If you spend your day in Claude Code, Codex, or OpenCode, the worst part of validating an idea is the context switch. The Idea Launch MCP server removes it. One command connects your agent to your workspace, and from then on you set up a test, check what is still missing, and read the result by asking in plain language — in the same window you are building in.
- One command to connect. No API key, no config file to hand-edit.
- Your agent can create the idea, draft the brief, and write the ad creative.
- It can read your Ad Runs and market reads, but it can never spend your money.
In Claude Code, connecting takes one line:
claude mcp add --transport http idea-launch https://www.idea-launch.io/api/mcpThe first call opens a browser sign-in. Your client registers itself through OAuth and stores the token — there is no key to paste or rotate. Codex and OpenCode take a few lines of config instead; both are on the connect page.
What it does
Eighteen tools, split into reading and preparing
Reading tools answer questions. Preparing tools move a test forward.
The reading half is what most founders use most days. It covers your ideas and their setup progress, every Ad Run, the market read behind a run, the run timeline that explains why something is not live yet, and the signups a run produced.
The preparing half is the setup work: create an idea, draft the brief fields, save them, generate the tagline, ad copy and hero image, and stage the run to see whether it would pass every launch gate. Staging is a dry run — it creates nothing and charges nothing.
Read
Ideas, Ad Runs, market reads, run timelines, signups, and the page your traffic lands on.
Prepare
Create an idea, draft and save the brief, generate ad creative, stage the run.
Interpret
Two packaged prompts — one takes an idea to launch-ready, one reads your results back as a decision.
No agent can spend your money.
Starting or boosting an Ad Run debits a credit, and those actions are deliberately absent from the tool surface. Your agent can get a test to launch-ready and tell you it is ready. You confirm the run yourself, in the app. That is a design decision, not a limitation we plan to remove.
Where each step happens
In your agent, or in the app
The split follows one rule: anything that spends a credit is confirmed by a person.
| Step | In your agent | In the app |
|---|---|---|
| Create an idea | Yes | Yes |
| Write the brief | Drafts it and saves it | Edit any field |
| Ad copy and hero image | Generates both | Review and swap variants |
| Check launch gates | Dry run, creates nothing | Same checklist, on the idea page |
| Start the Ad Run | Never | You confirm it |
| Read the result | Score, signals, next move | Full readout and signups |
How to use it
Ask in plain language, not in tool names
The agent picks the tools. You describe the situation.
You do not need to learn the tool names. The two packaged prompts — one for taking an idea to launch-ready, one for reading results — handle the orchestration, and a plain request usually routes to the right one on its own.
If you have already shipped something, say so and give the URL. That routes the test down the path where your existing page is the destination, instead of generating a new one you do not need.
- “I just built a scheduling tool for barbers. Set up an Idea Launch test for it.”
- “Is my idea ready to test yet, or is something still missing?”
- “How did my Idea Launch test go, and what should I do next?”
- “Why isn’t my ad run live yet?”
Reading the answer
What your agent will and will not tell you
The tools hand back a decision, not a dashboard.
A market read comes back as a validation score with a band and a confidence level, the signals underneath it, and a recommended next move. A read marked provisional is an early number on thin traffic — treat it as a direction, not a verdict, and let the run finish.
One thing is deliberately missing: cost. Ad spend, and any number spend could be derived from, never leaves the server. What you get back is click-through rate, landing page conversion, signup volume and the validation score — the numbers that answer whether the positioning is working.
- A score and band, not a raw metric dump.
- Confidence, so a thin read is labelled as one.
- A recommended next move you can act on today.
- No spend figures, ever — by design, not by omission.
Installing it as a plugin
Or skip the config entirely
The Agent Plugin connects the server and adds the judgement in one step.
Idea Launch also ships as an Agent Plugin — a vendor-neutral package that any conformant client can install. It wires up the MCP server and adds two skills that carry the parts a tool list cannot: how to read a market read honestly, and what separates a targetable brief field from a vague one.
In Claude Code it is two commands. In VS Code, Cursor, Copilot, Codex and other plugin-aware clients, it is whatever that client’s install flow looks like — the connect page has the details.
Founder questions
Questions you might still have
Do I need an API key?
No. The first call opens a browser sign-in, and your client registers itself through OAuth and stores the token. There is nothing to paste, store in an env file, or rotate.
Can my agent accidentally spend money?
No. Starting and boosting an Ad Run are not exposed as tools at all, so there is no call an agent could make that debits a credit. Those actions are confirmed by you in the web app.
Which clients work?
Any client that speaks MCP over streamable HTTP. Claude Code, Codex and OpenCode are documented on the connect page. Plugin-aware clients like VS Code, Cursor and GitHub Copilot can install the Agent Plugin instead, which connects the server and adds the skills in one step.
Why can't I see my ad spend?
Because spend is not the number that answers your question, and mixing it in tends to turn a validation decision into a media-buying one. The read you get back is built around click-through rate, landing page conversion, signups and the validation score. That constraint is enforced on the server, not left to the agent.
Does it work for a product I've already launched?
Yes. Give it the URL when you create the idea. The test then points traffic at the page you already have, which is the faster path when the thing exists and the open question is whether the current angle earns attention from strangers.
Can it start a test on a schedule for me?
No. Every run is confirmed by a person. If you want a repeating test cadence, the honest version is that you decide each time — which is usually the right call anyway, since each run should be asking a different question.