Model Boss

Model Boss is an AI coding-agent orchestrator that coordinates Claude Code and Codex across multiple models so developers can delegate bounded implementation work to lower-cost worker models while a stronger model approves plans and final evidence. The host-selected main conversation loop stays immutable, and external workers receive sealed task packets in disposable git worktrees rather than the full conversation history. Write-capable routes run under OS sandboxing (macOS sandbox-exec, Linux bwrap, or WSL), and workers are limited to Read, Glob, Grep, Edit, and Write tools with Bash, Web, and MCP access disabled. Approvals are hash-bound: final sign-off is tied to the source snapshot, worker delta, and projected task patch hashes, so stale approvals are rejected. It offers two topologies: Lite, where authority stays in the main loop with an optional worker, and Max, which requires a distinct external reviewer that checks worker claims against independently captured process and Git evidence. Safety gates cap self-healing at three attempts and revision rounds at two. Models are declared as authority, balanced, or fast and verified live before dispatch, with support for Claude models (Opus, Sonnet, Haiku, Fable) and Codex tiers, plus Kimi and GLM through verified wrappers. It is aimed at engineers running multi-file implementations, migrations, and mechanical refactors with testable acceptance criteria, and steps aside for tiny single-file edits or pure discussion.

Best for developers who want to delegate bounded, multi-file coding tasks to lower-cost AI worker models while a stronger model reviews and approves the results.