Skip to content
KendorDocs

How to let candidates use Claude Code or Codex

For owners and editors ·

Switch on AI assistants in the terminal for a coding screen or a live coding interview. On your keys candidates run Claude Code or Codex; on their own account, any of seven agents.

Turn on AI assistants in the terminal in a screen's settings, or leave the AI switch on for a live coding interview. Candidates then start an agent in the sandbox terminal and work with it on the real workspace.

Which agents they can run depends on whose account pays for them:

AI account Agents Token budgets
Your organization's API keys Claude Code (claude) with an Anthropic key, Codex (codex) with an OpenAI key Apply
The candidate's own account Claude Code (claude), Codex (codex), Antigravity (agy), GitHub Copilot (copilot), Cline (cline), OpenCode (opencode) and Cursor (cursor-agent) Don't apply

On your organization's keys, an owner has to connect at least one key first. See Connect your Anthropic and OpenAI keys.

On a coding screen

AI assistants are off for every screen until you switch them on.

  1. Open the screen in the builder.
  2. In the settings rail, find Options and turn on AI assistants in the terminal.
  3. Under AI account, choose Your organization's API keys or The candidate's own account.
  4. On your organization's keys, optionally enter a Token budget per candidate, in million tokens. Leave it empty for no limit. See Set token budgets for AI agents.

If your organization has no key connected yet and the screen uses your keys, the rail says so: "No API key is configured yet, so candidates will see the assistant as unavailable." The Add one in Integrations. link takes you to the right place.

The setting applies to every candidate on the screen. If you change it while candidates are mid-screen, their running sandboxes pick up the change without a restart. That covers turning the agents on or off, changing the budget, and an owner adding or removing a key.

Tip. When you try a challenge yourself in the builder, Claude Code and Codex are available in your own sandbox whenever your organization has keys connected, so you can check how the task feels with an agent before you invite anyone.

On a live coding interview

Live coding interviews start with AI agents on. You can change that at three levels:

Where What it sets
The live coding interview's page, AI assistants in the terminal The default for new sessions of that interview
The Invite to a live coding interview dialog, Claude Code and Codex in the terminal This one session
The room's top bar, AI on / AI off The running session, for everyone in it

In the room, interviewers can flip the switch at any time. It covers the whole shared sandbox, the change applies straight away without a restart, and the room notes who turned AI agents on or off. The candidate sees which way the switch is set but can't change it.

If a key is missing, a notice under the switch says which agent won't work, for example "There is no OpenAI key, so Codex won’t work — only Claude Code."

What candidates see

Before a screen starts, the start page lists AI assistants are available with the command to run. On your keys that's claude, codex, or claude or codex, depending on the keys you have. It tells them that their workspace files and what they type to the assistant go to your organization's AI provider under your account, and that reviewers can see their terminal, including the assistant's work. When conversations are recorded, it also says that the prompts, replies and the commands the assistant runs are saved and shown to reviewers.

The start page only shows this when the screen has assistants on and at least one key works.

Inside the editor:

  • A hint in the header reminds them: "AI assistant: run claude in the terminal".
  • A new terminal prints "AI assistant enabled for this assessment — run claude to start it." (or codex, or both).
  • A small ring in the header fills toward the token budget. Hovering it shows "AI tokens: 188k of 1M used", and it turns amber close to the limit.

On your organization's keys, Claude Code and Codex open already set up for the workspace, with no sign-in or onboarding. On the candidate's own account each agent asks them to sign in first; see below. If a candidate runs an agent that isn't enabled, the terminal answers with one line, such as "The AI assistant is not enabled for this assessment." or "codex is not enabled for this assessment — use claude."

On the candidate's own account

Under AI account in a screen's settings, choose The candidate's own account instead of Your organization's API keys. Candidates can then run any of the seven agents and sign in to their own Claude, ChatGPT, Google, GitHub Copilot, Cline or Cursor account in the sandbox, on their own plan. OpenCode can also use its free models with no account. None of your keys or budget is used, so this works even if your organization has no key connected, and Kendor doesn't count tokens. Their sign-in carries over between challenges and is deleted when they submit the assessment or their time runs out.

Reviewers still see the terminal and the recorded conversations. Antigravity's conversation isn't recorded on its own, so its work shows only in the terminal replay. The review page's header reads AI on own account instead of a token count.

When you try a challenge yourself in the builder, your sandbox always uses your organization's keys, so only Claude Code and Codex are there.

Send candidates How to connect your own AI account so they know what to expect.

What reviewers see afterwards

On a submission's review page:

  • The header shows the attempt's token use, such as "AI 1.4M / 2M tokens", or "AI 1.4M tokens" when there is no budget.
  • The AI agents tab shows each recorded conversation in order: the candidate's prompts, the agent's replies, the tools it called and their output, with the time and model on every turn.
  • The Terminal tab replays the terminal, including what the agent ran there.

The AI agents tab only appears when at least one conversation was recorded for that challenge. A live coding interview's result page has the same AI agents tab, or "No AI agent was used in this session" when nobody started one.

Note. Reviewers judge how someone worked with the agent, not only the code it left behind. Review a submission covers the rest of the review page.

Was this article helpful?