Claude Code with three brains: how to get GPT 5.5 and Gemini into your terminal at no extra cost

What just happened

Claude Code has been in a slump for the past few weeks. You noticed it, I noticed it, half of LinkedIn noticed it. You ask, it answers off target. You correct it, it answers off target again.

Anthropic admitted the performance drop itself on 23 April. They're working on it. But while we wait, there's a better way than throwing Cloud Code out or continuing to fight it.

The way is called a multi-brain setup. Three models, one terminal, each one doing what it was built for.

The picture behind it

Imagine you run a workshop. You have a builder who builds fast. A second man who checks after every step whether the structure holds. And a third who reads plans, big, thick plans that nobody else can get through.

When all three work together, the job site runs. When you only have one, sooner or later he builds himself into a wall. That's exactly what's happening right now with Claude Code running solo.

The three models and what each one can do

Claude Code, the builder. Writes the code, does the building itself. Stays the IDE harness that sits up front and talks to you.

GPT 5.5 (via Codex CLI), the reviewer. Cross-checks Claude's output and finds the logic errors Claude can't see in its own work. Adversarial review means devil's advocate at the push of a button. You get a list of high/medium/low findings instead of collecting all the bugs yourself.

Gemini 2.5 Pro, the eyes. Native video analysis up to 2 hours. PDFs with 200+ pages in one request. 1 million tokens of context. Audio up to 10 hours. Claude on its own quits long before that. Gemini reads it in one go.

Why three brains beat one good brain

Every model has blind spots. When you look at the same problem from three different angles, most of those spots close.

Claude is good at building, but not good at criticizing itself. When it goes down the wrong path, it doesn't notice on its own. GPT 5.5 sees that from the outside.

GPT is good at structuring and reviewing, but not as good at swallowing a whole codebase or an 8-hour audio file. Gemini swallows it without a pause.

Gemini has the largest context window, but not the finest feel for code at the detail level. Claude and GPT have that.

Three tools. Three strengths. One toolbox.

Setup in five steps

Step 1: Install Codex CLI

In the terminal:

npm install -g @openai/codex

Then check:

codex --version

If a version number comes back, it worked.

Step 2: Install Gemini CLI

npm install -g @google/gemini-cli

Then check:

gemini --version

Step 3: Log in to both

For Codex: codex login, which opens a login page where you connect your ChatGPT account. It uses your existing ChatGPT Plus or Pro subscription. No extra token pricing, no API keys.

For Gemini: gemini, and on the first run it walks you through the login. You connect your Google account, done. Completely free on the free tier.

Step 4: Claude Code plugins

For Claude Code to call the other two from inside your workflow, you need two plugins. There are two GitHub repos:

The Codex plugin for Claude Code. It lets Claude call the Codex CLI as a sub-tool, for example for "review this code" or "adversarial check".

The Gemini plugin for Claude Code. It lets Claude call on Gemini for long contexts, videos or PDFs.

You install the repos through the plugin marketplace in Claude Code. Each repo has a detailed how-to. I link my tested version below.

Step 5: Create a routing skill

The real lever: a skill in your .claude/skills folder that decides which model steps in when. The routing logic I run:

→ If Claude fails at a task twice (failure detection): hand over to Codex for a rescue review. Hard stop, no "one more round with Claude".
→ If the task involves video, audio, a PDF with 50+ pages or a long codebase search (risk path detection): straight to Gemini, because Claude's context window tears before that.
→ If code was written that is meant to go into the repo: run Codex over it automatically before Claude makes the commit. Adversarial review as the default layer.
→ For strategic questions ("Which architecture? Which stack? Which business model?"): parallel consensus. All three models get the same question and answer independently. Claude then merges the three answers into one.

The skill is a Markdown file with a few if-then rules. Write it once and it runs on its own.

What it costs per day

Claude Code: your existing plan, $20 or $200 a month depending on the tier.

ChatGPT Plus: $20 a month. It includes 50 GPT-5.5 messages every 3 hours. Codex draws from exactly this budget, with no extra pricing per token. If you're on the $200 Pro tier, you get what amounts to unlimited GPT-5.5 for code reviews.

Gemini: free. 1,500 Flash requests per day (for fast text and image tasks). 50 Pro requests per day, and every single one of those Pro requests can be a 2-hour video or a 200-page PDF. For 95% of users that's enough without an upgrade.

So you either pay nothing extra (if you already have ChatGPT Plus) or €20 more a month. In return you get access to two more frontier models, and in its own field (long contexts, multimodal) Gemini has no rival.

Real use cases from my week

Use case 1: Code review after every feature.

Before: Claude builds, I look it over, find 60% of the bugs, the rest shows up in production. Now: Claude builds, I say "Codex, review this in adversarial mode". Codex returns a structured list: security holes, missing validation, hardcoded secrets. It keeps finding things Claude missed. Bug rate in production cut in half.

Use case 2: Contract review for a client.

A 180-page service contract in legal German. Claude on its own would have buckled at the context window. Instead, Gemini swallows the whole PDF in one Pro request. It returns a structured summary, flags critical clauses, compares them with standard clauses. 15 minutes instead of half a day.

Use case 3: Reverse engineering a viral Reel.

A competitor breaks through on TikTok with a series. I give Gemini a 90-second video. Gemini returns a full transcript, a second-by-second visual overlay (which edit, which cut, which camera move) and the pacing patterns it can spot. We take the mechanics, pour them into our own brand, post the following week.

Use case 4: Multi-model debate for strategic decisions.

Question: "Which platform should my client prioritize in 2026: LinkedIn, TikTok, YouTube?" I start a parallel consensus session: all three models get the same question in the same form. Claude argues for deep audience loyalty, GPT for scalable discovery, Gemini brings in data points from current studies. I get three different lenses instead of one monolithic recommendation.

The truth about where Claude Code stands right now

I have no alliance with any model. Cloud Code was magic in February, noticeably worse in early April, and now they're in the middle of fixing it. While they fix it, solo Claude isn't the best workflow.

But Codex alone isn't the answer either. GPT 5.5 is great at reviewing, but Claude is still better at writing with style and nuance. Gemini has the biggest context, but in code detail it needs the other two.

The answer isn't one of the three. The answer is the combination.

My advice

If you work with Claude Code regularly, block out an afternoon. Install Codex and Gemini CLI. Do the plugin setup. Write yourself a small routing skill (Claude helps you with it, ironically).

After that you have a tool that leaves any solo Claude setup far behind. At no extra cost in most cases.

And an honest question: if you needed an operation today, would you take the one general practitioner, or the team of surgeon, anesthesiologist and imaging?

That's the exact question your terminal is asking right now.

FAQ

What is a multi-brain setup in Claude Code?

A multi-brain setup is a terminal configuration in which Claude Code works together with Codex CLI (GPT 5.5) and Gemini CLI. Claude stays the main tool for writing code. GPT 5.5 takes over the review, and Gemini handles long contexts like videos or PDFs. That way the weaknesses of the individual models cancel each other out.

Does the multi-brain setup cost extra money?

If you already have ChatGPT Plus for $20 a month, you pay nothing extra, because Codex runs on that budget. Gemini is free on the free tier with 1,500 Flash requests and 50 Pro requests per day. At most you need an existing Claude Code subscription and a ChatGPT account.

How do I install Codex CLI and Gemini CLI?

You install Codex CLI in the terminal with npm install -g @openai/codex and then check it with codex --version. You install Gemini CLI with npm install -g @google/gemini-cli and check it with gemini --version. Then you log in to both with your existing account: ChatGPT for Codex, your Google account for Gemini.

Why has Claude Code gotten worse lately?

Anthropic itself admitted a performance drop in Claude Code on 23 April and is working on a fix. Since early April, users have reported that answers miss the mark more often than they did in February. Until the fixes are in, running Claude Code solo isn't the most reliable workflow.

What does a routing skill in the .claude/skills folder do?

A routing skill is a Markdown file with if-then rules that decides which model gets which task. For example, it detects when Claude fails twice and then hands off to Codex for a review automatically. For video, audio or PDF tasks it routes straight to Gemini, because Claude's context window is too small for them.