Skip to main content

Build thunks from Claude Code or Codex

Install the Thunk.AI plugins so Claude Code, Codex, or the Claude and Codex apps can build, test and troubleshoot your thunks

You can build, change, test and troubleshoot thunks from an AI agent, Claude or Codex (in a terminal or in their desktop apps), as well as in the Thunk.AI app. Describe what you want ("build a thunk that triages support tickets", "why are this thunk's work items failing?") and the agent works on your thunks through your own Thunk.AI account.

The Claude and Codex buttons under New on your Thunk.AI home page show the setup steps for each agent, and examples of what to ask it.

This is the opposite direction from publishing your thunk as an MCP server: there, an AI agent calls the tools of a thunk you built; here, an AI agent builds the thunk.

What you install

Thunk.AI publishes two plugins:

  • thunk-prod connects the agent to the Thunk.AI FDA MCP, with tools to create and edit thunks, run and inspect work items, and read the Thunk.AI help center.

  • thunk-skills teaches the agent how to use those tools well, through its thunk-guide skill: which help articles to read for a task, the order to build a thunk in, and how to test and troubleshoot one. Its thunk-deployment-readiness skill checks whether a thunk is ready to go into production and fixes what you agree to.

The guidance in thunk-guide is also published in this help center, as Building thunks with the FDA MCP and Analyzing a thunk with the FDA MCP. The guidance in thunk-deployment-readiness is published as Checking deployment readiness with the FDA MCP. The FDA MCP points every agent at those guides when it starts a session, so an agent connected without the plugin gets the same guidance.

The agent acts as you. It sees the thunks you can see and can change only what your role allows. Your organization, a thunk's owner and you can narrow that further: see Control what Forward Deployed Agents can do.

Claude Code

Run these commands in a terminal:

claude plugin marketplace add ThunkAI/thunk-plugins
claude plugin install thunk-prod@thunk-ai

Installing thunk-prod installs thunk-skills too. If a Claude Code session is already open, run /reload-plugins in it.

Then sign in once. In Claude Code, run /mcp, select plugin:thunk-prod:thunk-fda, choose Authenticate, and sign in with your Thunk.AI account. The approval page shows which account and which Thunk.AI tenant (for example app.thunk.ai) the agent will use.

To pick up new versions, run /plugin marketplace update thunk-ai, or turn on auto-update for thunk-ai under Marketplaces in /plugin.

Claude app

In the Claude desktop app, click Customize, then + Add, choose Marketplace, and enter ThunkAI/thunk-plugins. Then install thunk-prod from that marketplace and sign in with your Thunk.AI account when its FDA server asks.

If you don't see Customize, use the Claude Code steps instead. Plugins you install there are available in the app too.

Build thunks from the app's Code tab. In the Cowork tab, Claude can't read the thunk-guide skill from these plugins yet, so it stops and asks you to switch to Code instead of building without the skill.

Your whole Claude organization

A Claude organization admin can make the plugins available to everyone in the organization, so nobody has to add the marketplace themselves. Open Organization settings, then Plugins & skills, then Marketplaces, choose Sync from GitHub, and enter ThunkAI/thunk-plugins. Each person still signs in with their own Thunk.AI account, and installs thunk-prod if it isn't installed for them already.

Codex

Run these commands in a terminal:

codex plugin marketplace add ThunkAI/thunk-plugins
codex plugin add thunk-skills@thunk-ai
codex plugin add thunk-prod@thunk-ai
codex mcp login thunk-fda-prod

Add both plugins: Codex does not install one plugin along with another. The last command signs you in with your Thunk.AI account. The approval page shows which account and which Thunk.AI tenant (for example app.thunk.ai) the agent will use.

The next time you start Codex, it says that hooks need review. Choose Review hooks and trust the two from thunk-skills, or Trust all and continue. They let Codex suggest a next step after it works on a thunk and welcome you the first time. Codex asks again when an update changes them. If you continue without trusting them, everything else still works.

To pick up new versions, run codex plugin marketplace upgrade thunk-ai, then run the two codex plugin add commands again.

If you added the guidance plugin before it was renamed, under the name thunk-builder, remove that copy with codex plugin remove thunk-builder@thunk-ai.

The server inside thunk-prod was called thunk-builder until November 2026 and is now thunk-fda. If you signed in before the rename, sign in once more after you update: in Claude Code, /mcp, then plugin:thunk-prod:thunk-fda, then Authenticate; in Codex, codex mcp login thunk-fda-prod.

Codex app

In the Codex app, open Settings, then Plugins. Open the menu next to the plugin search box, choose + Add More…, and enter https://github.com/ThunkAI/thunk-plugins.git. Then add both thunk-skills and thunk-prod, and sign in to thunk-prod with your Thunk.AI account.

If you don't see Plugins in Settings, or adding the marketplace fails, run the Codex commands in the Codex section instead.

If Codex asks you to review hooks, trust the ones from thunk-skills, as described in the Codex section.

Once the plugins are set up, the agent opens a thunk it creates in a panel beside the conversation, so you can watch it being built. It still makes every change through the FDA MCP tools, not by clicking in that panel, unless you ask it to.

In the Codex app, the plugin's suggested prompts offer first things to try, and after the agent works on a thunk, it ends its reply with what you could do next.

Other AI apps

Apps that don't install these plugins, such as ChatGPT, can still connect if they support custom MCP connectors. Add one with the URL https://postern.thunk.ai/api/builder/mcp and sign in with your Thunk.AI account. You get the FDA MCP tools without the thunk-guide guidance.

Tips

  • Start by naming the thunk you mean, or ask the agent to list your thunks.

  • When the agent builds a new thunk, it lays out all the steps first, then tests the thunk on sample work items as soon as it is built. While testing, it uses stand-in (mock) versions of outside systems such as a CRM or email; if your account already has a connection to one, it asks whether to use that instead.

  • A new workflow thunk comes with a webhook channel, Workflow requests, so other systems and people can send it work right away. The agent gives you its webhook URL and its form link. See Receive requests through a webhook or a shared form.

  • The agent keeps the thunks it creates for a project, including mock services and test copies, together in one folder on your home page, and tells you which folder. Name a folder if you want a particular one. See How to organize your thunks into folders.

  • After you change a thunk, ask the agent to test the change on sample work items before you use the thunk for real work.

  • For a thunk that talks to people, ask for a test work item with a simulated user playing the other side. See Test a conversation with a simulated user.

  • When you analyze a thunk that is already in use, ask the agent to work on a copy, so your live thunk is not changed.

  • Before you move a thunk to Production, ask "is this thunk ready for production?". The agent runs the design review, goes through the checklist in Move a thunk into production line by line, tells you which lines are met, which are not and which need your decision, and then makes the fixes you pick. It does not promote the thunk; you do that in Deploy → Phase.

Did this answer your question?