Prompt management, without the chaos.

Your prompts are scattered across code, notebooks, and someone’s head. LangWatch makes them one versioned source of truth: edit in a playground, ship through pull requests, A/B test in production, and tie every version back to its traces.

Used by thousands of AI developers shipping complex AI reliably.

System prompt

You are an accounting assistant. Categorize each expense into the right ledger account, return the GL code, and flag anything that needs manual review.

User

Categorize this expense:{{transaction}}

AWS invoice, $4,210.18, project Northwind

Cloud infrastructure · GL 6200 · confidence 0.98. No manual review needed.

The prompt problems teams bring us.

These come up in almost every conversation we have with teams putting AI into production. They are the reason prompt management exists.

• Prompts scattered everywhere
  Hardcoded in code, pasted in docs, tweaked in notebooks. No single source of truth, and no one is sure which version is actually live.

• No history, no rollback
  A prompt change breaks production and you cannot tell what changed, or get back to the version that worked.

• Changes ship unreviewed
  Anyone can edit a prompt and push it live. There is no review gate and no record of why it changed.

• Guessing which prompt is better
  You rewrite a prompt and hope it helps. There is no way to compare variants on real metrics like quality, cost, and latency.

• Only engineers can touch them
  Product and QA have the context to improve prompts, but every small change still needs a code deploy.

• No link to production behavior
  You cannot see how a specific prompt version actually performed once it shipped, so iteration stays guesswork.

Prompts as code, on the CLI.

Pull prompts into your repo as files, commit them next to your app, and sync changes with the platform. Lock versions for reproducible deploys and materialize remote prompts in CI.

• init, create, and add to manage prompt dependencies
• pull, push, and sync with the platform
• prompts.json and prompts-lock.json for reproducible installs

$ langwatch init
✓ created prompts.json

$ langwatch prompt add pizza-prompt
✓ added pizza-prompt@3

prompts/
├── .materialized/    # remote prompts fetched here
├── prompts.json      # prompt dependencies
└── prompts-lock.json # lock file

One source of truth for every prompt.

Manage, version, deploy, and measure your prompts in one place, with the workflow your engineers already use.

1. Versioned, with full history
   Every prompt is versioned with complete history and one-click rollback. Tag versions for production, staging, or any environment.

2. Prompts as code
   Manage prompts as .prompt.yaml files with the CLI: pull, push, and sync. Commit them next to your app and lock versions for reproducible deploys.

3. Reviewed through GitHub
   Gate prompt changes behind pull requests and let GitHub Actions sync on every push, so prompts get the same review as code.

4. A/B test in production
   Run variants live, randomized per request. LangWatch tracks quality, cost, and latency for each, so you pick the winner on data, not vibes.

5. Edit in the playground
   Iterate on prompts in an interactive playground with AI assistance, then deploy a new version without redeploying your app.

6. Linked to your traces
   Pull prompts at runtime via the API or SDK with dynamic variables, and tie every version back to its production traces.

And when manual iteration is not enough, optimize prompts automatically with DSPy, scored against the metrics you define.