AI agents in the terminal

Claude Code and Codex in the sandbox. On your keys, within your budget.

Q: Who pays for model usage?

Your organization, directly to Anthropic or OpenAI. Kendor doesn't resell tokens or add a margin.

Q: What does connecting without a key mean?

For Anthropic, your Claude Console organization can trust Kendor through workload identity federation, so no long-lived API key is created or shared. OpenAI uses an API key.

Q: Doesn't allowing AI make the score meaningless?

The grade is a verified re-run of what was submitted, so it reflects code that actually works. How the candidate got there — what they asked, what they accepted, what they rewrote — is the part your reviewers judge.

If the job involves working with an agent, screen for it. Candidates run Claude Code or Codex in their sandbox terminal, billed to your organization's own Anthropic or OpenAI key. The key never enters the sandbox, and you can cap each attempt with a token budget.

Request access See pricing

billing-service38:12AI tokens: 188k of 1M usedTokens the assistant has read and written across this screenAI assistant: run claude or codex in the terminal

send.gosend_test.go

42func (s *Sender) Send(ctx context.Context, inv Invoice) error {

43 if inv.Total <= 0 {

44 return nil

45 }

47 pdf, err := s.render(ctx, inv)

48 if err != nil {

49 return fmt.Errorf("render invoice %s: %w", inv.ID, err)

50 }

52 return s.mail.Send(ctx, inv.CustomerEmail, pdf)

53}

Terminalclaude

▸ Sandbox shell — you are in /workspace

AI assistant enabled for this assessment — run claude or codex to start it.

$ claude

> Invoices with a zero total are still being emailed. Find out why and fix it with a test.

● Read(internal/invoice/send.go)

⎿ Read 184 lines

● Read(internal/invoice/send_test.go)

⎿ Read 96 lines

● The guard checks inv.Total < 0 instead of inv.Total <= 0, so zero-total invoices get through.

● Update(internal/invoice/send.go)

⎿ Updated internal/invoice/send.go with 1 addition and 1 removal

● Update(internal/invoice/send_test.go)

⎿ Updated internal/invoice/send_test.go with 18 additions

● Bash(go test ./internal/invoice/...)

⎿ ok acme/billing/internal/invoice 0.311s

● Fixed, and the new test covers a zero-total invoice. Want me to check the other callers of Total?

mainGo 1.26UTF-8

Your key never enters the sandbox.

Claude Code and Codex in the sandbox talk to Kendor's gateway with a session token that opens nothing on its own. The gateway checks the session and its budget, adds your organization's key and forwards the call to Anthropic or OpenAI. There's nothing in the environment for a candidate to print or copy.

Keys are encrypted at rest, and only organization owners can change them
Anthropic can connect through identity federation, with no long-lived key at all
The gateway only reaches Anthropic and OpenAI. It isn't a way out to the internet

Sandbox$ env | grep -E 'ANTHROPIC|OPENAI'ANTHROPIC_BASE_URL=http://kendor-gateway:4102/anthropicANTHROPIC_AUTH_TOKEN=eyJhbGciOiJIUzI1NiIs…OPENAI_BASE_URL=http://kendor-gateway:4102/openai/v1OPENAI_API_KEY=eyJhbGciOiJIUzI1NiIs…a session token, not a vendor key

Kendor gatewayChecks the sessionChecks the budgetAdds your key

ProviderAnthropicOpenAIEnds in …q7Lm · updated 9/23/2026

A budget per candidate, a cap per month.

Give each screen a token budget per candidate, counted across all its challenges, and set a monthly cap for the whole organization. Candidates see a ring in their header fill toward the budget, and the gateway stops the agent as soon as either limit is reached.

Counts what the agent reads and writes, not cached context
Usage is billed to your own Anthropic or OpenAI account

AI assistantsLet candidates use Claude Code and Codex in the sandbox terminal, billed to Acme's own vendor accounts.

Billing modeWho pays for inference and where the key lives.

Your API keysInference is billed to your Anthropic / OpenAI account. Keys are encrypted at rest and never enter a candidate's sandbox.Included with KendorComing soon

Anthropic API keyPowers Claude Code. Create a key at console.anthropic.com.

Ends in …q7Lm · updated 9/23/2026API keyIdentity federationPaste a new key to replace itSaveTest connectionRemove

OpenAI API keyPowers Codex. Create a key at platform.openai.com.

Not configuredPaste a keySave

Monthly token capWhat all candidates together may spend through the assistant in a calendar month: what it reads and writes, cached context not counted. Leave empty for no cap.

20million tokensSave6.4M of 20M tokens used this month

Off until you turn it on.

AI assistants stay off for every screen until you switch them on in the builder. Candidates get the agents your organization has keys for, and their start page tells them before they begin what goes to your provider and that reviewers can see their terminal.

Options

ProctoringTrack tab switches, external pastes and fullscreen exits during the attempt.AI assistants in the terminalCandidates can run Claude Code and Codex in the sandbox terminal through your organization's API keys.

Token budget per candidate1million tokensWhat the assistant reads and writes, added up across all challenges. Cached context is not counted.

Candidate's start pageAI assistants are availableRun claude or codex in the terminal. Your workspace files and what you type to the assistant are sent to the hiring organization's AI provider (Anthropic or OpenAI) under its account, and your terminal — including the assistant's work — is visible to reviewers.

Watch the terminal live, read-only.

Open the Terminal tab on a candidate's review page while they work and you see their shell as it happens: what they ask the agent, what it runs and what they keep. Reviewers can watch, but they can't type into it.

SolutionTimelineTerminalAI agentsActivityComments Open in editor

shell-ghfd5ishell-trz3fkLive · read-only view of the candidate’s terminal

> Also raise the retry limit to 5 attempts

● Update(internal/mail/retry.go)

⎿ Updated internal/mail/retry.go with 3 additions and 1 removal

$ git diff --stat

internal/mail/retry.go | 4 +++-

1 file changed, 3 insertions(+), 1 deletion(-)

$ git checkout internal/mail/retry.go

Updated 1 path from the index

$ vim internal/mail/retry.go

Read every prompt afterwards.

The AI agents tab keeps the conversation in order: each prompt, the tools the agent called and what it answered, with the time and model on every turn. You judge how someone works with an agent, not only the code it left behind.

Candidates are told up front that the conversation is saved for reviewers

SolutionTimelineTerminalAI agentsActivityComments Open in editor

10:42:07 AMclaude-sonnet-4-5PromptInvoices with a zero total are still being emailed. Find out why and fix it with a test.Tool · Read{"file_path":"/workspace/internal/invoice/send.go"}▸ Tool output
10:42:19 AMclaude-sonnet-4-5AgentThe guard checks inv.Total < 0 instead of inv.Total <= 0, so zero-total invoices get through.Tool · Edit{"file_path":"/workspace/internal/invoice/send.go","old_string":"if inv.Total < 0 {","new_string":"if inv.Total <= 0 {"}▸ Tool output

Switch agents on mid-interview.

In a live interview or pair coding session, the host turns AI agents on or off for everyone from the room's top bar. The running sandbox picks up the change straight away, without a restart.

Everyone in the room sees whether AI is on
Same organization keys and gateway as your screens

← LeaveLive pair coding/Ada OkaforLiveWhiteboardCodeAI on24:10End session

retry.goretry_test.go

12const maxAttempts = 3

14func backoff(attempt int) time.Duration {

15 return time.Duration(attempt*attempt) * 200 * time.Millisecond

16}

Terminal

▸ Sandbox shell — you are in /workspace

AI assistant enabled for this assessment — run claude or codex to start it.

$ claude

How it works

Two settings, then it's in their terminal.

Add your provider key
An Anthropic or OpenAI key in Settings → AI assistants, once for the organization. Set a monthly cap there if you want one.
Turn it on for a screen
Switch on AI assistants in the screen builder and give it a token budget per candidate.
Candidates run it
They find claude or codex ready in their sandbox terminal, and you can watch from the review page.

Questions

Who pays for model usage?

Your organization, directly to Anthropic or OpenAI. Kendor doesn't resell tokens or add a margin.

What does connecting without a key mean?

For Anthropic, your Claude Console organization can trust Kendor through workload identity federation, so no long-lived API key is created or shared. OpenAI uses an API key.

Doesn't allowing AI make the score meaningless?

The grade is a verified re-run of what was submitted, so it reflects code that actually works. How the candidate got there — what they asked, what they accepted, what they rewrote — is the part your reviewers judge.

More of Kendor

All features

Run your next screen on Kendor. Read what the code says.

Kendor works with engineering teams in the EU, the UK and the US. Tell us what you're hiring for and we'll set up your organization with you.

Request access See pricing

Claude Code and Codex in the sandbox. On your keys, within your budget.

Your key never enters the sandbox.

A budget per candidate, a cap per month.

Off until you turn it on.

Watch the terminal live, read-only.

Read every prompt afterwards.

Switch agents on mid-interview.

Two settings, then it's in their terminal.

Add your provider key

Turn it on for a screen

Candidates run it

More of Kendor

Run your next screen on Kendor. Read what the code says.