Agents¶

Kube Foundry supports three agent runtimes. Each runs in its own container image with an entrypoint that handles git setup, skill injection, agent execution, and PR creation.

Supported agents¶

Agent	Image	API Key Secret Field	MCP Support
`claude-code`	`agent-images/claude-code`	`ANTHROPIC_API_KEY`	Yes
`codex`	`agent-images/codex`	`OPENAI_API_KEY`	No
`open-code`	`agent-images/open-code`	`ANTHROPIC_API_KEY`	Yes

Sandbox lifecycle¶

Every task goes through the same lifecycle inside the sandbox pod:

Configure git -- set credentials, author name/email, authenticate with gh
Clone repo -- shallow clone of the specified branch
Create work branch -- factory/<taskname>
Apply skills -- inject prompts, files, MCP config, run init commands
Run agent -- execute the AI agent with the task description
Commit and push -- stage all changes, commit, push to work branch
Create PR -- open a pull request against the base branch

Choosing an agent¶

Set the agent field in your SoftwareTask:

spec:
  agent: claude-code  # default

Claude Code¶

Best for complex, multi-file tasks. Supports MCP servers for external tool access. Uses ANTHROPIC_API_KEY.

Codex¶

OpenAI's coding agent. Uses OPENAI_API_KEY. Does not support MCP servers -- a warning is logged if MCP servers are configured.

OpenCode¶

Open-source coding agent that uses Anthropic's API. Supports MCP servers. Uses ANTHROPIC_API_KEY.

Credentials¶

The credentials secret must contain the appropriate API key for the chosen agent:

# For Claude Code or OpenCode
kubectl create secret generic factory-creds \
  --from-literal=ANTHROPIC_API_KEY=sk-ant-... \
  --from-literal=GITHUB_TOKEN=ghp_...

# For Codex
kubectl create secret generic factory-creds \
  --from-literal=OPENAI_API_KEY=sk-... \
  --from-literal=GITHUB_TOKEN=ghp_...

Custom resources¶

Resource	Description
`resources.cpu`	CPU limit for the sandbox pod (default: `2`)
`resources.memory`	Memory limit (default: `4Gi`)
`resources.timeoutMinutes`	Max execution time (default: `30`, max: `120`)