How to Cut Bolt.new Token Costs on Bigger Apps
Bolt.new bills per token, and the token counter climbs fastest on the boring stuff: re-reading files, reinstalling packages, and rebuilding screens that already worked. A multi-screen app that feels free on day one can cost $30–$60 in a week if you prompt like a chatbot.
This is the second stage. You already built something small. Now you want a real app without watching the quota meter.
Why do Bolt.new sessions get more expensive as the app grows?
Because the context grows too. Bolt.new has to hold the relevant files in context to make an edit, and a 40-file React project means every request carries a heavier payload than the 4-file prototype you started with.
Three multipliers in order of impact:
- File count. Splitting one 900-line component into 12 files feels cleaner and costs more per edit.
- Reinstalls. Any dependency change can trigger a fresh
npm installinside the WebContainer, which shows up as slow turns, not tokens — but you retry more, and retries cost. - Screenshots and errors pasted in. Pasting a full console dump of 300 lines is cheaper than pasting it three times because you forgot the fix failed.
How do I scope requests so Bolt only touches one file?
Name the file. Every time.
- Weak: “The chart looks wrong, make it show revenue per month.”
- Strong: “In src/components/RevenueChart.tsx, change the data source to
monthlyRevenuefrom useDashboardData and format Y values as$12k.”
The second one usually lands in one turn. The first often produces a new component, a new helper, and a duplicate of the chart.
Two habits that compound:
- Keep a short
NOTES.mdin the repo root listing the folder map. Ask Bolt to read it at the start of a session: “Read NOTES.md before changing anything.” - Batch cosmetic fixes. Three small styling changes in one message beat three separate turns, because the file is only read once.
How do I add a real database without breaking the build?
Bolt can connect to Supabase. The failure mode is schema drift: you ask for a feature, Bolt writes SQL, and the table in Supabase doesn’t match what the code expects.
Order of operations that works:
- Create the Supabase project manually, tables and columns included. Do not let Bolt invent your schema.
- Paste the schema (table names, column names, types) into the chat and ask Bolt to read it as the source of truth.
- Wire one read query first. Confirm the preview shows rows before asking for writes.
- Add auth last. Auth changes touch every fetch call and will re-generate files.
If the app uses Supabase auth, tell Bolt the env var names exactly (VITE_SUPABASE_URL, VITE_SUPABASE_ANON_KEY). Guessing names is a reliable way to get a blank preview.
When is exporting cheaper than staying in Bolt.new?
Sooner than most people expect.
Export to GitHub when any of these is true:
- The app has more than about 25 files
- You need a package Bolt struggles with (native modules, WebSockets, anything server-side heavy)
- You’re doing three or more small edits a day on the same files
- You need real CI or tests
Once exported, the same code works in VS Code or Cursor, where you pay per completion rather than per whole-file rewrite. Bolt remains useful for new screens: paste one component in, get a scaffold, paste it back. That hybrid is common among people shipping real products.
Deployment note: Bolt’s Netlify deploy is convenient, but environment variables set in the chat do not travel to Netlify. Set them in the Netlify dashboard before the first deploy or the site will build and return 500s.
Who should not push Bolt.new this far?
If your app handles payments, personal health data, or anything with an audit trail, build it in a normal repo with code review from the start. Bolt is a great place to prove an idea and a bad place to hold a schema you can’t explain.
Also skip the deep path if you don’t read JavaScript at all. At some point the chat stops working and the code tab is the only way out.
Quick Summary
- Token spend scales with file count and request size, not with how impressive the app looks.
- Always name the target file in a request; unnamed requests cause duplicate components.
- Create Supabase tables yourself, paste the schema in, wire reads before writes, add auth last.
- Export to GitHub at roughly 25 files or when edits become daily and small.
- Env vars must be set in Netlify separately; Bolt chat values don’t carry over.
