To learn more about large language models generally, see Introduction to LLMs.
Build with your preferred agent
Bolt gives you the choice of which AI agent powers your builds, so you can work in the way that suits you best. You can choose between two options:- Claude Agent – Our most capable builder in Bolt and the best choice for production-quality apps and larger development work. It produces more complete results, makes stronger decisions, and handles complex reasoning with fewer errors. It may take a little longer to run and can use more tokens, although the higher quality often saves time overall. Claude Agent is powered by the Claude model family from Anthropic, and you can choose the level of performance you want, including fast and efficient Haiku 4.5, the everyday Sonnet 4.6, or the powerful Opus 4.6. Plan Mode is also available with Claude Agent, which helps you develop detailed build strategies and execute them accurately.
- v1 Agent (legacy) – Based on the original Bolt experience, this agent is more likely to produce incomplete apps and require extra fixes, so it’s not recommended for production work. It is, however, faster, uses fewer tokens per prompt, and works well for quick prototypes or testing layouts and design ideas. It uses Anthropic Claude Sonnet as its LLM.
The v1 agent is being retired. Starting April 13, 2026, you won’t be able to select v1 Agent (legacy) when starting a project. After August 3, 2026, you won’t be able to access projects that use the v1 agent at all. Make sure to switch to Claude Agent before that date.
Claude Agent vs. v1 Agent (legacy) efficiency
While v1 Agent (legacy) might seem more efficient because it uses fewer tokens at the start, this can be misleading. For any project beyond a quick mockup or prototype, you will likely save tokens overall by using Claude Agent. Its more comprehensive approach to building leads to higher quality planning, greater reliability, and fewer mistakes. This means less rework, fewer wasted calls, and more efficient use of your token budget over time.Switch between Bolt agents and models
Not all models are available with the free plan. To be able to switch to any model, upgrade to a paid plan.
If you are already using Claude Agent, switching between models in Claude Agent does not clear your chat history.
Switch agents or models on the homepage
- Haiku 4.5
- Sonnet 4.6 (default)
- Opus 4.6
Switch agents or models in a project
- In the bottom-left corner of the chatbox, click the current model name.
- Select the agent or model you want to use.
- Confirm your choice.
Differences between agents
Bolt behaves slightly differently depending on which agent you choose. Some tools and interface elements adjust to match each agent’s strengths. Key differences include:- Claude Agent works with Plan Mode. v1 Agent (legacy) works with Discussion Mode.
- Claude Agent lets you choose between Haiku, Sonnet, and Opus models.
- Claude Agent can create Bolt Databases and connect projects to Supabase.
| Feature | Claude Agent | v1 Agent (legacy) |
|---|---|---|
| Plan Mode | ✔ | ✗ |
| Create Bolt databases | ✔ | ✗ |
| Select your model | ✔ | ✗ |
| Works with Bolt databases | ✔ | ✔ |
| Connects to Supabase | ✔ | ✔ |
| Discussion Mode | ✗ | ✔ |
Model selection inside Claude Agent
Model selection controls how Claude Agent behaves in your project. Each model reflects a different balance of speed, reasoning depth, and token usage. For most users, Sonnet 4.6 is the default and recommended starting point. It offers a strong balance of speed, cost, and reasoning depth, making it reliable for most development work. Haiku 4.5 fits lightweight or high-volume tasks. Opus 4.6 is best suited for the hardest problems, large architectural decisions, or work where accuracy matters more than speed.These models are not a simple ladder where each one is strictly “better” than the last. Each model reflects a different set of tradeoffs between speed, reasoning depth, token cost, and behavior. A model that excels at one type of task may not be the best fit for another. The best way to find what works for your projects is to experiment. Try different models on the kinds of tasks you do most and see which one gives you the results you want.
Haiku 4.5
Fast and token-efficient for simple tasks and rapid iteration.Best for:
- Quick UI or styling changes
- Content edits and translations
- Test or seed data
- Straightforward static pages
- Tasks where the solution is already clear
Sonnet 4.6
The default model in Bolt. Strong reasoning depth with speed and efficiency built for all-day use.Sonnet 4.6 is the model most users should start with, and the one Bolt selects by default. It offers a balance of speed, cost, and reasoning depth for everyday building tasks. It performs well across planning, coding, debugging, and iterative development. It can hold a complex plan across multiple steps and execute without losing the thread.It can handle multi-file refactors, end-to-end feature builds, and coordinated front-and-backend changes smoothly. For most development work, Sonnet 4.6 delivers excellent results while staying fast and token-efficient enough to use throughout the day.Best for:
- Scaffolding new apps and features
- Backend logic and database work
- Multi-step feature builds spanning front and backend
- Complex refactors touching many files
- General-purpose development where reliability and speed both matter
Opus 4.6
The most capable model available in Bolt, with the deepest reasoning and greatest sensitivity to your instructions.Opus 4.6 is tuned to respond more directly to how you prompt it. If you tell it to think carefully about a specific problem, it invests deeper reasoning in that area in a way other models may not. This makes it especially well-suited for navigating large, existing codebases and reasoning through messy or legacy code where understanding tangled dependencies and older patterns matters. For new projects or tasks where Sonnet 4.6 already performs well, Opus 4.6 may not show a meaningful difference. Its strength emerges in the hardest problems where precise, directed reasoning makes a real impact.Best for:
- The most complex architectural and system-wide decisions
- Brownfield development in large, tangled codebases
- Tasks where you want fine control over reasoning depth through your prompts
- High-stakes work where the extra reasoning investment is justified
Comparing models
| Model | Speed | Reasoning depth | Best use |
|---|---|---|---|
| Haiku 4.5 | Fastest | Light | Quick edits and iteration |
| Sonnet 4.6 (default) | Balanced | Deep | Most everyday development work |
| Opus 4.6 | Slower | Deepest | Advanced reasoning, legacy codebases, prompt-directed deep thinking |
Using claude.md files with agents
Bolt agents now support importingclaude.md files directly into projects. When a claude.md file is present, it is detected and used automatically, with no additional setup required.
To upload a file, follow the steps below:
- Log in to your project.
- Click the plus icon in the chatbox.
- Click Attach file, and choose your file.
claude.md file is especially helpful for users who rely on Project Knowledge settings. Instead of adding content manually through the UI, you manage project context in Markdown files. This makes it easier to include links, organize information across multiple files, and maintain project knowledge in a more flexible and scalable way.
Important: The
claude.md file acts as the entry point for agent instructions. While it can reference other files, such as text or additional Markdown files, the primary instructions must live in claude.md itself. The file name is required and cannot be changed. Agents only look for instructions starting from claude.md, then follow any links or references you include from there.