Every FunAI model,
inside your agent
Images, video, audio, 3D and text as tools your assistant can call. Same account, same credits, same models you use in the studio.
- 1
Open Settings, then Connectors
In Claude Desktop or on claude.ai.
- 2
Add a custom connector
Paste https://mcp.funai.art as the server URL.
- 3
Sign in to FunAI
Claude opens FunAI to authorise it. No key to paste.
https://mcp.funai.artAsk in chat. Get the finished file.
Your assistant picks the model, spends your credits and waits for the render, so the result lands in the same reply — not a job id you have to chase.
Make me a square poster of a lone starship falling toward a black hole, lit only by the accretion disc.
aspect_ratio: "1:1"
model_id: "flux-pro-1.1"
Rendered in 14s. Balance updated.
Nine tools. Four are free to call.
Browsing models, checking a price and checking your balance cost nothing. Generating costs what it costs in the studio — billed once, by the same pipeline.
| generate_image | Make an image from a prompt | spends credits |
| generate_video | Make a video clip from a prompt | spends credits |
| generate_audio | Make speech, music or sound effects | spends credits |
| generate_3d | Make a 3D model | spends credits |
| generate_text | Run a text model | spends credits |
| list_models | See which models are available | free |
| estimate_cost | Check what a generation will cost first | free |
| get_credit_balance | Check how many credits are left | free |
| get_job_status | Check on a generation already running | free |
Everything the studio makes, your agent can make
The same catalogue behind funai.art, reachable as a tool call.
It cannot charge you twice
Handing an agent a spending tool is the part worth being careful about. Here is exactly what is guaranteed.
- Generations are never retried
- A request that landed but timed out on the way back would start a second billable job, so the generation tools do not retry. Read-only calls do.
- A retry replays, it does not re-bill
- Each generation runs at most once per request. A client asking again for a reply it never saw gets the original outcome back.
- Nothing auto-approves a spend
- Every tool that costs credits is marked as writing, so your client asks you before it runs one.
- It waits, then it degrades
- Long renders stream until they finish — 131.5s on the longest verified run. Past three minutes you get a job id instead of a failure.
Connect it in a minute
Create a key, paste the URL, restart your client.