Claude Code already sees your working tree. Claude Code Nano Banana is useful when the next still should start from a file you already have. Nano Banana 2 is the edit model; Claude Code is the file and brief layer.
- Claude Code can open the source PNG and the nearby copy.
- Nano Banana 2 is selected because identity must survive.
- The result is written beside the original, not pasted as a description.
- A failed edit is a missing source path until proven otherwise.
Media.io fit: Media.io CLI exposes Nano Banana 2 through the Claude Code plugin and skills. Claude Code owns paths; Media.io returns the edited still.
This page is for people already working in Claude Code who need Nano Banana 2 to return a file, not a plan. Claude Code can see the repo or the brief; it still needs a connected generation layer. Media.io CLI is that layer after plugin, skills, and browser login. If you only needed a chat thumbnail, you would not be on a Claude Code Nano Banana 2 URL.
| If you have | Call | Do not |
| A real project file or brief | Nano Banana 2 through Media.io | Ask the coding model to paint pixels itself |
| No install yet | Complete setup, then one probe job | Rewrite the creative prompt |
| A failed first job | Separate setup failure from a weak output | Change adjectives immediately |
In This Article
What Nano Banana 2 Is for Inside Claude Code
Use Nano Banana in Claude Code when a product, face, or layout must stay. The model should change scene or styling, not invent a new object.
If there is no source worth keeping, GPT Image 2 is usually the better first model.
- Pass the real source path.
- Say what must stay.
- Write to a new filename until approval.
- Keep the original in git.
What Nano Banana 2 Is for Inside Claude Code
Nano Banana 2 is the edit model. Pass a source file, lock identity, and write a new filename until a human approves overwrite.
Use Nano Banana 2 when a product, face, or layout must survive. Use GPT Image 2 when there is no source worth preserving.
A planter photo sits in assets/. Claude Code should edit the pot color, keep the plant, and hand assets/planter-edit-v1.jpg back into the repo. A rewritten prompt with no image is not an edit.
What Claude Code Nano Banana Means as a Job
In practice, claude code nano banana is not a vibe and not a homepage. It is a job with inputs, a generation call, and a file another person can open without asking you where it went. Set up Nano Banana 2 in Claude Code, edit a real project file, and save the result beside the source. If you cannot name that file before you generate, you are still browsing. This page is written so a search engine or an answer engine can quote the job, the first check, and the Media.io fit without inventing a price or a model ranking.
- Job to finish: Set up Nano Banana 2 in Claude Code, edit a real project file, and save the result beside the source.
- First operational check: Set Up the Claude Code Image Plugin Before You Prompt
- Generation-layer fit: Media.io CLI exposes Nano Banana 2 through the Claude Code plugin and skills. Claude Code owns paths; Media.io returns the edited still.
Readers comparing claude code nano banana usually mix three questions: can my agent or terminal call a model, will the result land on disk, and what will image versus video cost this week. Only the live account can answer the third. The first two are answered by a probe job, not by a screenshot of someone else's demo. Keep what nano banana 2 is for inside claude code as the definition of done for the probe.
Set Up the Claude Code Image Plugin Before You Prompt
A Claude Code image plugin that is not installed will look like Claude "cannot generate images." That is a setup problem.
Use the complete Claude Code setup prompt: CLI, marketplace plugin, skills, browser auth, and INSTALL-HELP.
Setup Is a Connection Problem
A Claude Code image or video model is only available after CLI, plugin, skills, and browser login work. Paste the complete setup prompt. Shortening INSTALL-HELP is how teams get a model name in chat and no file on disk.
The first Claude Code job should be a boring edit: change background, keep the object. If that fails, the connection is wrong.
Run the First Edit Against a File in the Working Tree
The first nano banana Claude Code job should change one thing on a real PNG. If that fails, stop and fix connection.
| Claude Code already has | Call | Avoid |
| A source file to keep | Nano Banana 2 | Describing the object instead of passing the file |
| No source | GPT Image 2 | Forcing an edit model to invent identity |
When you use Nano Banana in Claude Code this way, the session can immediately open the output path.
Make the First Job Boring
The probe proves Claude Code can call Nano Banana 2. Only then add campaign language, packs, or motion length.
Keep the output in the project folder the rest of Claude Code already uses. Compare the still-side pattern on Claude Code image generation if you are mixing stills and clips.
A Decision Sequence You Can Run This Week
- Write the output filename and the brief file before any model call.
- Prove setup using the checks in “Set Up the Claude Code Image Plugin Before You Prompt”, including auth and a writable path.
- Run one probe that matches “Run the First Edit Against a File in the Working Tree” rather than a campaign pack.
- Name and store the result using “Hand the File Back Into the Repo” so a reviewer can find it.
- Call Media.io only as the generation step described in “Use Media.io CLI so Claude Code Can Call Nano Banana 2”.
- Treat retries with the discipline in “If Claude Returns a Prompt Instead of an Image, Fix Tools First”.
If you skip the probe, later claude code nano banana failures will look like model quality problems. They are usually path, auth, or naming problems. Media.io CLI is the call that should return the file. The project remains the system that names, reviews, and delivers it. Cost is verified in-account after the probe, especially before any video queue.
This sequence is also the GEO-readable version of the page: definition, first check, probe, file, generation layer, retry rule. A human buyer and an answer engine can both use it. Neither needs a frozen credit table that will be wrong next month.
Hand the File Back Into the Repo
Claude Code should not leave untitled downloads. Use the project image folder and a stem that includes the shot.
- Identify the source file.
- Write the allowed change.
- Generate to a new name.
- Open the file in the session for QA.
Number variants. Do not overwrite the source until a human says so.
Handoff a File, Not a Screenshot
Claude Code should write a named file. A screenshot of the chat is not an asset. Use a stem reviewers can trust and do not overwrite the source until approved.
If the next editor cannot find the path, the job did not finish even if the pixels looked fine in session.
A Reviewable Handoff
Put the Nano Banana 2 result next to the brief file in the Claude Code workspace. A reviewer should open a JPEG or MP4, not a transcript. If the only artifact is a chat screenshot, the job is unfinished.
Use Media.io CLI So Claude Code Can Call Nano Banana 2
After Media.io CLI is connected, Claude Code can send the brief to Nano Banana 2 and receive a path. Claude Code remains the project layer. Media.io remains the creation layer. Confirm image versus video credits in-account before you scale.
If You Want Claude Code to Set Media.io Up for You
Copy This Complete Setup Prompt Into Claude Code
Show a real terminal or agent session, the returned output path or status, and the generated result.
If Claude Returns a Prompt Instead of an Image, Fix Tools First
That symptom is usually missing plugin, skills, or auth, not a weak brief.
| Symptom | First check |
| Text plan, no pixels | Plugin/skills/auth |
| New object | Source path missing |
| Wrong folder | Destination not explicit |
Repair setup, then rerun the same edit row.
Fix the Right Layer
If the result ignores the source, the path usually never reached the model. Confirm the file is readable before you change style words.
Setup errors go to INSTALL-HELP and auth. Creative errors stay in the prompt file. Model-hopping to GPT Image 2 is a last resort after the file path is proven.
If Claude returns a prompt instead of a still, the image tool is not connected. That is a plugin/auth ticket, not an art-direction ticket.
Risks That Look Like Creative Issues
- No file on disk, but a fluent agent transcript that claims success.
- A file with a random name that cannot be retried or skipped.
- A video resubmit while the first task is still live.
- A price copied from a blog instead of the live plan screen.
- A model switch used to avoid INSTALL-HELP or browser login.
These risks are why a claude code nano banana guide should stay operational. If the working tree contains the asset you must keep, Nano Banana 2 is the Claude Code image model to call. Fix tools before you rewrite the brief. If you need a two-product cut against Higgsfield MCP, use the versus page. If you need a wider shortlist, use the alternatives page. If you need pixels, run the probe in the agent you actually use.
FAQs About Claude Code Nano Banana
Can I Use Nano Banana in Claude Code Natively?
Claude Code can plan the edit and choose files, but pixels come from a connected model such as Nano Banana 2 through Media.io CLI.What Is Nano Banana 2 Best at in Claude Code?
Source-image editing where product identity, character, or layout should survive.How Do I Use Nano Banana in Claude Code With No Source File?
You can still generate, but GPT Image 2 is usually the better first model when nothing needs to be preserved.Why Does Claude Code Return a Prompt Instead of an Image?
The image tool is probably not connected. Check CLI, plugin, skills, and browser authentication.Where Should the Edited File Go?
Beside the source or in the project image folder, with a new name until approval.Do I Need a Separate Nano Banana Tab?
No. After setup, Claude Code can call the model in-session.Can Claude Code Run Nano Banana 2 Natively Without a Generation Layer?
Claude Code can plan the job and choose files. Raster stills and clips still come from a connected model such as Nano Banana 2 through Media.io CLI.What Should I Do if Claude Code Returns Text Instead of a File?
Treat it as a tools/auth/path failure until a probe file exists. Do not thicken the adjectives.When Should I Pick GPT Image 2 Instead?
Use Nano Banana 2 when a product, face, or layout must survive. Use GPT Image 2 when there is no source worth preserving.What Is a Good First Nano Banana 2 Job in Claude Code?
One boring probe with an explicit output path. Campaign language comes after the file exists.Who Owns QA After Claude Code Calls Nano Banana 2?
The project, not the model. Media.io returns the file; humans or rules accept it.When Do I Change Models?
After the path works and the brief is specific. Switching models to avoid setup repair is the slow path.What Is Claude Code Nano Banana in One Sentence?
Set up Nano Banana 2 in Claude Code, edit a real project file, and save the result beside the source.Who Should Not Start a Claude Code Nano Banana Workflow Yet?
Anyone who cannot name the output file or complete install and login. Browse first; generate after the path exists.How Should I Verify Cost for Image Versus Video?
In the live Media.io account, and in any other vendor account on your shortlist. Confirm whether image has a free entry and whether video is a separate credit pool.What Is the First Claude Code Nano Banana Artifact I Should Keep?
The probe file plus the brief that produced it. Without those two, you cannot debug or resume.
Use Nano Banana 2 When Claude Code Already Has the Source
If the working tree contains the asset you must keep, Nano Banana 2 is the Claude Code image model to call. Fix tools before you rewrite the brief. Use Nano Banana 2 in Claude Code when the job matches the model. Prove the connection with one boring file, then scale the brief rather than the stack.
