Skip to content

Prompt management for LLM apps: Change a prompt. Every app follows.

Publish a change in the dashboard and every running copy of your app picks it up on its next refresh. No rebuild, no redeploy.

  • Free plan
  • No card required
  • TypeScript SDK
  • Any LLM provider

Support bot/support-reply

Published
Live · v3 — Your app receives exactly this text.
285 characters
  • api-1
    v3us-east-1
  • worker-1
    v3eu-west-1
  • cron-1
    v3us-west-2
  • api-2
    v3us-east-1
  • api-3
    v3eu-west-1
  • worker-2
    v3ap-south-1
Every instance on v36 of 6

Why PromptHelm

A prompt change shouldn't be a release.

Today one new word means a pull request, a review, CI and a deploy. With PromptHelm, it's a publish.

Without PromptHelm

A prompt change is a code change.

  1. Edit the string in your code
  2. Open a pull request
  3. Wait for review
  4. Wait for CI
  5. Deploy every service that uses it
  6. Spot a typo. Start again.

With PromptHelm

A prompt change is a publish.

  1. Edit the prompt in the dashboard
  2. Publish
  3. Live in every app

Live in 6 of 6 instances

Meanwhile, the release on the left is at step 1 of 6: edit the string in your code.

How it works

Three steps to a live prompt.

Move a prompt out of your code once. After that, every change is an edit in the dashboard.

Support bot · New prompt

Slug

support-reply

Your code asks for the prompt by this name.

Prompt

You are the support assistant for {{company}}. Answer {{customer_name}}'s question in two or three sentences.

companycustomer_name

Or hand it to your coding agent

Paste this into Claude Code, Cursor or any agent that edits your repository.

Move this app's hard-coded LLM prompts to PromptHelm…

A/B tests

A/B test a prompt on real calls.

Give a new version a share of your calls and see whether it does better, with tokens and latency side by side. Then ship the winner in one click.

Support bot/support-reply

Day 1 of 7

Your app

0

calls

  • v3

    Live80%

  • v4

    Test20%

20% get v4, 80% keep v3

A new share starts the count over.

A/B test

v4 on 20% of calls · stops if v4 is clearly worse

Not enough outcomes yet

Getting to an answer0%
v3Live80%

—

0 outcomes

Tokens per call—

v4Test20%

—

0 outcomes

Tokens per call—

  • Split on your server

    The SDK picks the version for each call, keeping each user on one if you pass their id. We get totals only, never your text or your users.

  • Look as often as you like

    The range stays honest however often you check, so an early peek can't talk you into shipping.

  • Stops if it's worse

    Turn the guard on, and a new version that is clearly worse sends every call back to the live one.

A/B tests come with Pro, $19 a month. See plans

Built in

Everything a prompt needs.

Versions, review, fallbacks and privacy, there from the first publish.

  • Variables, filled in on your server

    Mark what changes with double braces. The SDK fills the values in where your app runs, so they never reach us.

    What the model receives

    You are the support assistant for Acme Cloud. Answer Maya's question in two or three sentences.

  • 60 s

    From publish to production

    Every running copy of your app picks up a change on its next refresh. Pass ttl for a different pace.

  • Review before it ships

    Publish opens the exact change from live, word by word, and Undo stays a click away.

    Publish v4 of support-reply?

    Changes from live v3

    Answer in two or three sentences. Be concise warm and concise.

  • One-click rollback

    Every publish is a numbered version. Make any of them live again, and your app follows on its next refresh.

    support-replyHistory
    v3 compared with live v4
    You are the support assistant for {{company}}. Answer {{customer_name}}'s question in two or three sentences. Be concise. If you're not sure, say so and offer to connect them with a person.
  • Falls back, never falls over

    A timeout after 3 seconds, an outage or a missing variable returns your fallback and calls onError.

    [PromptHelm] support-reply:
    No response within 3000 ms
    → served your fallback
  • Never sees your data

    Your app calls your LLM provider directly. We get a slug and an API key, nothing else.

    Your servertoPromptHelm

    a slug and your API key

    Your servertoYour LLM

    your keys, user data, traffic

  • 0

    Dependencies

    The SDK is MIT licensed, ships as ESM and CommonJS and runs on Node.js 18 and up. Any other language fetches over HTTP.

FAQ

Questions developers ask.

Anything else is in the quickstart and the guides.

What happens if PromptHelm is down?

Your app keeps answering. After the first fetch the SDK serves prompts from memory and keeps the last good version when a refresh fails; before that, a call with a fallback returns your fallback. Either way, onError tells you.

Do you see my users' data or my LLM traffic?

No. Your app calls your LLM provider directly, as it does today. The SDK fetches the prompt text with your API key and fills in the variables on your server, so their values never reach us.

How quickly does a change reach my app?

On its next refresh: 60 seconds by default, or whatever you set as ttl. A rollback reaches it the same way.

Can I A/B test a prompt?

Yes, on Pro. Give a new version a share of the calls, say 20%, while the live version keeps the rest. The SDK picks the version for each call on your server and keeps each user on one when you pass a unit. Your code reports how each call went and its token usage, and the prompt's page tells you whether the new version did better, at what cost in tokens and latency, with a range that stays honest however often you look. PromptHelm gets totals only, never the text. Then you ship the new one or stop. See how.

What if PromptHelm shuts down?

Your prompts aren't locked in. The pull command writes every published prompt into a file in your repo that your app also uses as its fallback, so it keeps answering without us. Settings exports a whole project, drafts and every version, as JSON.

Which languages can I use?

There's a TypeScript SDK for Node.js 18 and later. Any other language can fetch a prompt with one HTTP request; the quickstart shows how.

Does it work with OpenAI, Anthropic and the others?

Yes. PromptHelm supplies the text and, if you set them, the model, temperature and output limit to call it with; your code passes them to whichever provider and client library you already use. Model settings.

What does it cost?

The Free plan covers 10 prompts and 2 projects, and signing up needs no card. Pro is $19 a month or $190 a year for 200 prompts and A/B tests. Requests are never billed; see pricing.

Stop redeploying for a sentence.

Free for 10 prompts. Sign up with your email, add a prompt and fetch it from your app.