Skip to content

Code Mode

Normally an AI client uses your connected apps one call at a time. It finds a tool, runs it, reads the whole result, then decides what to do next. That works, but a job like "find last week's unread emails from customers and summarize each one" means a dozen round trips. Every full result passes through the conversation, including the parts the model never needed.

With Code Mode, the client can also send Elaichi a short program instead. The program calls your tools, loops over the results, picks out what matters, and sends back only the answer. The intermediate data stays out of the conversation.

What changes for the client

The client gets one more tool beside search_tools and execute_tool: run_code. It takes the body of a JavaScript function that calls tools as await tools.<name>(args), using the exact names search_tools returns. A program passes each tool the same arguments execute_tool would. Nothing else changes: search_tools and execute_tool behave exactly as they did before run_code existed, and the client still picks whichever fits the task.

You don't write the programs; the AI client does. You'll see the program in your client's tool-call details.

What a program can and cannot do

  • The same access as a direct call. Every tool call a program makes goes through exactly the checks a direct call goes through: your own toolboxes and connections, your organization's restrictions, and the permissions you gave the client on Elaichi's consent screen. A program can't reach anything you couldn't reach yourself.
  • Deleting still needs permission to delete. A program can only call a tool that deletes something if the client was allowed to delete when you connected it, the same as a direct call.
  • In the Elaichi assistant, changes still ask you. A running program can't stop and wait for your approval. So a call that would normally show you an approval card is refused inside a program, and the assistant makes that call on its own, where you can approve it. Changes you've set to "always allow" run as usual.
  • No internet. A program has no network access of its own. It can only reach the outside world through your connected tools.
  • Everything is logged. Each run is one entry in your audit log, and each tool call inside it is its own entry, linked to the run. The program text itself isn't stored, because it often contains your data. The log keeps a fingerprint of it instead.

What you see in chat

In Claude and other clients that show Elaichi's cards, each run gets a short receipt: whether the program finished, every change it made as a line of its own ("Created a contact in Work"), and how many lookups it did. A change that failed, or was still running when the program ended, says so. In the Elaichi assistant, the same summary appears on the tool card, and pressing stop ends a running program straight away. Changes it had already made are not undone.

Limits

Each run can make up to 50 tool calls, run for up to 30 seconds, and return up to 100,000 characters. An organization can start up to 20 runs a minute and 1,000 a day; past either, clients do the rest with execute_tool, one call at a time. A run that hits a limit stops and says which one. The tool calls it already made are not undone.

Availability

Code Mode is part of every organization's connected tools, the same way execute_tool is. There is nothing to turn on: any AI client that can run your connected tools, and the Elaichi assistant, gets run_code beside them. Clients that never call it see exactly what they saw before.