An agent framework that ships with a debugger, a trust model, and a web UI

What it does ConnectOnion is a Python framework that wraps raw LLM calls into autonomous agents. You write tools as ordinary Python functions, hand them to an Agent instance, and the framework supplies the surrounding infrastructure: an HTTP backend, a ready-made chat frontend, built-in tools for email and file access, and a plugin system with 12 lifecycle hooks. It also exposes a multi-agent trust layer and a P2P relay so agents can discover and call each other remotely.

The interesting bit The project treats Claude Code’s internal feature set as a target rather than a competitor, packaging its capabilities—context compression, sub-agent spawning, and autonomous work mode—as open plugins that work with any agent. It even loads Claude Code skill files directly from .claude/skills/ without conversion, and offers an @xray decorator that drops you into an interactive debugger mid-execution to inspect or rewrite local variables.

Key highlights

• Ships with a built-in web UI (chat.openonion.ai) and API layer; calling host() turns an agent into a remotely callable service over HTTP and P2P.
• Includes ready-made tool integrations (Gmail, Outlook, Google Calendar, browser automation, shell) and a co copy CLI command to fork any built-in tool into your own codebase for modification.
• Approval workflows are plugin-based: dangerous operations like shell commands automatically prompt for human sign-off, but you can swap or disable the logic entirely.
• Skills auto-discover from three levels (project, user, built-in) and reuse Claude Code’s skill format without conversion.
• Trust decisions between agents use fast preset rules (open, careful, strict) that resolve before burning tokens on an LLM call.

Verdict Developers who want to move from a Python script to a deployed agent team without wiring up FastAPI or React will find this a convenient all-in-one choice. If you prefer composing your own stack from specialized libraries, the bundled frontend and opinionated tool ecosystem may feel like overkill.