MCP & Claude
Connect GoodTake to Claude using the Model Context Protocol
Overview
GoodTake exposes an MCP (Model Context Protocol) server at https://mcp.goodtake.ai/mcp. When connected, Claude can call GoodTake tools directly — no copy-pasting API keys, no leaving the conversation.
There are three ways to connect:
| Method | Best for | Effort |
|---|---|---|
| CLI skill | Claude Code — easiest way to start | npx skills add + done |
| Claude.ai Connectors | Claude.ai web app | ~1 min in Settings |
| Claude Code OAuth | Claude Code CLI, desktop, or IDE | OAuth browser flow |
The CLI skill is the fastest path if you already have gt installed — no OAuth, no settings UI, just one command. The Connectors and Claude Code OAuth methods go through the MCP server and use OAuth 2.1 + PKCE to mint a gt_ API key scoped to your active organization; the key appears in Settings → Developer Settings like any other key and can be revoked there.
MCP tools
Once connected, Claude has access to:
| Tool | Description |
|---|---|
generate_image | Submit an image generation and return signed image URLs (or the job ID if the connection times out) |
generate_video | Submit a video generation and return a signed video URL. Slower models finish after the call returns — see below |
upload_image | Open an upload widget so you can send images from your own device into the conversation |
get_generation | Retrieve status and signed asset URLs for an existing generation by ID |
list_generations | List your organization's recent generations, image and video |
list_models | List available models with their credit costs — pass type: "video" for video models |
get_balance | Show your organization's current credit balance |
Both generation tools wait up to 5 minutes for a result. generate_video names its model explicitly — ask Claude to list the video models first if you are not sure which to use.
Using your own images
generate_image takes reference images through image_urls, and generate_video takes a single source image through image_url. Any publicly reachable https URL works — and for images on your own machine, upload_image is the way in. Claude cannot read chat attachments or open files on your computer, so asking it to "use the photo I just attached" will not work on its own.
Just mention the image and Claude will open the widget:
"Here's a photo of my dog — turn it into an oil painting"
"Use these three product shots as references for a lifestyle image"
What happens next:
- An upload area appears in the conversation. Drag files onto it, or click to pick them.
- Each image uploads and shows Ready. Remove one with the ✕ in its corner.
- Press Continue. A message is prefilled in your composer containing the image links.
- Read it, then press send. Claude passes the links straight into the generation.
Step 4 is yours to make — Claude cannot send that message for you, which is what keeps a page in an iframe from putting words in your mouth.
You can request up to 14 images at once. The upload links are valid for 15 minutes; once uploaded, the image URLs themselves work for 7 days, so you can reuse them later in the same conversation.
The widget renders in Claude.ai on the web and in the macOS desktop app. On
surfaces that don't render it, upload_image returns a curl command per
slot so you can upload by hand — the rest of the flow is identical.
Prompting Claude
Once connected you can ask naturally:
"Generate a 16:9 product shot of white sneakers on a marble surface"
"Create three variations of a sci-fi cityscape using the seedream model"
"Get the status of generation 550e8400-…"
"What video models can I use, and what do they cost?"
"Animate this photo into a 5-second clip"
"How many credits do I have left?"
Also: Claude Code skill
If you prefer a lighter integration without setting up the MCP server, the goodtake-generate skill wraps the gt CLI instead of calling MCP directly.