolt doesn’t just send your message to the AI. Behind the scenes, it bundles three massive payloads into every single execution cycle: 1. Your Entire Conversational History: Every prompt, every response, and every console error message since you opened the tab. 2. The Full Project File Tree: All directory paths, package configs (package.json, tsconfig.json, tailwind.config.js), and environment states. 3. The Raw Content of Every Touched Component: If your prompt touches the navbar, Bolt often reads the parent layout, the router, and related components into the model's context window. We call this the Context-Sync Tax. [Your 15-Word Prompt] + [Entire Accumulated Chat Log: ~35,000 tokens] + [Project File Tree & Configs: ~25,000 tokens] + [5 Source Code Files Re-Read: ~40,000 tokens] ════════════════════════════════════════════ Total Burn per Single Prompt: ~100,000+ Tokens! The Real Data From Our Benchmark Lab We didn't just guess this math; we measured it with a stopwatch and token counters in our systematic test (Lovable vs. Bolt.new Real Build Benchmark): - Prompt 1 (Initial App Scaffolding): Consumed 101,000 tokens (~33.6% of daily cap). - Prompt 2 (Adding Authentication & Persistence): Consumed 95,000 tokens (~31.6% of daily cap). - Total 2-Prompt Burn: 196,000 tokens. !Bolt.new Token Consumption Dashboard Figure 1: Our empirical benchmark dashboard showing 196,000 tokens burned across just two prompts on Bolt.new's free tier. Notice what happened: On Prompt 2, we only asked for a basic auth modal and local storage, yet it burned nearly 100,000 tokens. Why? Because the project had expanded to 27 files. Once your project reaches 20+ files, a one-line CSS color tweak can easily burn 120,000 to 150,000 tokens. On a free plan with a 300,000 daily cap, you hit an immovable brick wall by prompt #3. --- The 5 Workarounds to Keep Building for $0 --- Workaround 1: The "Code View" Free-Editing Hack (0 Tokens) The single biggest mistake developers make on Bolt.new is asking the AI chat box to make trivial adjustments. If you type: "Fix the typo in the header on line 14" or "Change the background from slate-900 to zinc-950", Bolt re-indexes your entire project context. You just burned 110,000 tokens for a five-second manual edit! The Fix: Bolt is not just a chatbot—it embeds a full Monaco code editor (the exact same engine powering VS Code) directly in your browser. !Bolt.new Code View Manual Editing Figure 2: Clicking the </> Code toggle at the top bar opens the full file tree and code editor, allowing manual edits with 0 token consumption. How to use it: 1. In the top navigation bar, toggle from the Preview eye icon to the </> Code button. 2. In the left-hand sidebar, click into any file (e.g., src/components/PromptCard.tsx, index.html, or .env). 3. Make your styling changes, fix typos, or paste logic directly into the editor. 4. Save the file (Ctrl + S on Windows or Cmd + S on Mac). 5. Look at the bottom terminal: Bolt's built-in Vite development server immediately hot-reloads your changes in real-time. Token Cost: Exactly 0. You can spend six hours polishing UI, tweaking colors, and refactoring logic without burning a single token from your daily allowance. --- Workaround 2: Use a .bolt/ignore File to Slash Context Burn by 50% By default, Bolt’s indexing crawler reads almost everything in your workspace when packaging context for the LLM. If your project contains large documentation files, sample data, mock JSON arrays, or heavy SVG icons, Bolt blindly packages all of that raw text into every prompt. You can blindfold the AI to non-essential files using a .bolt/ignore file—similar to how .gitignore works for git. How to set it up: 1. Open the Code View tab. 2. In your root directory, create a new file named .bolt/ignore (or place it inside the .bolt folder). 3. Add paths that the AI does not need to analyze on every prompt: # Ignore documentation and markdown guides README.md docs/ *.md # Ignore static mock data and heavy JSON files src/data/mockData.json public/assets/ # Ignore compiled build outputs and locks dist/ package-lock.json # Ignore test suites during UI prototyping **/*.test.ts **/*.spec.tsx The Impact: Shrinking your active context tree prevents Bolt from sending unnecessary kilobytes of text to the LLM. In our testing, properly excluding heavy assets and lockfiles reduced prompt context weight by 35% to 50%, effectively doubling your daily prompt capacity. --- Workaround 3: Use Version History Instead of "Error-Loop Prompting" Here is a horror story that drains millions of tokens every day: 1. Bolt generates a component that introduces a TypeScript error. 2. The user types: "Fix the error." (Cost: 110,000 tokens) 3. Bolt tries to fix it, but breaks another file. 4. The user types: "You broke the layout, undo that and fix it." (Cost: 130,000 tokens) 5. Result: 240,000 tokens burned in 3 minutes, your daily limit is gone, and the code is worse than when you started. Never use AI chat prompts to undo mistakes. Bolt.new includes a native Version History checkpoint system built directly into the UI. !Bolt.new Version History Rollback Figure 3: Bolt's Version History drawer lets you review and revert to any prior working checkpoint instantly for 0 tokens. How to use it: 1. On the left chat panel, click the clock icon or look at your prompt history checkpoints. 2. Click Version History. 3. You will see a chronological list of every stable checkpoint Bolt saved (e.g., “Create PromptVault prompt library - Oct 01”). 4. Click on the last working checkpoint and select Revert to this version. Bolt instantly rolls your entire WebContainer back to the exact working state. Total token cost: 0. --- Workaround 4: The External LLM Debugging Trick When your app throws a complex runtime error or a broken React hook, do not feed the error back into Bolt’s chat. Instead, let free external models do the heavy lifting: 1. Copy the error message and the broken file’s code from Bolt’s Code View. 2. Open a separate tab with ChatGPT (Free GPT-4o), Claude.ai, or Google Gemini. 3. Prompt the external model: > "I have a Vite + React component throwing this TypeScript error: [Paste error]. Here is the file: [Paste code]. Give me ONLY the corrected code block to paste back in." 4. Copy the fixed code from ChatGPT, switch back to Bolt's Code View, paste it in, and save. Why this is brilliant: You get the debugging power of frontier models without burning a single token on your Bolt account. --- Workaround 5: The Local Eject Pipeline (Continue in Cursor / VS Code for $0) When your application exceeds 20 files, continuing to build inside a browser chat window is an anti-pattern. The Context-Sync Tax becomes too heavy, and browser memory limits start causing latency. This is the exact point where professional developers eject from Bolt.new and transition to their local machine. !Bolt.new GitHub Sync and Eject Figure 4: The GitHub sync modal in Bolt.new. One click exports the entire repository to GitHub so you can clone it locally. The Step-by-Step Ejection Workflow: Step 1: Sync to GitHub In the top right corner of Bolt, click the GitHub icon. Authorize your account and click Create Repository. Bolt will automatically push all source files, configurations, and assets into a fresh, clean GitHub repository. Step 2: Clone to Your Laptop Open your terminal and clone the repository: git clone https://github.com/your-username/your-bolt-app.git cd your-bolt-app Step 3: Install and Run Locally Because Bolt scaffolds standard Vite + React applications, there is zero proprietary lock-in. Run: npm install npm run dev In our local eject audit of Bolt.new's code, the entire application installed in under 60 seconds and compiled cleanly with zero terminal errors into a featherlight 185 KB bundle: !Bolt.new Local Terminal Build Figure 5: Bolt.new local build terminal: Clean exit code 0, 4 dependencies, compiled in 6.17 seconds. Step 4: Continue Building with Free AI in Cursor or VS Code Once the project is on your laptop, you no longer depend on Bolt's 300,000 token limit: - Cursor Free Tier: Download Cursor IDE. It includes generous free AI completions and chat requests. - VS Code + Roo Code / Cline: Install the free Roo Code or Cline extension in VS Code. Connect it to cheap model endpoints (like DeepSeek-V3 via OpenRouter) where 100,000 tokens cost less than $0.02—a fraction of the cost of SaaS subscriptions. - Local LLMs (Ollama): If you have an Apple Silicon Mac or Nvidia GPU, connect VS Code to a local open-weights model (like Qwen 2.5 Coder) for completely free, infinite offline coding. --- Bonus Alternative: Self-Host "Bolt.diy" (Zero Daily Limits) If you love the Bolt.new in-browser WebContainer interface but refuse to pay $25/month for arbitrary token markups, the developer community created the ultimate solution: Bolt.diy. Bolt.diy is an open-source, community-maintained clone of Bolt.new that runs locally on your machine or private server. Why Bolt.diy Eliminates Token Walls: 1. Bring Your Own Key (BYOK): Instead of paying StackBlitz a flat monthly fee, you enter your own API keys from OpenRouter, Groq, Anthropic, or OpenAI. 2. True Pay-As-You-Go Economics: On Bolt.new, a heavy project can burn a $25/month plan in a week. On Bolt.diy using DeepSeek-V3 or Gemini Flash via API, building that exact same app typically costs between $0.30 and $1.50 in total API compute. 3. Local Offline Models: Bolt.diy natively connects to Ollama and LM Studio. You can run models like qwen2.5-coder:14b or deepseek-r1 directly on your hardware with zero token limits and zero cost. How to Launch Bolt.diy in 2 Minutes: git clone https://github.com/stackblitz-labs/bolt.diy.git cd bolt.diy pnpm install pnpm run dev Open http://localhost:5173 in your browser. You get the exact same multi-file AI coding workspace, WebContainer execution, and live preview—with zero daily caps. --- The Strategic Cheat Sheet: What to Do at Every Stage Bookmark this table. The next time you build on Bolt, follow this roadmap to stretch your tokens 5x: | Project Stage | Token Danger Level | Recommended Action | | :--- | :--- | :--- | | 0 – 5 Files (Initial MVP) | 🟢 Safe (Low token burn) | Use AI for broad architecture & full-page scaffolding. | | 5 – 15 Files (Iterating) | 🟡 Medium (Context rising) | Add .bolt/ignore. Use Code View for styling & text tweaks. | | Code Breaks / Syntax Bug | 🔴 DANGER (Token trap) | Use Version History rollback. Do NOT re-prompt the AI. | | 15 – 25+ Files (Advanced) | 🔴 Critical (Context-Sync Tax) | Eject to GitHub. Open in Cursor or VS Code. | | Daily Limit Hit (0 Left) | ⛔ Blocked (Daily cap hit) | Edit manually in Code View, or clone locally with npm run dev. | | Long-Term Scaling | 🟢 Infinite Runway | Transition to Bolt.diy with your own OpenRouter/DeepSeek keys. | --- Frequently Asked Questions (FAQ) When does the Bolt.new daily token limit reset? The Bolt.new daily free token limit (300,000 tokens) resets every 24 hours on a rolling window from the time your account registered its first prompt, or at 00:00 UTC depending on your regional datacenter. Can I buy more tokens without subscribing to Pro? Currently, StackBlitz does not sell standalone one-off token top-ups for free accounts. To increase your hosted token pool on Bolt.new, you must upgrade to the Pro tier ($25/month), which grants 10 million monthly tokens and removes the 300,000 daily cap. Does editing code manually in Bolt.new use tokens? No. Toggling the Code View and typing directly in the Monaco editor uses zero AI compute. You can write hundreds of lines, add dependencies, and tweak CSS for free. Why did my token limit burn on a tiny prompt? Because Bolt.new charges for the entire context window on every execution. Even if your prompt was only 10 words, Bolt re-read your conversation history, file tree, and related components into the model, burning between 80,000 and 150,000 tokens behind the scenes. --- Final Takeaway: Build Smarter, Not More Expensive AI web builders are remarkable for rapid zero-to-one prototyping. Bolt.new proved in our benchmark lab that it can scaffold a production-ready, 4-dependency React MVP in under 9 minutes. However, treating an in-browser AI tool like a full-lifecycle development environment is a recipe for depleted wallets and frozen previews. The winning workflow in 2026 is simple: 1. Scaffold the core MVP on Bolt.new within your 300,000 daily free allowance. 2. Make micro-adjustments in Code View for $0. 3. The moment your project stabilizes, click the GitHub icon and eject. Take your clean 185 KB codebase into Cursor or VS Code, and finish your application with complete developer freedom. --- Authored by the AISparkHub Technical Research Team. Codebases, terminal logs, and live benchmarks verified across Bolt.new and StackBlitz WebContainers. Related benchmarks: Lovable vs. Bolt.new Head-to-Head and v0 vs. Lovable Code Quality Audit.
Build and run web apps from plain-language prompts with bolt.diy, choosing the AI model you want instead of being tied to one provider. stackblitz-labs/bolt.diy t.co