SakanaAI Fugu Guide: How to Enable a Multi-Agent Model in Codex

Wed, 24 Jun 2026 07:41:25 +0800

The most interesting thing about SakanaAI/fugu for everyday developers is not only that it is a multi-agent model. It is that it already provides a direct entry point for using it with Codex.

If you are searching for:

How to install Fugu
How to use SakanaAI Fugu
How to connect Fugu to Codex
What is codex-fugu

This guide is written as an installation and first-use tutorial.

Official links:

GitHub repository: SakanaAI/fugu
Product page: Sakana Fugu

What Fugu Is Good For First

Before installing it, be clear about one thing: Fugu is not necessarily the best choice for every task.

It is better suited to complex, multi-step development tasks that benefit from model coordination and verification, such as:

Reading a codebase and summarizing its structure;
Locating bugs;
Modifying code and continuing to run tests;
Reviewing code;
Analyzing complicated errors;
Designing engineering plans;
Handling tasks that require long-chain reasoning;
Paper reproduction, research analysis, security analysis, and other hard tasks.

If you only ask something simple like “How do I read a file in Python?”, a normal model is enough. Fugu is more useful when you want the AI to think through more steps and check more rounds.

Check Before Installation

According to the SakanaAI/fugu README, the official one-line installer supports:

Ubuntu;
macOS.

Windows users, or users whose one-line installation does not finish, should follow the official manual instructions.

Before installing, check a few things:

Your environment can access https://sakana.ai/;
curl is available in your terminal;
You know whether you are on Ubuntu, macOS, or Windows;
You already have a working Codex environment;
If you are behind a company proxy or firewall, confirm that the installer can be downloaded.

If you are in the EU/EEA, also note the availability limitation mentioned on the official page: Fugu is currently unavailable in the EU/EEA while Sakana works on GDPR and EU-specific compliance.

Install Fugu With One Command

The installation command shown in the official README is:

`1`	`curl -fsSL https://sakana.ai/fugu/install \| bash`

This command downloads the installer from Sakana’s official address and passes it to bash.

If you prefer a more cautious approach, you can first print the script content without executing it:

`1`	`curl -fsSL https://sakana.ai/fugu/install`

After confirming the source and content, run the official installation command.

Commands like curl | bash are convenient, but they execute a remote script. Copy them only from the official README or official documentation, not from random comments, screenshots, or reposts.

Start codex-fugu

After installation, the official README shows this launch command:

`1`	`codex-fugu`

If the command starts normally, Fugu has been connected to your Codex workflow.

Start with a small task rather than a huge repository problem. For example:

`1`	`Please read the current project structure and explain what the main directories do.`

Or:

`1`	`Please review the latest changes and identify risks that may affect build or deployment.`

These tasks are specific enough, but not too heavy. They are good for observing Fugu’s response style, speed, and stability.

Suggested First Tests

When you use codex-fugu for the first time, try this order:

Ask a repository structure question;
Ask it to inspect a specific file;
Ask it to analyze a small bug;
Then try a cross-file task.

For example:

`1`	`First browse this repository and tell me where the entry files, content directories, and build scripts are.`

Then:

`1`	`Please check whether content/post/2026/06/127/index.zh-cn.md has complete front matter.`

Then go one step further:

`1`	`Please identify whether the title, slug, and description of this article are inconsistent, and give me revision suggestions.`

This gradual approach is safer. You can first judge whether it understands the project before handing it larger tasks.

Codex Tasks Suitable for Fugu

Fugu packages multi-agent coordination as a model/API, so it is more suitable for tasks that require several rounds of reasoning.

1. Code Review

You can ask:

`1`	`Please review the current uncommitted changes from a code-review perspective. Prioritize bugs, regression risks, and missing tests.`

This kind of task requires reading multiple files, understanding context, and judging risk. It is a better fit for multi-agent coordination than simple Q&A.

2. Bug Diagnosis

You can ask:

`1`	`This feature is failing. Please use the logs to identify possible causes, inspect the related files, and then propose the smallest fix.`

Bug diagnosis usually requires step-by-step investigation. A system like Fugu is suitable for chaining analysis, verification, and repair suggestions.

3. Cross-File Refactoring

You can ask:

`1`	`Please change this config value from hardcoded logic to environment-variable loading, and check whether related tests need updates.`

Cross-file changes easily miss boundaries. When using Fugu, ask it to explain the impact area before editing.

4. Documentation and Tutorial Writing

You can ask:

`1`	`Please use the current README and code structure to write a beginner installation and run guide.`

This type of task requires reading materials, extracting structure, and writing executable steps, which also fits Fugu well.

5. Hard Analysis Tasks

For research, paper reproduction, security analysis, or long engineering investigations, consider Fugu Ultra.

For day-to-day development, you do not need to start with Ultra. Test regular Fugu first for cost, speed, and quality.

How to Choose Fugu or Fugu Ultra

Sakana’s official page mentions two models:

Fugu
Fugu Ultra

You can think of them this way:

Model	Suitable scenarios	Orientation
Fugu	Daily coding, code review, chatbots, ordinary complex tasks	Balance performance and latency
Fugu Ultra	Hard reasoning, research, competitions, security analysis, long-chain tasks	Prioritize output quality

Practical advice:

Use Fugu first for normal development tasks;
Use Fugu first even for cross-file tasks;
If the result is not deep enough, consider Fugu Ultra;
For high-cost or high-risk tasks, do not rely only on the model answer. Use tests and human review.

What About Windows?

The official README says the one-line installation supports Ubuntu and macOS. Windows users, or users whose installation does not complete, should follow the official manual guide.

On Windows, consider:

Using WSL Ubuntu;
Installing on a macOS or Linux machine;
Checking Sakana’s official console or documentation for Windows manual steps;
Avoiding direct use of Linux/macOS shell commands in PowerShell.

In particular:

`1`	`curl -fsSL https://sakana.ai/fugu/install \| bash`

This is a bash command, not a native PowerShell command. If Windows does not have a bash environment, it will likely fail.

Common Issues

1. The curl command fails

Check network access first:

`1`	`curl -I https://sakana.ai/fugu/install`

If it cannot connect, the issue may be network access, proxy, certificate, or firewall configuration.

2. bash is not found

This usually means the current environment does not have bash, common in Windows PowerShell.

Possible solutions:

Switch to WSL;
Use macOS or Linux;
Follow the official manual installation guide.

3. codex-fugu is not found

Possible causes:

Installation did not succeed;
The command was not added to PATH;
The current terminal has not reloaded environment variables;
The installation script only affected the current shell.

Try reopening the terminal, then run:

`1`	`codex-fugu`

If it still fails, return to the official installation guide and check path configuration.

4. It is slower than a normal model

That is not necessarily abnormal.

Fugu coordinates multiple models behind the scenes, so complex tasks may involve more internal steps. It is useful when you trade speed for stronger complex-task handling, but it is not ideal for every lightweight question.

5. Can I upload company code?

Be careful.

The official page mentions that Fugu supports excluding specific providers or models based on data, privacy, and compliance requirements. Still, if you handle company code, customer data, security vulnerabilities, or patent materials, confirm your organization’s data policy and model-pool configuration first.

Usage Tips

For your first Fugu session, do not start with the largest and hardest task.

A safer flow is:

Test with a small repository;
Ask it to read the structure;
Give it a small bug;
Try a cross-file change;
Then use it for complex engineering tasks.

Make prompts specific, for example:

`1`	`Please read the related files first, then propose a change plan. After confirming the impact area, only modify the necessary files. When done, explain what you checked.`

These constraints make the result easier to control.

Summary

If you only want to understand what Fugu is, think of it as “a multi-agent system packaged as a model/API”.

But the part worth trying is its Codex entry point:

1
2

curl -fsSL https://sakana.ai/fugu/install | bash
codex-fugu

For everyday developers, Fugu is best used for code review, bug diagnosis, cross-file refactoring, complex documentation, and long-chain engineering tasks. Start with small tasks, then decide whether it belongs in your daily Codex workflow.

SakanaAI on KnightLi Blog

SakanaAI Fugu Guide: How to Enable a Multi-Agent Model in Codex

What Fugu Is Good For First

Check Before Installation

Install Fugu With One Command

Start codex-fugu

Suggested First Tests

Codex Tasks Suitable for Fugu

1. Code Review

2. Bug Diagnosis

3. Cross-File Refactoring

4. Documentation and Tutorial Writing

5. Hard Analysis Tasks

How to Choose Fugu or Fugu Ultra

What About Windows?

Common Issues

1. The curl command fails

2. bash is not found

3. codex-fugu is not found

4. It is slower than a normal model

5. Can I upload company code?

Usage Tips

Summary