--harness, -H | choice | Harness running the agent (choices: claude-code, claude-desktop, codex, codex-app, cursor) |
--model, -m | string | Model used by the harness |
--file, -f | file | Local file uploaded and placed in the sandbox before setup, its path given to the agent with the prompt; repeat for several |
--repository, -r | string | Git repository cloned into the sandbox |
--env, -e | string | Environment variable set in the agent’s processes, as KEY=VALUE, or KEY to pass its local value |
--setup | string | Shell script run in the sandbox after checkout, before the first prompt; from a file: —setup ”$(cat setup.sh)“ |
--teardown | string | Shell script run in the sandbox after the run ends, before the sandbox is destroyed; from a file: —teardown ”$(cat teardown.sh)“ |
--mcp | string | MCP server the agent can use, as NAME=URL, called without credentials, or NAME=CONNECTION_ID, called as the user who connected it |
--mcp-header | string | Header added to every request to the MCP server NAME given by URL, typically its API key, as NAME=HEADER: VALUE, such as ‘docs=Authorization: Bearer KEY’ |
--browser | boolean | Give the agent a browser, a headless Chromium in its sandbox whose pages can offer it tools through WebMCP. The Codex app and Claude Cowork always use their own browser instead |
--budget, -b | number | Most the run may spend, in US dollars |
--one-shot | boolean | End the run once its first prompt finishes (default); —no-one-shot waits for follow-ups |
--idle-timeout | integer | Seconds an idle run waits for a follow-up before it ends |
--time-limit | integer | Longest the run may last, in seconds, before it is stopped |
--idempotency-key | string | Retry safely: running again with the same key returns the run already created |
--label | string | Label attached to the run to find it again, as KEY:VALUE, such as ticket:ENG-42; repeat for several |
