PromptOT MCP: Manage and version LLM prompts from your AI tools

PromptOT has released an MCP server that allows developers to manage and version LLM prompts directly from AI tools like Claude Desktop, Cursor, and Codex. The server exposes 23 tools for creating, editing, versioning, and testing prompts, enabling prompt operations as part of the development workflow without redeploying applications.

Prompts often start as simple strings in code. Then the product grows. You add a better system prompt. Then a guardrail. Then a different version for production. Then a customer-specific variable. Then test cases. Then someone wants to roll back to the version from last week. At that point, prompts are no longer just text. They become part of the application layer. That is the problem PromptOT is built around. PromptOT is a prompt management platform for LLM applications. It lets teams build, version, evaluate, and deliver system prompts through an API without redeploying the app every time a prompt changes. The new part I want to share is the PromptOT MCP server . It lets you manage your PromptOT prompts directly from MCP-compatible AI tools like Claude Desktop, Cursor, Codex, ChatGPT, claude.ai, and other clients. MCP stands for Model Context Protocol . It gives AI tools a standard way to connect with external systems. Instead of only chatting with an assistant, you can give the assistant controlled tools for reading and changing real application data. For PromptOT, that means your AI assistant can help with things like: The goal is not just to generate prompt text. The goal is to manage prompt operations as part of the development workflow. PromptOT lets you structure prompts as versioned assets instead of hardcoded strings. A prompt can be composed from typed blocks, such as: It also supports variables like: {{customer name}} {{tone}} {{product context}} Those variables can be resolved when your application fetches the prompt. So instead of redeploying your app every time you adjust a prompt, your app can fetch the latest published prompt through an API. The PromptOT MCP server exposes 23 tools across five areas: | Area | What you can do | |---|---| | Prompts | List, get, compile, create, update, and delete prompts | | Blocks | Create, edit, delete, and reorder typed prompt blocks | | Variables | Manage {{key}} placeholders | | Versions | Save drafts, publish, diff, and roll back | | Test cases | Create and manage prompt evaluation inputs | That makes it possible to work on prompts from inside the AI tools developers already use. For example, you can ask your assistant to: Show me the current support-agent prompt. Then: Add a stricter refund-policy guardrail as a new block, but save it as a draft. Then: Compare this draft with the currently published version. Then: Publish the draft if the changes look safe. The assistant is not just writing text in a chat window. It is using scoped PromptOT tools to work with real prompt objects. The MCP package is available as: npx @prompt-ot/mcp You will need a scoped PromptOT API key from your PromptOT dashboard. Add this to your Claude Desktop MCP config: { "mcpServers": { "promptot": { "command": "npx", "args": "-y", "@prompt-ot/mcp" , "env": { "PROMPTOT API KEY": "pot ...", "PROMPTOT MCP CLIENT": "claude-desktop" } } } } PromptOT also provides a desktop extension bundle for Claude Desktop: https://www.promptot.com/downloads/promptot.mcpb Add this to your Cursor MCP config: { "mcpServers": { "promptot": { "command": "npx", "args": "-y", "@prompt-ot/mcp" , "env": { "PROMPTOT API KEY": "pot ...", "PROMPTOT MCP CLIENT": "cursor" } } } } Add this to your Codex config: mcp servers.promptot command = "npx" args = "-y", "@prompt-ot/mcp" env = { PROMPTOT API KEY = "pot ...", PROMPTOT MCP CLIENT = "codex-cli" } PromptOT also has a hosted MCP endpoint: https://mcp.promptot.com/mcp This can be used by clients that support hosted MCP connections with OAuth. Here is a realistic workflow. You are building an AI support agent. The current system prompt is hardcoded in your app. Every time you want to update the refund policy, escalation rules, or tone, you need to change code, open a pull request, wait for CI, and deploy. With PromptOT, you can move that system prompt into a managed prompt. You split it into blocks: Role Context Instructions Refund policy Escalation rules Output format Guardrails Your app fetches the published version through the PromptOT API. Then, with the MCP server connected, you can ask your AI tool to help manage changes: Open the support-agent prompt and create a draft version with a stricter escalation rule for billing-related complaints. Then: Show me the diff between the published version and this draft. Then: Create a test case for an angry customer asking for a refund outside the refund window. Then: Publish the draft. This keeps prompt work closer to the tools where developers and AI engineers already think through changes. Prompt management gets messy because prompts sit between product, engineering, support, and evaluation. They are not exactly code, but they affect production behavior like code. A few things become important quickly: PromptOT is designed around those questions. The MCP server adds another layer: instead of only clicking through a dashboard, you can let your AI development tools interact with the prompt system directly. PromptOT uses scoped API keys for MCP access. That matters because an MCP server can expose powerful actions. PromptOT separates tool capabilities so clients can understand whether a tool is read-only, destructive, or idempotent. The goal is to make prompt operations useful without turning every assistant action into an unsafe production change. LLM apps are moving from simple prompt strings to real prompt infrastructure. Once prompts need versions, variables, publishing, rollback, evaluations, and team workflows, they deserve their own operational layer. PromptOT MCP is one step toward making prompt management feel native inside the AI tools developers already use.