NewMCP now available for Claude, Cursor, Codex & Zed→

Version, Eval & DeployPrompts done right.

One platform for every production prompt, delivered via a single API call.

Get started free See how it works

No credit cardWorks with your stack

v004 · published

Synced via API & MCP

Compiled · 6 blocks

editor-assistant.prompt

v004

Role · system

You are an assistant to the editor of a literary quarterly.

Context · user scope

Reader submissions arrive in batches of {{batch_size}}. Prior editorial standards apply.

Instructions

Classify, summarize in two sentences, and suggest a disposition for each submission.

Guardrails

Never reject on tone alone. Flag for a human when the writer is under {{min_age}}.

Output_format · JSON

{ disposition, confidence, rationale }

The shift

LLM prompts are production code. Manage them like it.

The problem

Your prompts live in a Google Doc. And a Slack thread. And someone's Notion.

Every model call in production depends on prose, and nobody knows which version is the good one. When something breaks, you roll back code but the prompt is gone.

No version history, no diffs, no one to blame (or thank)
No way to evaluate a rewrite before shipping it
Copy-pasted across clients, SDKs, and scripts
Legal and brand review happens (if at all) in DMs

The current state

"final_prompt_v2_REAL.txt"Google Drive

"Hey can someone send me the latest…"Slack · #eng

Prompt draft, April 11Notion

const SYSTEM = `You are a helpful…`src/api.ts

Hard-coded in prodVercel env

The solution

One source of truth, shipped by API or MCP.

PromptOT is the CMS prompts deserved. Author in typed blocks, version with real history, evaluate against saved test cases, and deliver to any model without changing your app.

Six typed blocks compile into one deterministic prompt
Semantic versioning with diffs, rollback, and approval workflows
Evaluate every version against your test suite, not guesses
Environment-scoped keys: production, staging, or latest draft

With PromptOT

editor-assistant.promptv004

support-triage.promptv012

onboarding-coach.promptv007

search-reranker.promptv003

legal-summary.promptv002

Features

The prompt management platform, end to end.

Everything teams need to author, version, evaluate, and deliver LLM prompts in production. All in one workflow.

Scroll to exploreSwipe to explore→

Feature 01

Typed blocks.

Six first-class block types: role, context, instructions, guardrails, output_format, and custom. Each is typed, validated, and versioned independently.

Compose once, reuse across prompts
Variable interpolation with {{key}} syntax
Compile-time schema validation

Feature 02

Semantic versioning.

Every change is a versioned artifact with an author, a reason, and a diff. Roll back in one click, or publish through review.

Full history with diffs and commit messages
Published vs. draft environments
One-click rollback, no redeploy

Semantic prompt versioning with diff view and rollback

Feature 03

Run Evaluations.

Run any version against your saved test cases across the models you use in production. Compare pass/fail, latency, and token cost side by side before you ship.

Reusable test cases with expected outputs
Multi-model comparison in a single run
Rule-based criteria (contains, regex, similarity)

Run evaluations against saved test cases across multiple models

Feature 04

API delivery.

Fetch the compiled prompt from a simple REST endpoint. Environment-scoped keys decide whether you get the published version or the latest draft.

Production key → published version
Development key → latest draft
HMAC-signed webhooks for changes

Prompt delivery API with environment-scoped keys

Feature 05

MCP native.

Your prompts are also an MCP server. Use them directly in Claude Desktop, Cursor, Codex CLI, Windsurf, Zed, or ChatGPT. No SDK, no code change.

One-line config to connect clients
Environment-aware: dev clients see drafts
Secure by default with signed requests

Claude DesktopLIVE

CursorLIVE

ChatGPTLIVE

Codex CLILIVE

WindsurfLIVE

ZedLIVE

Feature 06

AI co-pilot.

An editor that reads with you. Ask it to tighten a guardrail, explain why v003 underperformed, or draft the next iteration. Pick your model.

GPT-4o · Claude Sonnet 4 · Claude Haiku 4 · Gemini 2.5
Suggestions as tracked diffs you can accept or reject
Per-prompt model routing for cost control

AI co-pilot suggesting prompt improvements

How it works

From prompt draft to production API in three steps.

You don't need a rewrite. Connect PromptOT to your existing app with a single API call and migrate one prompt at a time.

Step 1

Write in blocks

Compose each prompt as six typed blocks. Add variables, examples, and guardrails. The editor validates everything as you type.

Step 2

Test with evals

Run drafts against your saved test cases across the models you use in production. When v005 passes, promote it. When it doesn't, see exactly which test cases failed and why.

III

Step 3

Ship by API or MCP

Swap your hard-coded string for a fetch call. Or connect via MCP and let Claude, Cursor, or Codex pull the latest themselves.

Use cases

One prompt platform, every AI workflow.

From two-person startups to platforms with millions of calls, PromptOT is the prompt layer between your team and every model.

Startups

Solo & small teams

Enterprise

Large scale teams

Agencies

Per-client voice

Platforms

Multi-tenant

PromptOT

The prompt layer

Ship faster

Iterate without redeploy

Per-client libraries

Brand voice, versioned

Branded editor

Namespaced & scoped

Pinned reproducibility

One endpoint replay

Blog

Latest from the blog.

Tips, tutorials, and insights on prompt engineering and LLM management.

View all articles →

Pricing

Simple pricing. Free to start.

Every plan includes the full editor, six block types, unlimited versions, and API + MCP delivery. Scale when you're ready.

Free

For individuals and side projects

$0forever

Start free

Up to 3 projects
5 prompts per project
1,000 API calls / month
50 AI credits / month
Up to 5 team members
Community support

Prompt management FAQ.

Still stuck? Write to us at hello@promptot.com.

No. Swap one hard-coded prompt string for a fetch() to our API and you're done. You can migrate prompts one at a time with no big-bang cutover. For MCP clients like Claude Desktop or Cursor, it's a single config line.

All of them. PromptOT delivers a compiled string that you send to whatever model you're already using (OpenAI, Anthropic, Google, open-source, local). For the co-pilot that helps you edit prompts, we use GPT-4o, Claude Sonnet 4, Claude Haiku 4, and Gemini 2.5 Flash.

Save reusable test cases for each prompt: input text, optional expected output, and optional criteria (contains, regex, similarity). Trigger a run on any version against any combination of models. PromptOT records pass/fail, latency, and tokens for each result so you can compare versions or models side by side. Live-traffic A/B routing is on the roadmap.

Yes, the Free plan is free forever. Up to 3 projects with 5 prompts each, 1,000 API calls and 50 AI credits per month, unlimited versions, and full API + MCP delivery. No credit card required to start.

Every payload is encrypted in transit (TLS) and at rest via Supabase. Webhooks are HMAC-signed (X-PromptOT-Signature). SOC 2 Type II, SSO/SCIM, and EU data residency are on the roadmap for our enterprise tier. Contact us for the current status if these are blockers.

Self-hosting isn't available today. PromptOT is a managed service hosted on Railway with Supabase. If your team needs in-VPC deployment, get in touch. We're tracking interest for an enterprise self-host option.

Ship LLM prompts the right way. Start free today.

Three minutes to your first compiled prompt. Five to your first API call. No credit card required.

Start for free →