This week AI coding agents such as Claude Code and Codex can build, test and connect thunks with less help from you, list properties get their own formats and views, and the run explanations and Rewrite became conversations you can continue.
New
Build thunks from Claude Code or Codex. Install the Thunk.AI plugins in Claude Code or Codex with two commands, sign in with your Thunk.AI account, and ask the agent to build, test or troubleshoot your thunks. The agent creates a thunk's steps and workflow state in one go, tests the new thunk right away with test work items, and stands in mock versions of outside systems while it does. If your account already has a connection to one of those systems, it asks you whether to use that instead. In the Codex app, the thunk opens in a panel beside the conversation. The plugins' building guidance is also published as two guides for AI agents, and every agent connected to the Builder is pointed at them, with or without the plugins. See Build thunks from Claude Code or Codex and Analyzing a thunk with the Builder MCP.
AI coding agents can set up connections, files and requests. Through the Builder API (including from an AI coding agent over MCP), an agent can now:
create an MCP server connection, and an OAuth connection that needs no sign-in (client credentials). The agent never handles the secret: the connection is created without it, and you enter it on your Account page.
refresh an MCP connection's tools when the server changes them, and see which tools were added, removed or changed. Every thunk that uses the connection picks up the change. It only happens when you ask.
send a file from its own machine, such as a sample invoice, straight to a thunk as a work item's file or image or into a content folder, through an upload link that lasts 30 minutes and takes up to 5 files.
start each new workflow thunk with a Workflow requests webhook channel that only adds new work items, so other systems and people with the form link can send requests right away.
Format a list property as a table or a list, with your own views. A property that holds a list of values now has a Format. Choose whether it opens as a table (one row per item) or as a list, and add named views, such as Open line items, that each keep the items matching a condition, show the fields you choose, and use their own layout. Viewers switch views with a row of buttons above the list, and each viewer's choice is remembered. In grids and reports, a list cell counts the items in the chosen view. Stored values and what the AI agent sees do not change. See Format a list property.
Tools tell outside AI agents how they behave. When your thunk is published as an MCP server, each tool now tells the connecting agent, such as ChatGPT or Claude, whether it only reads, whether it can overwrite or delete data, whether it is safe to repeat, and whether it reaches outside systems or people. Agents use this to decide which tools they can call without asking first. Built-in tools set these hints themselves. For a custom tool, set them under Behavior hints for AI agents, or let the tool builder set them. They do not change approval. See Publish your thunk as an MCP server for AI agents.
Richer chat conversations. A new Send structured data setting on a conversation property lets a custom tool attach structured data, such as an appointment card or a set of choices, to the messages it sends. It is off by default. In the hosted chat app, Image and File fields on the start form and on forms the agent sends now let the end user pick a file from their computer, with a preview for images. Before, an Image field showed two text boxes, src and anchor. See Conversation properties for email and Set up the hosted Chat UI for your thunk.
Improvements
Ask follow-up questions about a run, and rewrite instructions in a conversation. On a failed run, What went wrong? is now a labelled button beside the error. It and Why so long? open a conversation where you can ask follow-up questions after the explanation. A Rewrite button on a step's directions and on a custom AI tool's instructions opens a conversation that rewrites them to follow How to write effective directions. Each change has a side-by-side comparison and a Revert button. These are available to the thunk's owner and admins. See Analysis of AI behavior and How to Improve AI Instructions with Rewrite.
More room to work, and thunks open where their work is. The Planning tab and a custom-built tool's page now have an AI button at the right end of their tabs (its tooltip reads Hide AI). It hides the AI chat on both pages, and Show AI brings it back. While the planning agent is updating the plan, its chat stays open. You can also drag the left edge of the details panel beside a grid or saved report to resize it. A link to a built thunk now opens Work Items if it has any, the Tests page for thunk admins if it only has test work items, and otherwise the Planning tab with the AI chat closed.
Monitor tabs have shorter names, and Pause / Restart is now in Deploy. The Monitor tabs are now Errors, Cost, Time, Reports, Tracking and Log. Pause / Restart moved to the Deploy pane, where it is the first tab. Old links to it still work. See The Monitor pane: what each tab is for and Pause and restart a thunk.
Safer, simpler MCP sign-in. Quick Connect now signs in to MCP servers that look up a published description of the application instead of letting it register itself, so you no longer set up an OAuth credential by hand for them. Existing connections keep working. When Claude, Claude Code or another AI app connects to your thunk's MCP server, the sign-in page now shows the app's verified web address and where access will be sent. Apps that cannot be verified are labelled unverified, and you get a warning when access goes to an app on your own computer. See Manage OAuth credentials and authorized thunks and Publish your thunk as an MCP server for AI agents.
Gemini 2.5 Pro is retired. Gemini 2.5 Pro no longer runs. No newer Gemini model is a Pro model, so thunks are not moved automatically: a thunk still set to Gemini 2.5 Pro shows it as unavailable until you choose another model, such as Gemini 3.5 Flash or GPT-5.1. See Supported AI models and what happens when they change.
Fixes
Recent changes shows what each change did. Many change-log entries said no differences were found even when something had changed. Each entry now names what changed, or says the save changed nothing. Changes to members, workflow inputs and outputs, and a custom tool's intent and approval settings show old and new values, and older entries in a long history open correctly. See Versioning Support for your Thunk.
Tools fail less often, and say why when they do. Spreadsheet folder tools no longer vanish when a spreadsheet cannot be opened. Each call now fails with an error that names the folder and the reason, and a call naming a missing column asks you to re-index the folder. After an ownership transfer, a custom code tool that needs a connection the new owner lacks fails with an error naming it, instead of
tools is not defined, and the step and tool editors mark that connection Not available. When a model such as muse-spark-1.3 leaves out a level in a tool's nested input, Thunk.AI now restores it when there is only one way to read the call, instead of the step retrying without finishing. See Spreadsheet Content Folders and How custom code tools work.Admins are told when any Thunk email cannot be delivered. When our email provider blocks an address after an earlier hard bounce or spam complaint, your organization's admins now get a notice for error alerts, tool-approval requests, thunk invitations and new-thunk emails, as they already did for conversation replies. The notice names the blocked address and the kind of email, never its content.
A new webhook channel keeps the access you chose. When you added a webhook channel in Inbound Requests, the AI was always allowed to add and update data, whatever you picked. The channel now keeps your choice, including the default, AI can only add new data. Existing channels keep their current setting, so check them under Instructions and narrow any that should be. See Receive requests through a webhook or a shared form.
Published thunks are listed by their name. When you pick a plug-in connection or add another thunk's published tools, each thunk is now listed under its current name instead of the text it was first created with. The Name on a thunk's Export Interface tab is now the thunk's own name, so changing either renames both. See Publish your thunk as an MCP server for AI agents.