ChatGPT apps and MCP servers
ChatGPT apps and MCP servers, built and taken through review
This is for companies that want their product to work inside ChatGPT, Claude and other AI assistants: a SaaS people ask questions about, a booking or quoting flow, a catalog or a knowledge base.
We've done it for ourselves. Pulore App Planner is our ChatGPT app. It's an MCP server with five tools that plan software projects, check website speed and send us a brief only after the user approves it. It passed OpenAI's review and is listed in ChatGPT's plugin directory.
Fixed price: a ChatGPT app in two weeks
For an app with up to five tools over an API you already have, we design it, build it and submit it to OpenAI in two weeks. It's a fixed price, from $5,000, and you get the exact figure in writing before we start. An app that signs users in and acts on their behalf, with up to six tools, starts at $8,000 and takes three weeks.
Tool design
What each tool does, when ChatGPT should call it, and how it's annotated.
The MCP server
Typed, tested and rate-limited, in your repository and deployed on your infrastructure, or ours if you prefer.
The review package
Test cases, a demo recording, listing copy and privacy policy updates, submitted for you, with the reviewer's feedback handled.
Claude too
The same server set up as a Claude connector and documented for other MCP clients.
Approval and placement in the directory are OpenAI's decisions, so neither is guaranteed. Hosting costs are separate, and extra tools or later changes are billed hourly.
What a ChatGPT app is
A ChatGPT app, which OpenAI's directory now calls a plugin, is an MCP server: a set of tools with names, descriptions and input schemas that ChatGPT can call in the middle of a conversation. It can read live data from your product and take actions the user asks for.
The same server can be added to Claude as a custom connector and to other MCP clients, so one build reaches several assistants.
What OpenAI's review looks at
We took our own app through review, so we build to these from the first day.
Useful on its own
OpenAI rejects apps that exist mainly to advertise. Our contact and booking tools only run when the user asks, and plans carry no sales pitch.
Honest tool annotations
Every tool declares whether it only reads, whether it changes or sends anything, and whether it reaches outside services. We write a justification for each one.
Descriptions that say when
Each tool description says when to use it and when not to, and we test them against direct, indirect and unrelated prompts so the app isn't called by mistake.
Approval before anything is sent
A tool that sends data shows the user exactly what will go out and waits for a yes. Nothing from the rest of the conversation is sent.
Sensitive data turned away
Inputs are checked for card numbers, keys and tokens and refused. Logs keep error codes, not what people typed.
The review package
Listing copy, privacy policy and terms, positive and negative test cases, a recorded walkthrough and domain verification.
How a project runs
Pick the tools
We list what your users would ask an assistant and choose the three to six tools that answer it.
Build the server
A TypeScript MCP server with typed inputs, rate limits and, if your data is private, OAuth sign-in so each user sees only their own.
Test it in ChatGPT
We run every test prompt in developer mode on desktop and mobile, and fix descriptions that trigger at the wrong time.
Submit and launch
We prepare the review package, answer the reviewer's feedback, then publish, link it from your site and set it up for Claude.
What you get
Tool design
A short spec for each tool: when ChatGPT should call it, its inputs and outputs, and its annotations.
MCP server in your repository
Typed, tested and deployed on your infrastructure, or ours if you prefer.
Review package
Listing copy, test cases, a demo recording, privacy and terms updates and domain verification.
Claude and other clients
The same server set up as a Claude connector and documented for other MCP clients.
Usage you can measure
Tagged links back to your site and counts of tool calls, without logging what people type.
What we've done for clients
Emblem
AI due diligence that cites its sources
Emblem's AI agents run on LangGraph, with LangSmith tracing, test datasets and regression runs. We bring the same habits to MCP tools: clear schemas, tests for when a tool should and shouldn't run, and answers that cite their source.
Read the Emblem case studyHow a project starts
From a free call to the end of the second week.
A free 30-minute call
Tell us what your users would ask an assistant about your product. We'll say whether a ChatGPT app is worth it, or whether a feature on your own site would serve them better.
A fixed price, or hourly
After the call we send a fixed-price quote for a clear scope, or agree on hourly work at $100 to $150 an hour. An app with up to five tools is a fixed price from $5,000, and one with sign-in and actions from $8,000. Approval is OpenAI's decision, so nobody can promise it, but we build to its guidelines from the start.
The first two weeks
- We agree on the tools, their inputs and outputs, and how each one is annotated.
- We build the MCP server and test it in ChatGPT's developer mode on desktop and mobile.
- We write the test cases and listing copy reviewers ask for.
Frequently asked questions
What's the difference between a ChatGPT app and a custom GPT?
A custom GPT is instructions and files inside ChatGPT. An app connects ChatGPT to your own server through MCP, so it can use live data and take actions, and it's listed in ChatGPT's plugin directory after OpenAI reviews it.
Will OpenAI approve our app?
We can't promise it, because OpenAI decides. Apps that mainly advertise get rejected, so we design tools that are useful on their own, prepare the review package and handle the feedback.
Does it work in Claude too?
Yes. The server speaks MCP, so it can be added to Claude as a custom connector and to other MCP clients. Listing it in Claude's directory is a separate submission, which we can do too.
Do users need an account?
Only if your tools need one. Ours needs none. If your app reads private data, we add OAuth sign-in so each user only reaches their own account.
How much does it cost to build a ChatGPT app?
From $5,000 for an app with up to five tools over an API you already have, built and submitted in two weeks. With sign-in and actions on the user's behalf, it starts at $8,000 and takes three weeks. Both are fixed prices agreed before we start.
What does it cost to run?
The server is a small web service, so hosting is usually modest. Costs grow with what the tools call, such as your own APIs or a third-party service. We estimate it from your expected usage before we build.