# PromptHelm

> Prompt management for LLM apps. Change the prompts your LLM app sends without a redeploy. Publish in the dashboard, roll back in one click, and every running copy picks it up within a minute.

PromptHelm stores the prompts your app sends to an LLM. You edit and publish them in a dashboard; your app fetches the live version at runtime with the TypeScript SDK (`@prompt-helm/sdk`, Node.js 18 and later) or one HTTP request. Your app still calls the LLM itself: PromptHelm never sees your provider keys or your traffic.

- Every publish is a new version that never changes, and rolling back is one click.
- The SDK caches each prompt for 60 seconds and gives up on a request after 3 seconds. With a fallback, any failure returns your own text, so your app keeps working.
- Variables (`{{name}}`) are filled in on your server; their values never reach PromptHelm.
- A version can also carry model settings (`model`, `temperature`, `maxTokens`), which the SDK's `resolve` returns as `config`: moving to another model is a publish that rolls back in one click and can be A/B tested. Your code still makes the model call.
- API keys are read-only and belong on your server.
- On Pro, an A/B test gives a new version a share of the calls while the live version keeps the rest. The SDK picks the version per call on your server (a `unit`, such as a user id, keeps each user on one), and the app reports each call's outcome and token usage; PromptHelm gets totals only and says, with an interval that holds however often it's checked, whether the new version did better, and at what cost in tokens and latency.
- `npx @prompt-helm/sdk pull` writes every published prompt to a file in your repo, to pass as the fallbacks; Settings exports a whole project, every version included, as JSON.
- Pricing: Free covers 10 prompts and 2 projects, with no card. Pro is $19 a month or $190 a year for 200 prompts, unlimited projects and A/B tests, paid through Paddle. Requests are never billed.

## Docs

- [Quickstart](https://prompthelm.app/quickstart.md): set up the SDK or the HTTP API, with the full reference and notes for coding agents
- [Everything in one file](https://prompthelm.app/llms-full.txt): this summary, the quickstart and the FAQ
- [Pricing](https://prompthelm.app/pricing): plans and what counts as a prompt

## Guides

- [How to change an LLM prompt without redeploying](https://prompthelm.app/guides/change-prompts-without-redeploying): A prompt in your code needs a release for every word. Where else it can live, what each option costs, and the runtime pattern that keeps your app up.
- [Prompt versioning and rollback for LLM apps](https://prompthelm.app/guides/prompt-versioning): Why every published prompt should be a version that never changes, how rollback becomes a pointer move, and how to tell which version gave an answer.
- [What is prompt management? A guide for LLM apps](https://prompthelm.app/guides/prompt-management): What prompt management is, the signs your LLM app needs it, what a tool should do, and what to ask before you put one on your app's request path.
- [Prompt management for the Vercel AI SDK](https://prompthelm.app/guides/vercel-ai-sdk-prompt-management): The Vercel AI SDK has no built-in prompt management. Fetch prompts at runtime, pass them to generateText, and change them, or the model, without a redeploy.
- [A/B testing LLM prompts without false winners](https://prompthelm.app/guides/ab-testing-llm-prompts): Checking a prompt test as results arrive finds winners that are not there. How always-valid intervals, split checks and guardrails keep the verdict honest.

## Optional

- [Privacy policy](https://prompthelm.app/privacy)
- [Terms of service](https://prompthelm.app/terms)
