Harnesses and models

This page covers choosing the agent (harness) a sandbox runs, passing flags through to it, and choosing how its model calls are paid: your own Claude or ChatGPT plan, or your own API key.

1. Pick an agent. Name it after run.

   droneship run claude      # Claude Code
   droneship run codex       # OpenAI Codex CLI

2. Choose the model with --model: an alias such as sonnet or opus, or a full model id. It works for claude and codex, in a live session or a prompt run (-p).

   droneship run claude --model opus
   droneship run claude -p "fix the flaky test in tests/test_api.py" -d --model sonnet

   In a live session, pass other flags through to the agent after --. A prompt run passes no flags to the agent, and refuses any after --.

   droneship run claude -- --verbose

3. Choose how it pays for the model: your own API key, or your own plan.

   # your Anthropic API key
   droneship run claude --byok ANTHROPIC_API_KEY
   # your Claude Pro/Max plan
   droneship run claude --subscription

   With no key sent, claude and codex use the API key you saved in the console, if you saved one (below). Otherwise, on a terminal, they use your plan: you sign in inside the sandbox. For codex, the plan is your ChatGPT plan.

The agents

Agent	What runs	How you reach it
claude	Claude Code	Your terminal
codex	OpenAI Codex CLI	Your terminal

Registry ids such as claude-code also work.

• --ask is for claude only. Claude Code runs without asking before each command, because the sandbox is the boundary. --ask brings the prompts back.

How the agent reaches its model

Every agent calls its model through the Highflame AI Gateway unless you pass --direct. The gateway applies your policy and measures spend, which is what fills the AI spend column in the console.

You pay for the model calls, with your own API key or your own plan. On the gateway, your key is held by a sidecar outside the sandbox.

The choices:

Your own plan (--subscription). For claude (Claude Pro or Max) and codex (ChatGPT). You sign in inside the sandbox, as you would on your laptop, and model calls bill to your plan. The sign-in stores a long-lived credential in the sandbox, so a plan launch can’t run at --isolation KERNEL. Interactive sessions only.

Your own API key (--byok). Name the environment variable that holds the key. The CLI reads the value from your environment, not from the command line, and remembers the name for this profile, so you pass it once.

export ANTHROPIC_API_KEY=sk-ant-...
droneship run claude --byok ANTHROPIC_API_KEY

The variables each agent accepts are ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN (below) for claude, and OPENAI_API_KEY for codex. Your key still goes through the gateway. --no-byok stops sending it and forgets the remembered name.

A key saved in the console. On the console’s Settings page, under Model keys, save your own Anthropic or OpenAI API key for the project. Only you can change it, and nobody can read it back. A claude or codex launch that sends no key uses it, and says so:

Using your saved Anthropic key; --sub uses your plan instead.

It works for prompt runs too. On an interactive launch, --byok and --subscription take its place. On a prompt run, --byok and an exported key do. --direct never uses it.

Your plan’s token (CLAUDE_CODE_OAUTH_TOKEN). For claude only. Run claude setup-token on your laptop and export the token it prints. Model calls then bill to your Claude Pro or Max plan, with nobody signing in inside the sandbox, so this is how a prompt run uses your plan. Like a key, the token stays in the gateway sidecar, outside the sandbox.

claude setup-token
export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat01-...
droneship run claude -p "fix the flaky test in tests/test_api.py" -d

An interactive session takes it with --byok CLAUDE_CODE_OAUTH_TOKEN. It doesn’t work with --direct.

Direct to the provider (--direct). --direct skips the gateway. It needs --byok or --subscription, and model calls then bypass gateway policy and spend metering (the KERNEL isolation floor also drops). Routing through the gateway is the default and needs no flag.

Note

Prompt runs (-p) run with nobody attached to the sandbox to sign in, so they use your API key or your plan’s token. When ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN is exported, a prompt run sends it with no flag. If both are, it sends the API key. With neither, it uses the key saved in the console. With no saved key either, it stops before anything starts:

No ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN is set. Nothing was started.
  For your subscription plan: run claude setup-token and export the token as CLAUDE_CODE_OAUTH_TOKEN.

Prompt runs refuse your plan and --direct:

Invalid value for --prompt with --subscription: a plan login needs a person in the sandbox, and a prompt run has none. Set your provider key (for example ANTHROPIC_API_KEY), or export a `claude setup-token` token as CLAUDE_CODE_OAUTH_TOKEN, instead

Invalid value for --prompt with --direct: a prompt run can only reach models through the gateway. Drop --direct

With no terminal (in CI, say), nobody can sign in to a plan, so send a key or your plan’s token with --byok, or save a key in the console. With no key sent and none saved, the launch stops before anything starts. --direct never uses a saved key, so with --direct you must send one.

More: droneship run, Harness reference, Isolation and identity, Tools and MCP.

Next

• Background runs and logs: hand Claude a task with -p.
• Pair with an agent: work alongside Claude live.
• Egress: what the sandbox can reach.