Paperclip

Paperclip starts its agent over ACP with a managed config directory.

Claude Code

boxer install claude-code
git add .claude .mcp.json && git commit -m "Run agent commands in boxer"

claude-agent-acp loads the project layer, so this is enough.

Codex

boxer install codex --user

Use the user layer. Paperclip does not bypass Codex's hook trust, and it builds its managed CODEX_HOME from ~/.codex/config.toml, which is where --user writes.

Secrets

Paperclip starts the agent with its own environment, not your shell's. A variable listed in secrets must be set where Paperclip runs, or in Paperclip's own settings for the agent. For Codex, also see Codex secrets.

Check it

Run boxer integrations to confirm the user-level Codex hook is in ~/.codex/config.toml. Then ask an agent to run python3 -c 'import sys; print(sys.platform)' in a task. It prints linux.

Turn on isolated workspaces in Paperclip's experimental settings. Without them, every run shares the project checkout, and so shares one sandbox.

Tested in the matrix.