LaunchRanked

Launch guideAI products

How to launch an AI agent

An agent takes actions, so buyers ask two questions a chatbot never faces: does it actually finish the task, and what can it touch? Launch with recorded end-to-end runs, a plain list of permissions, and a clear way to stop or undo. If your agent exposes tools over MCP or runs inside ChatGPT, the directories for those ecosystems are launch channels in their own right.

Updated · platform facts checked against the official docs linked below

At a glance

Proof buyers want
Full task runs, including failures
Ecosystem directories
MCP Registry, ChatGPT app directory[1]
Trust must-haves
Permissions, audit log, human approval for risky steps

Where people find AI agents

  • MCP Registry: If your agent ships an MCP server, publishing there lets MCP clients discover it. It launched in preview.
  • ChatGPT app directory: If you build with the Apps SDK, the directory is where ChatGPT users find apps.
  • AI tool directories: People browse by task; list under the job the agent completes.
  • Hacker News (Show HN): Developers will test the edge cases. Only post if they can try it.
  • LaunchRanked: A permanent page to link from every directory.

Assets to prepare

Assets to prepare before launch
AssetSpec or guidance
End-to-end recordingsThree real tasks from start to finish, uncut, including one where the agent asks for help.
Permissions pageEach system it can read or change, why, and how to revoke access.
Audit log screenshotShows buyers they can see what the agent did.
server.json (if MCP)The metadata file the MCP Registry's publisher CLI submits.[1]
Demo accountFull-featured test credentials; ChatGPT app review rejects submissions reviewers can't log into.[2]

Store and listing rules

What the platform itself says, with a link to the page we checked.

  • The MCP Registry verifies namespace ownership through GitHub login, GitHub Actions OIDC, or DNS/HTTP proof for your domain. It launched as a preview and warns of possible breaking changes.[1]
  • ChatGPT app submissions must come from verified individuals or organizations, need a published privacy policy and working demo credentials, and trial or demo apps are rejected.[2]
  • ChatGPT app tools must be annotated correctly (read-only vs destructive vs open-world actions) or face rejection.[2]

A 4-week launch timeline

  1. Week 1 · four weeks out

    • Pick three tasks the agent reliably completes and measure how often it does.
    • Write the permissions page.
    • Add an approval step before anything destructive.
  2. Week 2 · three weeks out

    • Record end-to-end runs, including a failure and how it recovers.
    • Complete identity verification if you'll submit to the ChatGPT directory.
    • Prepare server.json and publish to the MCP Registry if you ship an MCP server.
  3. Week 3 · two weeks out

    • Write the launch thread around one recorded task.
    • List on AI directories under the job it completes.
    • Schedule LaunchRanked and Product Hunt.
  4. Week 4 · launch week

    • Run a closed beta with ten users and watch where the agent stalls.
    • Write the FAQ from their questions.
    • Set cost and rate limits for launch day.

Launch-day checklist

  1. Lead every post with a recorded task, not a feature list.
  2. Show the permissions page link in the first comment.
  3. Watch runs for loops or runaway costs; cap them if needed.
  4. Reply to security questions quickly and specifically.
  5. Collect failed runs as your next week's roadmap.

After launch: SEO steps

Launch boards reset; search pages don't. These steps keep people finding you after launch week.

  1. Write one page per task the agent automates, with a real run and the time it took.
  2. Publish an integration page for each tool it connects to.
  3. Write a clear, quotable explanation of how it works and what it can't do.
  4. Check AI crawler access so assistants can read those pages.

Common mistakes

  • Showing only a polished happy path; buyers assume agents fail and want to see how.
  • Asking for broad permissions it doesn't need.
  • No kill switch or undo.
  • Submitting to app directories with a demo account reviewers can't use.

Templates for this launch

FAQ

How do I publish an AI agent to the MCP Registry?

Describe your server in server.json, then use the registry's mcp-publisher CLI and verify your namespace with GitHub or your domain. The registry launched in preview, so check its README for current steps.

What does ChatGPT's app directory require?

Identity verification, a published privacy policy, working demo credentials, correct tool annotations and a complete app, not a trial. Read OpenAI's submission guidelines before you build the listing.

How do I get buyers to trust an agent?

Show full recorded runs, list every permission and why, provide an audit log and require approval before destructive steps.

Should an agent launch on Product Hunt?

It can, with a recorded task as the first gallery item. Pair it with ecosystem directories where users of that platform look.

Sources

  1. MCP Registry (GitHub: modelcontextprotocol/registry)
  2. OpenAI Apps SDK: App submission guidelines

Checked . Platforms change their rules; open the source before you rely on a number.

← All 35 launch guides