Security

Local-first control for AI coding agents, with explicit approval boundaries.

AgentChat is designed so coding agents and project execution remain on the user's Runtime. The web app provides remote visibility and control without becoming the owner of local processes.

Local execution

Claude Code, Codex, and project work run on the user's machine through Runtime-owned sessions.

Structured approvals

Approval events are modeled as structured agent events, not trusted terminal text.

Session boundaries

Runtime Session IDs are immutable, and exited sessions are not resurrected.

Read architecture