✨ feat: Added Support for Flagship Gemini Image (🍌 Nano Banana) models for Image generation and editing

[devilb2103]

Added Gemini Flash Image Generation Tool (Nano Banana + Nano Banana Pro)

trim.6A4F5154-48D1-4547-ABEF-C3A188733218.MOV

Summary

This PR (addresses #9283 and #6065) implements a new image generation tool using Google's Gemini 2.5 Flash Image Model with Vertex AI integration. The tool supports both text-only image generation and image context-aware generation for editing/modification requests. It includes handling of rejected tool calls due to content policy violoations, multiple file storage strategies, and seamless integration with LibreChat's existing agent tool ecosystem.

Key Features:

• Text-to-image generation using Gemini 2.5 Flash Image Model
• Image context support for editing existing images
• Multiple storage strategies (local, S3, Azure, Firebase)
• Safety filtering with user-friendly error messages

Configuration Requirements:

• Google Cloud service account JSON file required at api/data/auth.json
• Docker volume mapping: ./api/data/auth.json:/app/api/data/auth.json
• Vertex AI API permissions for the service account
• LibreChat YAML configuration update required

Change Type

• New feature (non-breaking change which adds functionality)

Testing

The feature has been tested for:

• ESLint compliance (all files pass)
• Tool registration and integration
• Docker container compatibility
• File path resolution in containerized environment

Test Configuration:

• Environment: Docker containerized deployment
• Dependencies: @google/genai: ^1.17.0 added to api/package.json
• Authentication: Google Cloud service account with Vertex AI permissions
• Storage: Local and S3 file strategy tested (other strategies compatible - follow Open AI Image Tools Logic)

Test Steps:

1. Build Docker container with new dependencies
2. Configure Google service account credentials
3. Enable gemini_image_gen tool in an agent
4. Test text-to-image generation
5. Test image context-aware generation
6. Verify safety filtering functionality
7. Verify if generated Images show up in attachments drop down in the collapsable LLM Tools section at the right hand side of UI

Checklist

• My code adheres to this project's style guidelines
• I have performed a self-review of my own code
• I have commented in any complex areas of my code
• My changes do not introduce new warnings
• Local unit tests pass with my changes
• Any changes dependent on mine have been merged and published in downstream modules.

Files Modified

Core Implementation

• api/app/clients/tools/structured/GeminiImageGen.js - Main tool implementation
• api/package.json - Added @google/genai dependency
• packages/data-provider/src/config.ts - Added to imageGenTools set

Integration

• api/app/clients/tools/index.js - Tool export
• api/app/clients/tools/manifest.json - Tool manifest entry
• api/app/clients/tools/util/handleTools.js - Tool registration and configuration
• api/server/services/ToolService.js - Import statement

Documentation

• api/app/clients/tools/structured/README-GeminiNanoBanana.md - Implementation guide

Breaking Changes

None. This is a purely additive feature that doesn't modify existing functionality.

Configuration Notes

Required Docker Configuration:

volumes:
  - ./api/data/auth.json:/app/api/data/auth.json

LibreChat YAML Configuration:
Add gemini_image_gen to the includedTools array in your librechat.yaml:

includedTools: ['gemini_image_gen']

Environment Variables

Add these to your .env file to configure the Gemini Image Generation tool:

Variable	Description	Default	Required
GEMINI_IMAGE_PROVIDER	Provider to use: vertex (service account) or gemini (API key)	Auto-detected	No
GEMINI_IMAGE_MODEL	Model ID for image generation	gemini-2.5-flash-image-preview	No
GEMINI_API_KEY	Gemini API key (required if using gemini provider)	-	If provider=gemini
GOOGLE_SERVICE_KEY_FILE	Path to service account JSON (for Vertex AI)	api/data/auth.json	If provider=vertex
GOOGLE_CLOUD_LOCATION	Google Cloud region for Vertex AI	global	No
GOOGLE_APPLICATION_CREDENTIALS	Vertex AI Service account credentials path	None	Yes

Note: If GEMINI_IMAGE_PROVIDER is not set, the tool auto-detects:

• Uses Vertex AI if GOOGLE_SERVICE_KEY_FILE exists or api/data/auth.json is present
• Uses Gemini API if GEMINI_API_KEY is set

Note: Service Account Credentials must be manually set in .env in the following manner:

GOOGLE_APPLICATION_CREDENTIALS="data/auth.json"

Google Cloud Setup:

1. Create a service account with Vertex AI permissions
2. Download the service account key as JSON
3. Place the file at api/data/auth.json
4. Ensure the project ID matches your Google Cloud project

[falkenbt]

Really nice, I got this to work on my machine. There is just a small issue with the credentials path, see inline comment.

[devilb2103]

@falkenbt The issues are addressed. Thank you for the comments

[lukaswelte]

It might only be my local configuration, but for me it says the image is generated, but it then does not show. I also can't find it anywhere on the local disk (using local storage strategy)

Do you have good pointers how I could further debug?

Thanks a lot for contributing this btw. - eagerly awaiting it :)

[petervcook]

A future related feature request after this ships: Nano Banana is supported by OpenRouter and it would be great to support Nano-Banana via OpenRouter in addition to Vertex API.

[devilb2103]

It might only be my local configuration, but for me it says the image is generated, but it then does not show. I also can't find it anywhere on the local disk (using local storage strategy)

Do you have good pointers how I could further debug?

Thanks a lot for contributing this btw. - eagerly awaiting it :)

Could you please share the logs that appear in the librechat container when you try to generate an image?

[lukaswelte]

@devilb2103 thanks for that trivial hint - sorry I hadn't checked that 🤦
Error was in IAM permissions on GCP - problem solved, it works like a charm 🚀

[falkenbt]

Looks good to me now, thanks a lot. I cannot approve though, I'm just a user without elevated access.

[falkenbt]

@danny-avila any chance this could make it into the 0.80 release? This is really useful for a lot of people.

[devilb2103]

hey @danny-avila do let me know what can be improved here so that this contribution can be merged without much efforts from the team 😄

[Dual-0]

Why is Vertex AI API used here? This also works with Gemini API

curl -s -X POST
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "parts": [
        {"text": "Create a picture of a nano banana dish in a fancy restaurant with a Gemini theme"}
      ]
    }]
  }' \
  | grep -o '"data": "[^"]*"' \
  | cut -d'"' -f4 \
  | base64 --decode > gemini-native-image.png

[devilb2103]

Why is Vertex AI API used here? This also works with Gemini API

curl -s -X POST

  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image-preview:generateContent" \

  -H "x-goog-api-key: $GEMINI_API_KEY" \

  -H "Content-Type: application/json" \

  -d '{

    "contents": [{

      "parts": [

        {"text": "Create a picture of a nano banana dish in a fancy restaurant with a Gemini theme"}

      ]

    }]

  }' \

  | grep -o '"data": "[^"]*"' \

  | cut -d'"' -f4 \

  | base64 --decode > gemini-native-image.png

It should be fairly easy to have another librechat manifest var to switch between the already supported VertexAI option with GeminiAPI

This implementation makes it convenient for those who have access to Googles VertexAI platform

[falkenbt]

@danny-avila is this on your radar? It would be great to have an indication when this could be merged and released, from the reactions here and my surroundings a lot of people are eagerly awaiting this. Thanks!

[djp1971]

We Would love to see this in the next release. i have a ton of users calling for it. tried making a action for this usinga. OpeAPI spec and it fails as it passes back base64 in a format that it cant decode. so this woudl be great as a tool.

[marlonka]

@danny-avila hate to annoy you but once again asking if you could accept this PR into main plz since there is huge demand for nano banana in our enterprise

[raymond9zhou]

@danny-avila Just wanted to reiterate the importance of this feature. Adding native support for Gemini 2.5 is a major value-add for the platform. The PR is stable, tested, and ready to go. Merging this now would greatly benefit the user base who are waiting to utilize these image generation tools. Hoping to see this live soon!

[ZenDevMaster]

It must be discouraging for contributors to submit pull requests and then be ignored.

[RepLicanT-UHD]

@devilb2103 thanks for that trivial hint - sorry I hadn't checked that 🤦 Error was in IAM permissions on GCP - problem solved, it works like a charm 🚀

@lukaswelte Could you please tell the exact problem with IAM Permissions? I have Vertex AI Admin role and I'm still experiencing the same problem.

@devilb2103 Thank you for the great job! Just like to ask if this normal or not:

warn: [GeminiImageGen] Missing required parameters for storage, falling back to data URL
warn: [GeminiImageGen] Could not save to storage, using data URL

[lukaswelte]

@RepLicanT-UHD in my case I literally didn't have the role on the service account that I gave to librechat. Are you sure it picks up the service-account.json in your case?

[RepLicanT-UHD]

@RepLicanT-UHD in my case I literally didn't have the role on the service account that I gave to librechat. Are you sure it picks up the service-account.json in your case?

Thanks for your response! Yes it does. When I select any model other than 2.5 Flash for the Agent to execute requests via the 'gemini_image_gen' tool, the errors disappear (surprisingly, o4-mini handles this role particularly well). The only warnings remaining are warn: [GeminiImageGen] Missing required parameters for storage, falling back to data URL and warn: [GeminiImageGen] Could not save to storage, using data URL. Meanwhile, my File Strategy is set to local, and images are displayed and edited correctly.

However, as soon as I switch back to 2.5 Flash with any combination of reasoning and input/output tokens, it claims inside the thought process that the image was generated and shown to the user, even though that's not the case.

[marlonka]

Does it support Gemini 3 Pro Image (aka Nano Banana Pro) already? Would be awesome for admins to choose in env or librechat.yaml the model or have a dropdown in the tool UI: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro-image

[Copilot]

Pull request overview

This PR adds a new image generation tool that integrates Google's Gemini 2.5 Flash Image Model with Vertex AI, supporting both text-to-image generation and image context-aware editing. The implementation follows existing LibreChat patterns for image generation tools and includes comprehensive safety filtering and multi-storage strategy support.

Key Changes

• Added Gemini image generation tool with support for text prompts and image context editing
• Integrated with existing LibreChat file storage strategies (local, S3, Azure, Firebase)
• Implemented safety filtering with user-friendly error messages for content policy violations

Reviewed changes

Copilot reviewed 8 out of 10 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
api/app/clients/tools/structured/GeminiImageGen.js	Core implementation of Gemini image generation tool with image context support and safety filtering
api/package.json	Added @google/genai dependency for Vertex AI integration
package-lock.json	Locked dependency versions including @google/[email protected] and unrelated dicebear changes
api/app/clients/tools/index.js	Exported GeminiImageGen tool for registration
api/app/clients/tools/manifest.json	Added Gemini Image Tools manifest entry
api/app/clients/tools/util/handleTools.js	Registered tool constructor with custom initialization logic and unrelated serpapi configuration
api/server/services/ToolService.js	Added commented-out import (cleanup needed)
packages/data-provider/src/config.ts	Added gemini_image_gen to image generation tools set
api/app/clients/tools/structured/README-GeminiNanoBanana.md	Comprehensive documentation for setup and usage

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review for a chance to win a $100 gift card. Take the survey.

[danny-avila]

I understand the frustration with review delays. As maintainers, we need to balance responsiveness with ensuring thorough reviews, since incomplete features can create significant long-term maintenance burden. I'm reviewing this PR now and will provide detailed feedback shortly.

To start, please see the GitHub Copilot reviews, which are actually worth considering.

I will add a configurable yaml key for this soon.

This is a "must implement" for merging and corresponding documentation is required as well in https://github.com/LibreChat-AI/librechat.ai

[devilb2103]

@danny-avila Thanks for the comments ✌️.

[usnavy13]

@danny-avila Thanks for the comments ✌️. I plan to implement said changes after a couple of weeks

I can take a shot at it this week if you add me

[usnavy13]

@devilb2103 @danny-avila
I implemented the suggested changes in #10676 since I couldn't commit to his branch.
I also added geminiAPI support and some other small improvements.

[Copilot]

Pull request overview

Copilot reviewed 15 out of 17 changed files in this pull request and generated 5 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review for a chance to win a $100 gift card. Take the survey.