If you searched for a Claude MCP GPT Image 2.0 API workflow, the same three routes now run GPT Image 2.5 Flare at 3 credits: connect Claude Imagine's hosted MCP with OAuth, install a local Claude GPT Image 2.5 Skill with one Claude Imagine key, or call the HTTP endpoint directly. Claude plans the request while the external model renders the pixels.
This hosted path does not require a separate OpenAI API key or billing account.
Independent service · 3 credits per image · 9 signup credits · No card required
Live product
This is the real Claude Imagine generator, preselected to GPT Image 2.5 text-to-image. Enter a prompt, choose one of five aspect ratios, and generate with the same account balance used by MCP and the Claude Code Skill. The preview gallery offers prompt starters; your completed request returns its own model-generated image.
Generate multiple images per click on Pro and Studio. Upgrade
Tips for better results
Made with Claude Imagine — click a sample to try its prompt. Your result will appear here.
Three working paths
Choose the route that matches where the output needs to land. Remote MCP is the shortest setup and keeps credentials out of your shell. The Skill saves the returned image into a local project. The HTTP API is the smallest building block for your own automation. All three call the same text-to-image service and share one Claude Imagine credit balance.
OAuth MCP
claude mcp add --transport http claude-imagine https://claudeimagine.com/api/mcpRun the command once. Claude Code discovers OAuth 2.1, opens the browser login and consent screen, and registers generate_image, edit_image, generate_video, check_generation, quote_generation, list_generations, list_models, and get_credits. You paste no token into the MCP configuration.
Claude Code Skill
git clone https://github.com/lumian2015/nano-banana-claude-skill .claude/skills/nano-banana-claude
export CLAUDEIMAGINE_API_KEY="sk-..."
python3 .claude/skills/nano-banana-claude/scripts/generate.py "editorial product photo of a lime desk lamp" --model gpt-image-2.5 --aspect 16:9 --out hero.pngThe public MIT Skill uses only Python's standard library. Create a Claude Imagine key once, then ask Claude Code for an image or call the script directly. The --out flag makes the destination explicit.
HTTP API
curl https://claudeimagine.com/api/v1/images/generations \
-H "Authorization: Bearer $CLAUDEIMAGINE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"gpt-image-2.5","prompt":"editorial product photo of a lime desk lamp","aspect_ratio":"16:9"}'POST a prompt, optional model, and optional aspect_ratio. A successful response contains image URLs, the friendly model id, and credits_used. The endpoint accepts text-to-image requests only.
Claude does not render images itself, so using GPT Image 2.5 in Claude means connecting Claude Imagine once and then asking for images in plain language. Every route below charges 3 credits per 1K image and runs the same GPT Image 2.5 Flare model.
In Claude Code run claude mcp add --transport http claude-imagine https://claudeimagine.com/api/mcp and sign in when the browser opens. In Claude Desktop add the same URL as a connector. No OpenAI key is involved; the 9 signup credits are enough for three GPT Image 2.5 images.
Say something like "generate a 16:9 hero image of a lime desk lamp with gpt-image-2.5". Claude calls generate_image with model gpt-image-2.5 and the aspect ratio you asked for; leave the model out and it uses Nano Banana 2 instead.
Reply with the change you want, such as "make the lamp red, keep everything else". Claude passes the previous image URL to edit_image with model gpt-image-2.5, which keeps the original framing and returns a new image. Old model id gpt-image-2 still works as an alias.
All three paths reach the same model through Claude Imagine, but authentication, client support, and output handling differ. Pick the route that matches where the image must appear.
Use the remote connector when the image should appear in Claude Desktop, claude.ai, or Claude Code. OAuth links your Claude Imagine account without exposing an API key, and Claude can check models and credits before it generates.
Use the Skill when Claude Code should write a PNG directly into a repository. The script accepts a model, aspect ratio, prompt, and output path, making it useful for web assets and documentation.
Use the API when another script, agent, or backend owns the workflow. A bearer key authenticates the request, and the JSON response reports the exact model and credits consumed.
Use the browser when you want a visible model selector, prompt starters, batch controls, task progress, downloads, and the image-to-image mode that the current MCP and public API do not expose.
MCP, Skill, API, and web generations deduct from the same Claude Imagine credit ledger. A subscription grants recurring credits; a top-up pack adds one-time credits without changing your plan tier.
When you request gpt-image-2.5, the service maps that friendly id to the GPT Image 2.5 text-to-image implementation. If a request cannot start, it returns an error instead of quietly choosing another model.
Claude remains the orchestrator and Claude Imagine remains the image service. Keeping those roles separate makes authentication, billing, and troubleshooting easier to understand.
Tell Claude the subject, visual style, framing, lighting, aspect ratio, exact text, and constraints. Claude can turn that brief into the final prompt before it calls an external tool.
MCP accepts model gpt-image-2.5, the Skill uses --model gpt-image-2.5, and the API uses the same friendly id. The live generator exposes GPT Image 2.5 directly in its selector.
The service validates the prompt, account balance, concurrency limit, and aspect ratio, then creates a tracked image task. Two credits are deducted atomically for each started GPT Image 2.5 generation.
MCP returns the image URL to Claude, the API returns JSON, the Skill downloads the file to your chosen path, and the web generator displays a downloadable result in the task panel.
Claude Imagine charges a predictable credit amount instead of passing through a separate OpenAI invoice. These values come from the live model catalog and billing code checked on August 24, 2026.
A GPT Image 2.5 text-to-image request costs 2 Claude Imagine credits on free, Basic, Pro, and Studio accounts. Batch generation multiplies that amount by the number of images requested.
A new account starts with 9 credits and needs no payment card. That is enough for three GPT Image 2.5 generations, or one Nano Banana 2 image with four credits left over.
Monthly or annual subscriptions grant the plan's recurring credit allowance and plan benefits. Basic includes 450 monthly credits, Pro 1,300, and Studio 4,800 at the current monthly prices.
A credit pack adds credits without creating a recurring subscription. It does not unlock Pro-only models because a pack buyer remains on the free plan tier, but GPT Image 2.5 itself is available.
The hosted MCP, public Skill, and direct API accept 1:1, 4:3, 3:4, 16:9, and 9:16. Unsupported values fail validation before an image task is created.
The live limits are one active generation on free accounts, two on Basic, three on Pro, and eight on Studio. A concurrency error does not silently queue an unlimited number of jobs.
Clear boundaries matter more than feature inflation. The current public integrations deliberately expose a smaller, predictable surface than the web application.
Claude writes or refines the prompt and invokes a connected tool. GPT Image 2.5 performs the image generation through Claude Imagine, so this is connected image generation rather than a native Claude raster model.
The public Skill script and the /api/v1/images/generations endpoint accept a text prompt only. The MCP connector also has edit_image, which takes a public image URL rather than an upload.
The Claude Imagine web generator exposes a GPT Image 2.5 image-to-image mode. Use that browser workflow when you need a reference image; do not assume the same input is available in MCP or the API.
The HTTP endpoint polls the task for up to 90 seconds. If generation is still running, it returns a 504 response that points you to Activity; the tracked task can still complete after the request stops waiting.
When the image task reports a failed state, deducted credits are refunded automatically. A completed upstream task that returns no image URL is a separate support case and is reported explicitly.
Claude Imagine is an independent service and is not affiliated with, endorsed by, or sponsored by Anthropic or OpenAI. Claude and GPT Image are trademarks of their respective owners.
Direct answers about hosted connectors, local skills, API keys, Claude Desktop, pricing, and output limits.
Need the raw endpoint instead? See the Claude Image Generation API quickstart.
Checked September 12, 2026. OpenAI documents GPT-Image-2.5 Flare generation and editing capabilities; Anthropic documents Claude's image understanding and remote MCP connectors. Claude Imagine product claims above are verified against the deployed code.
Connect with OAuth for the shortest path, or create one Claude Imagine key when the output needs to land directly in code. Your 9 signup credits are enough for three GPT Image 2.5 generations.