# Bolt.new AI App Builder: How to Build Without Burning Tokens

> How Bolt.new's AI agents spend tokens and how to build with fewer: Plan Mode first, Standard vs Max, one feature per prompt, /clear and free restores.

_Source: https://professionalstoolkit.com/articles/bolt-new-ai-app-builder — The Professional's Toolkit · updated November 2026_

---

Bolt.new (Bolt) is a browser-based AI app builder whose agent writes and runs JavaScript code from your prompts, and it meters all of that work in tokens. Building well means spending tokens on decisions, not rework. **Plan in Plan Mode, build one feature per prompt on the Standard agent, and undo through version history rather than by prompting again.** The Free plan ($0 a month) gives you 1M tokens a month with a 300K daily limit. Pro starts at 10M tokens a month for $25 per month billed monthly or $18 per month billed yearly ($216 a year). For scale, Bolt's own quickstart example spent about 204K tokens on the first build of a small app.


**~204K** — Tokens Bolt's quickstart example spent on a small app's first build (vendor figure).



- **1M / 300K** — Free tokens a month / daily limit
- **20** — Prompts the queue can hold while Bolt builds
- **30 a day** — Project security audits on paid plans, no tokens


## Why a prompt gets pricier as the app grows

Tokens measure the work Bolt does: reading your prompt and the project files, thinking, planning and writing code. The detail that shapes every habit below is where most of that work goes. Bolt's help center says most token usage comes from reading and syncing project files, so a larger project uses more tokens per message. In the same quickstart, a small title change cost about 70K tokens. That was a tiny app. The context Bolt carries with each prompt keeps growing as the codebase does, and nothing in a long chat is free to reread.

Two kinds of file inflate that context quietly. Bolt may add Markdown record-keeping files that it reads on every prompt, and files of about 700 lines or more fill the context fast; refactoring them costs tokens once but lowers later use. [1] Typing /clear in the chatbox resets the conversation: Bolt keeps the codebase but forgets recent chat history and how the code changed, so requests use fewer tokens. The same command often unsticks an unresponsive session. A .bolt/ignore file keeps files out of view, and running Knip in the Code view terminal to strip unused files costs nothing.

Connectors deserve the same discipline. Each connector (Bolt's MCP integrations with tools such as Notion, Linear or GitHub) adds information to the context, and several switched on together can slow Bolt, lower output quality and raise token use. Turn one on only where needed. The plan selector's estimate of about 100 messages for 10M tokens is Bolt's guess for a typical project, not an allowance. Paid tokens stay usable for up to two months from the start of the cycle they were issued in, only while you stay subscribed, and Free tokens never roll over, as our [Bolt.new pricing guide](/articles/bolt-new-pricing) explains.

## Choosing between Standard and Max for each task

Since May 2026 Bolt no longer has you pick a language model; it offers two agents and chooses the model behind them. **Standard** is built for speed and token efficiency on small or medium apps and well-defined tasks. It runs on every plan and is the only agent on Free. **Max** reasons more deeply for large apps, refactoring and open-ended work, uses more tokens per prompt, and needs a paid plan (Pro, Teams or Enterprise). Bolt's own guidance is blunt: Max on a simple task wastes tokens. Default to Standard.

One setting trips people up. New projects start on Standard, but an existing project keeps the last agent you chose, so a project you moved to Max for a refactor stays there until you switch it back. There are two sensible exceptions to the Standard habit. Bolt recommends Max when you build against a design system (adding your own needs a paid Teams plan; the pre-loaded ones work on any plan), because it follows complex instructions more consistently. And Dynamic reasoning, an account setting on paid plans that is off by default, adds deeper reasoning to complex prompts at the cost of more tokens. Leave it off until a specific job needs it.


> 💡 **Diagnose bugs in Plan Mode:** Ask Bolt to find the cause in Plan Mode, which changes no code and uses fewer tokens, then switch to Build Mode for one targeted fix.


## Planning before Bolt touches the code

Plan Mode lets you plan and ask questions without changing a line of code, using the codebase and your recent messages as context. It can also pull real-time information from trusted web sources and lists them at the top of the reply. It uses fewer tokens than Build Mode, though not zero. [2] When the plan looks right, the Implement this plan button switches to Build Mode and applies it. You can even start a new project in Plan Mode from the homepage; if you want a mobile app, say so in that first prompt. Older tutorials call this Discussion Mode, which Bolt retired in August 2026.

Debugging is where planning pays most. Every fix attempted in Build Mode spends tokens, so Bolt advises diagnosing the cause in Plan Mode first and only then building the fix. The Attempt fix button looks harmless. It is not: each click bills tokens, and Bolt's blog suggests restoring an earlier version and describing the problem once instead of clicking repeatedly. Ginger Pulse, a nutrition app whose first version came from the 2025 Bolt hackathon, follows a similar rhythm: its founder plans in Plan Mode before writing code and uses GitHub branches so beta users never see a broken build.

## Prompts that carry one feature and nothing extra

Bolt's blog recommends building one feature at a time, and its prompting guide adds the other half: many small prompts about the same screen cost more than one prompt that gathers them. So scope each prompt to a single feature but put every instruction for that feature into it. Narrow Bolt's reading as well. Type @ to tag a specific file or folder, lock a file or a whole folder against changes from Code view, and name components precisely; Bolt says specific instructions improve accuracy and token efficiency because it reads less of the code. The built-in Enhance prompt feature asks guided questions to tighten a vague request.

Rules you would otherwise repeat belong somewhere permanent. Knowledge holds standing instructions at account, project or team level, applies to every prompt and survives a /clear. An agents.md file in the project is picked up automatically. Skills, available on every plan, apply only when a prompt matches their description, and Bolt reads just the name and description until then, so turning on more skills barely moves token use. In team workspaces only admins add skills to the shared library, though any member can write a project skill. A skill cannot call outside tools; that takes a connector.

While Bolt builds, you need not wait. The prompt queue, added in August 2026, holds up to 20 prompts (five visible at a time) and each keeps the agent it was written for, so a Max refactor can sit behind two Standard tweaks. Dictation is convenient for long prompts, but it bills tokens for transcription even if you never send the message.

## The undo and edit paths that cost nothing

| Action in Bolt | Tokens |
|---|---|
| Editing in Code view, running terminal commands | None |
| Restoring a version from version history | None |
| Publish and Update buttons | None |
| Run security audit button and its fixes (paid plans) | None |
| Visual edits before you save | None |
| Saving a batch of visual edits | Yes, fewer than separate edits |
| Each Attempt fix click | Yes |
| Asking Bolt by prompt to publish, audit or undo | Yes |
| Plan Mode prompt | Yes, fewer than Build Mode |

Code view is a full editor and terminal on desktop Chrome or another Chromium browser; on Safari or a phone it is read-only. It is also the escape hatch on Free: at the daily limit you can keep editing code directly until the AI bar refreshes. Two cautions apply. Deleting a file there is immediate, with no undo, and version history does not track direct code edits, so commit to GitHub or download a copy before large manual changes.

Visual edits, opened with the Select button on every plan, let you change text and styling in the preview (colors, font sizes from 8 to 96 px, links, alt text) for free until you save. One batch can be saved per page. There is no dragging, no layout rearranging and no bulk change across many elements, and a saved batch can only be rolled back through version history.

Bolt saves a version after every processed prompt, under the clock icon. Restoring one is free, while prompting Bolt to undo a change charges you again. **Restoring never rolls back your database**, whether Bolt Database or your own Supabase, and a version cannot be restored while Bolt is still processing a prompt. For risky experiments, duplicate the project first; the copy takes the database schema and starter data but not users, user data, files or secrets.


> ⚠ **Restores leave your data alone:** Restoring a version rolls back code but never your Bolt Database or Supabase data, so export or back up data before risky changes.


## A security pass before real users arrive

Fast building leaves gaps, and Bolt ships two checks. The database Security Audit runs on every plan, looks at the database only and uses the same checks as the Supabase Security Advisor. The project security audit, in the Publish menu on paid plans, reviews code and database together: access control, sign-in, business logic, Row Level Security, information disclosure, secrets and injection. It fixes most issues on its own and flags the architectural ones. [3] The button costs no tokens and allows up to 30 audits a day, resetting at midnight UTC; asking Bolt to audit by prompt does cost tokens.

Bolt describes the audit as a first pass, not a replacement for enterprise security tooling, and the Terms leave reviewing and testing AI output with you. Two defaults help. Tables created from late October 2026 on are private until your app grants explicit read and write permissions, and Bolt turns on Row Level Security for tables it creates, while tables you add by hand need checking. Secrets live with server functions and never reach the browser. Policy limits, from refunds to the AI-training default, are covered in our [Bolt.new review](/articles/bolt-new-review).


> 💡 **Restore instead of re-prompting:** Restoring a version from the clock icon uses no tokens, while asking Bolt to undo a change or clicking Attempt fix bills tokens every time.



## FAQ

**What happens when I run out of tokens in Bolt?**

On Free, AI prompts stop at the 300K daily limit until the daily bar refreshes, and once the 1M monthly allowance is spent they stop until the monthly bar resets on the 1st; you can keep editing in Code view for free either way. Paid plans have no daily limit; more tokens come from a higher token level or a reload, which only the highest monthly Pro level or an annual Pro plan can buy.

**Can I control Bolt's agent from my own editor or an API?**

Bolt has no public API or SDK for controlling it from outside. An invite-only Bolt CLI beta, free while in beta, can push a local folder into Bolt, pull a project back and fetch chat history.

**Does Bolt train AI models on my prompts?**

Unless you opt out, content created from October 2026 on an individual account outside the EEA, UK and Switzerland may be used for AI training and licensed as de-identified datasets. The opt-out is free on any plan and does not undo training already done; Teams and Enterprise workspaces are excluded.

**Can I build with Bolt on a phone?**

Bolt works in mobile browsers for starting projects, chatting, previewing and publishing. Code view editing, visual edits and the database screens need a desktop browser.


## The bottom line

Bolt's agent rewards builders who treat tokens as a budget for thinking, not for trial and error. Plan the feature in Plan Mode, build it on Standard in one well-scoped prompt, keep standing rules in Knowledge or skills, and clear the chat before the context bloats. When something breaks, restore a version instead of clicking Attempt fix, and keep your data backed up because restores leave it untouched. Save Max for refactors and design-system work, and run the security audit before you share the app. Do that, and the same token allowance goes noticeably further.



## Sources

[1] Bolt help: Manage project context — https://support.bolt.new/prompting/manage-context (2026-11)
[2] Bolt help: Plan before building with Plan Mode — https://support.bolt.new/prompting/plan-mode (2026-11)
[3] Bolt help: Check your project's security — https://support.bolt.new/building/security (2026-11)
[4] Bolt.new pricing page — https://bolt.new/pricing
[5] Bolt help center: tokens — https://support.bolt.new/account/tokens
[6] StackBlitz Terms of Service — https://stackblitz.com/terms-of-service
[7] StackBlitz Privacy Policy — https://stackblitz.com/privacy-policy