See what your LLM prompts cost — token count and dollar cost — inline, above every call site, as you write them.
This extension wraps PromptScan, a static analyzer for LLM call sites. It reads your code (no runtime, no API key) and annotates each OpenAI / Anthropic / LangChain / litellm / Vercel AI SDK call with its estimated input cost.
What you get
- Inline cost CodeLens above every detected call site:
⚡ ~$0.0029/call · ~587+ tok · claude-opus-4-8 · +2 tools. Markers match the CLI —~an approximate (Anthropic proxy) tokenizer,+a floor where only static content could be counted. - Diagnostics in the Problems panel: prompt-caching opportunities, possibly-unused prompts, and duplicated prompts, each on the line that owns it.
- PromptScan: Scan Workspace command for a whole-project summary.
Because it's static analysis, it reports only what the source makes knowable: a prompt built at runtime is marked unresolved rather than guessed. Notably, a call whose prompt is a runtime value but whose tool schemas are static still gets a real number — the tools are often the larger cost.
Settings
| Setting | Default | Description |
|---|---|---|
promptscan.enableCodeLens |
true |
Show the token/cost CodeLens above each call site. |
promptscan.enableDiagnostics |
true |
Report caching opportunities, dead prompts, and duplicates as diagnostics. |
Notes
- Results reflect the saved file — the scanner reads from disk, so annotations refresh on save.
- Anthropic token counts use a calibrated
cl100kproxy (Anthropic publishes no tokenizer); OpenAI uses the exacttiktokenencoding. See the main README for the accuracy details.
Development
npm install
npm run build # tsc → out/
npm test # headless unit tests of the presentation logic
npm run package # build a .vsix (vsce)
Press F5 in VS Code to launch an Extension Development Host and try it on a real file.