The ecosystem surrounding Claude Code is evolving rapidly, moving beyond simple chat interfaces into complex orchestration layers. Developer rbinar has released CLI-Dispatch, an open-source plugin that fundamentally changes how Claude Code handles task delegation. Previously, Claude Code's subagents were locked into Anthropic's proprietary models. This new tool shatters that constraint, allowing the primary Claude session to offload specific coding tasks to DeepSeek, Gemini (via Antigravity), OpenAI Codex, OpenCode, and GitHub Copilot.

Architectural Isolation and Deterministic Verification

The core innovation here isn't just model swapping; it's the rigorous isolation of the workflow. CLI-Dispatch employs a deterministic runner that spins up each delegated job in a separate git worktree. This ensures that experimental or potentially destructive code changes from external models do not pollute the main repository state. The system runs a user-defined verification commandβ€”such as node --testβ€”and returns a concise verdict alongside a patch. This creates a clear separation of concerns: the external model performs the heavy lifting, while Claude Code acts as the final reviewer and integrator.

Practical Implementation and Security

Installation is streamlined through Claude Code's plugin marketplace, requiring only a few commands to set up the necessary wrappers. Security is a primary focus; API keys for the external backends are stored in a local browser form and never pass through the Claude context window, mitigating data leakage risks. The tool supports five distinct backends, identified by short codes like ds for DeepSeek and cx for Codex. Users can invoke these workers via natural language prompts or direct CLI commands like /cli-dispatch:run, which bypasses LLM token costs for the orchestration layer itself.

Key Takeaways

  • CLI-Dispatch breaks the Anthropic-only limitation of Claude Code's native subagents.
  • Work is isolated in git worktrees to prevent repository pollution.
  • API keys are handled locally, never exposed to the Claude model context.
  • Supports five major external AI CLIs: DeepSeek, Gemini, Codex, OpenCode, and Copilot.
  • The orchestration layer itself consumes no LLM tokens, reducing costs.

The Bottom Line

This is a pragmatic step toward model-agnostic development environments. By treating external LLMs as specialized workers rather than competitors, CLI-Dispatch offers a realistic path for teams to leverage the best model for each specific task without abandoning their Claude-centric workflow.