Claude Code just got 10X Better (Codex + Gemini)
The video demonstrates how to combine Claude Code with OpenAI Codex and Google Gemini CLI to overcome Claude Code's recent performance regression. The presenter shows how to install both tools, connect them via plugins, and use a custom 'three-brain auto-router skill' that automatically delegates tasks to the most appropriate model.
Summary
The presenter, Jack Roberts, opens by addressing a recognized decline in Claude Code's performance, referencing Anthropic's own acknowledgment of this on April 23rd. His core argument is that no single AI model is perfect, but each has unique strengths — and combining them produces better outputs than relying on any one alone.
The video covers the installation process for two CLI tools: OpenAI Codex (via 'npm install openai codex') and Google Gemini CLI, both verified via version checks in the terminal. The presenter then walks through authenticating each tool by logging into the respective accounts (ChatGPT and Gmail), emphasizing that no additional API keys or per-token costs are incurred if users already have existing subscriptions.
Two GitHub plugins are then installed into Claude Code (demonstrated in the Antgravity IDE): one for Codex and one for Gemini. The Codex plugin is framed primarily as a code reviewer capable of adversarial critique, while the Gemini plugin is highlighted for its unique multimodal capabilities — specifically native video, audio, and large PDF analysis up to a 1-million-token context window.
The presenter demonstrates the Codex review feature by having Claude write JWT middleware, then asking Codex to review it. Codex returns a structured high/medium/low severity breakdown of issues, illustrating how it catches problems Claude might miss or loop on. A Gemini demo follows, where a short video is analyzed and described in detail — including visual elements, ambient sounds, and scene composition — entirely through a single Claude prompt.
The centerpiece of the video is a custom 'three-brain auto-router skill' (a markdown file available for free download) that Roberts built to automate model selection. The skill assigns Claude as the primary builder/IDE driver, Codex as the reviewer and rescue model (triggered automatically after Claude fails twice on the same problem), and Gemini as the specialist for video, audio, and large document tasks. A 'no self-review' rule is hardcoded to prevent Claude from looping on its own errors.
A full end-to-end demo shows the router being given a single prompt to analyze a video from Alex Hormozi, extract both transcript and visual overlay data, build an interactive HTML document from the combined content, and have Codex review the final output — all without manual model switching. The presenter also demonstrates using all three models simultaneously in a debate format (e.g., best social media platform to start a business in 2026), noting that Claude sub-agents work best up to around five agents before diminishing returns set in.
On pricing, the presenter outlines that ChatGPT Plus ($20/month) provides ~50 GPT-4.5 messages per three hours with no extra token cost via Codex CLI, while Gemini offers 1,500 Flash requests and 50 Pro requests per day for free — each request potentially covering a full two-hour video or 200-page PDF.
Key Insights
- Anthropic effectively confirmed a performance regression in Claude Code around April 23rd, which the presenter argues is the motivation for supplementing it with other models rather than abandoning it entirely.
- Gemini's free tier offers 50 Pro requests per day where each request can cover a full two-hour video or a 200-page PDF, which the presenter describes as 'incredibly generous' and sufficient for most users at zero cost.
- The presenter argues that Claude Code's core failure mode is that when it gets something wrong, it doesn't recognize the error and instead loops — so the auto-router skill enforces a hard stop after two failed attempts and automatically hands off to Codex.
- The presenter claims that using Claude sub-agents to debate a question works best capped at around five agents, after which diminishing returns appear — attributed to the main Claude client's ability to manage multiple concurrent threads.
- Using Codex CLI with an existing ChatGPT subscription incurs no additional per-token API costs, meaning the code review layer is effectively free for anyone already paying the $20/month ChatGPT Plus tier.
Topics
Full transcript available for MurmurCast members
Sign Up to Access