Get second opinions from rival AI models inside Claude Code

Dispatch is a Claude Code plugin that delegates tasks to external AI CLIs from inside the current session, triggered by naming the tool in natural language - saying check with codex, ask gemini for a second opinion, or have agy review this - rather than requiring a slash command, though /codex and /agy work as deterministic alternatives. It supports the OpenAI Codex CLI, defaulting to gpt-5.5 at medium effort in a read-only sandbox, and Google Antigravity CLI, defaulting to Gemini 3.5 Flash or a named model such as Claude Opus or GPT-OSS, and is agentic by default with no read-only mode. After explicit user approval for that delegation, Claude runs codex exec or agy -p through the Bash tool, always passing delegated prompts through stdin or a temp file with quoted here-doc delimiters so untrusted issue text, PR descriptions, diffs, or chat messages can never let the shell expand variables, backticks, or globs. Rather than echoing the raw response, Claude synthesizes and critiques it: summarizing key findings, stating explicit agreement or disagreement, and recommending next steps, treating the other model's output as a peer opinion rather than authority. Follow-up requests like continue with codex or what did gemini think about the other issue resume the right external session by its stored topic ID using a delta bridge that sends only what changed since the last exchange, not a full replay.

Worked examples cover a pre-merge Codex review that flags three issues where Claude agrees with one and identifies two as false positives, a cross-model architecture challenge where Gemini raises a hot-partition concern Claude hadn't flagged while dismissing its other two points as inapplicable, and resuming a prior Codex thread by UUID to confirm a traced code path. Best practices require naming the tool explicitly so dispatch never hijacks ordinary requests, requiring explicit approval before any external CLI delegation, confirming before write-mode since Codex workspace-write and all agy calls can edit files, and using dispatch for genuine second opinions where models disagree rather than routine validation.

Use it to get a second opinion from a different model family before merging or shipping, cross-check Claude's analysis against Codex or Gemini, delegate a side task like research or image generation without leaving Claude Code, or triangulate a decision across multiple models. Do not expect agy to stay read-only when asked to analyze only - it can still edit files and run commands, so analysis calls should run from a clean git state with a git status check afterward. Session topic IDs live only in conversation memory and are lost on context compaction, and agy's first call in a session can take two to three minutes to cold-start.

Given a natural-language delegation request, it produces a synthesized critique of the external model's output - agreement, disagreement, and recommended next steps - rather than the raw response, plus optional image generation via gpt-image-2 or Nano Banana Pro depending on the CLI used.

OpenAI Codex CLI (OAuth via codex login) and Google Antigravity CLI (free Google account sign-in); the plugin never stores, reads, or passes API keys, and treats external model output as data rather than instructions. Related skills cover dispatching multiple parallel subagents and Codex-integrated code review.

Developers using Claude Code who want a second model's opinion on a review, architecture decision, or side task, and want Claude to adjudicate disagreements rather than just relay them.