crux task run¶

Run an agent in an isolated sandbox with scoped MCP access.

Usage¶

crux task run <prompt> [options]
crux task run --file <run.json>

Arguments¶

Argument	Description
`prompt`	Task string describing what the agent should do

Options¶

Option	Description
`--mcp <names>`	Space-separated MCP server names to enable in the sandbox
`--file <path>`	Load task configuration from a `run.json` manifest
`--keep`	Preserve sandbox directory after completion
`--no-stream`	Collect output instead of streaming it

Run Manifest Format¶

{
  "name": "weekly-research",
  "task": "Search for latest MCP security papers",
  "mcps": ["wikijs", "filesystem"],
  "skills": ["autoresearch"],
  "timeout_minutes": 30
}

Examples¶

# Ad-hoc run with MCPs
crux task run "Research MCP security" --mcp wikijs filesystem

# From a run.json manifest
crux task run --file tasks/research.json

# Keep sandbox for debugging
crux task run "Debug issue" --mcp filesystem --keep

# Collect all output at end instead of streaming
crux task run "Summarize repo" --mcp filesystem --no-stream