Building agents from Claude Code¶

You can build and improve Donkit agents from Claude Code instead of the Builder chat. Claude does the thinking and writing on your own Claude plan; the agent itself lives and runs on Donkit, exactly like an agent made in the Builder.

What you need¶

A Donkit account, signed in in your browser.
Claude Code installed on your computer.
Access to the feature. It is rolling out gradually, so it may not be switched on for every workspace yet.

Install the plugin¶

Add the donkit-builder plugin to Claude Code from its plugin menu. It gives Claude the instructions and tools it needs to create, inspect, change, test and publish Donkit agents.

Connect your account¶

In Claude Code, open /mcp and connect Donkit.
A browser page opens and asks you to allow access. Check the workspace name, then allow. New agents Claude creates go to that workspace.
Return to Claude Code. You are connected.

Each connection is listed under Devices → Apps. Open it there to see what is connected, and remove a connection to cut off access at once.

Other apps¶

Other coding apps that can connect to MCP servers with a sign-in, such as Codex, can build Donkit agents too. Add https://app.donkit.ai/api/mcp-server/mcp as an MCP server in the app, then allow access on the page that opens, as above. The app gets the same tools, access and limits as Claude Code, and it appears under Devices → Apps under its own name. The Builder also sees which app made each change, so it builds on that work instead of redoing it.

The donkit-builder plugin works only in Claude Code, so another app does not get its ready-made instructions. Ask the app to read Donkit's build guides before it starts: the server offers them through its get_guide tool.

What the plugin can and cannot do¶

It edits agents you are allowed to build. Your own agents, and agents shared with you at Can edit, in any workspace you can edit them in, with the same rules as in the Builder (for example, only one person edits at a time).
It does not keep agent files on your computer. The agent's files stay on Donkit; Claude reads and changes them there.
It can create new agents. A new agent gets a name, a description and an empty interface, and goes to the workspace you approved.
It publishes only after every check passes. If a check fails, Claude is told what to fix first.
It cannot see the platform internals. Claude can read your agent's own files and the SDK and templates it builds on, but not how the Donkit platform itself runs.

What Claude can see¶

Claude works as you, with the same access to the agent's data that you would have as one of its users:

Your own records in tables where each person sees only their own, and the agent's shared data: tables open to everyone who uses the agent, and the structure of every table.
Not other people's records. Claude never sees the records of the people who use your agent, even in tables you can read on the Data tab. It cannot open the Data tab or the list of people who use the agent. To look at other people's records, use the Data tab yourself.
Tests run as you. When Claude tries the agent's backend or opens the preview, it is signed in as you and sees what you would see as a user.

Some Donkit installations keep agent data without separating it per person. There, Claude cannot use the tools that would show records or open the preview with live data, and it tells you so.

Preview¶

Claude can open the agent's interface in a real browser once its data has loaded, click, type and press keys in it (keyboard shortcuts included), and read what changed and which errors appeared. Opening the preview counts against the hourly limit on heavy actions (see Limits); clicking, typing, pressing keys and reading the page after that do not. When Claude waits for something to appear, it waits at most 30 seconds. A preview that sat idle for a few minutes closes, and Claude opens it again.

Checkpoints¶

Before Claude's first change to the agent's interface in each hour, Donkit saves a checkpoint of the interface. If you don't like what changed, ask Claude to go back: it can list the saved checkpoints and restore the one you choose. Restoring saves the current state first, so you can undo the restore too.

Checks and publishing¶

Before publishing, Claude runs every check that applies to the agent: whether the interface builds, opens and responds, fits a phone screen and follows the design rules; whether the backend starts and its queries match the agent's tables; and whether every workflow is ready to run. Publishing waits until every check passes. A check that could not reach a verdict (for example, a preview that did not load) also holds publishing back until it is run again and passes.

Credits¶

Building from Claude Code does not use Donkit credits for Claude's own work, because that runs on your Claude plan. Test runs of the agent during the build use credits like any other run of the agent.

Limits¶

Calls from Claude Code are limited per person: per minute for all calls, and per hour for heavy actions such as checks, opening the preview, installing packages and changing tables.
Creating agents is limited per day.

When you reach a limit, Claude shows which one and when it resets; try again after that.

Troubleshooting¶

Sign-in expired. Reconnect through /mcp in Claude Code.
"Someone else is editing this agent." Another person (or the Builder) is working on it right now. Wait until they finish, then try again.
A rate-limit message. You made too many calls in a short time, or created too many agents today. The message says which limit and when it resets.
"Not available". The tool Claude tried is not offered for this agent, for example because the agent has no interface yet or your plan doesn't include agent teams. The message says what to do.
The connection is missing under Devices → Apps. It was removed. Connect again from /mcp.

To build without Claude Code, see The Builder.