For agents
Your agent can run 50 image and video models on your Bildy credits
Bildy has an MCP endpoint, and MCP is the protocol Claude, Cursor and most other agents use to pick up tools. You paste one address into your client, sign in with the Bildy account you already have, and your agent picks up the image and video tools. Everything it makes comes out of the same credit balance you spend in the browser.
The address you paste into your client
https://bildy.ai/mcpPaste it without the www. A client that registers www.bildy.ai gets a redirect to this address, and the sign-in then fails, because the two names do not match when your client checks them.
A browser gets nothing back from that address, because it only answers the kind of request an agent sends. The Notes memory tools live at a separate address, https://bildy.ai/api/mcp/notes, so a client you connect for picture work picks up the media tools and nothing else.
What MCP is, if you have not run into it yet
An agent on its own can only write text back to you. MCP is the standard way to hand it tools it can actually call, and a server like this one is a list of tools plus the code behind them. Anthropic published it in late 2024, and Claude, Cursor, Windsurf, Zed and a long list of others read it now.
Once Bildy is connected, you type "make me a square photo of a candle on a marble counter" at your agent, and it calls images_generate and hands you back the picture once the job finishes. You never open bildy.ai for it.
Connecting takes one address and a sign-in
Claude on the web, and the Claude desktop app
Open Settings, then Connectors, then Add custom connector, and paste the address. Claude opens a browser window, you sign in with your Bildy account, and the tools appear. You never touch an API key.
Claude Code, in a terminal
Run this once and sign in when it asks.
claude mcp add --transport http bildy https://bildy.ai/mcpCursor, Windsurf, Zed and the rest
Add an MCP server of type http with that same address. Some clients ask for a key instead of opening a browser, and those take a Bildy API key as a bearer token. You can make one on the API docs page.
An agent you wrote yourself
It is plain JSON-RPC over HTTP POST with a bearer token, so you can also call it with curl to check the connection before you wire anything up.
curl -s -X POST https://bildy.ai/mcp \
-H 'Authorization: Bearer bld_YOUR_KEY' \
-H 'Content-Type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'The 16 tools your agent picks up
| images_generate | Makes an image from a prompt, on whichever model you name. |
| images_edit | Changes pictures you hand it, following a prompt. |
| images_get_job | Reports on an image job and gives you the download link once it is done. |
| images_list_models | The image models you can name, and what one image costs on each. |
| videos_generate | Makes a clip from a prompt, or animates a picture you supply. |
| videos_get_job | Reports on a video job and gives you the download link once it is done. |
| videos_list_models | The video models, the clip lengths each allows and what they cost. |
| videos_extract_frame | Takes one frame out of a clip as a picture, usually the last one. |
| videos_compose | Joins pictures and clips end to end, each picture held for as long as you say. |
| videos_trim | Cuts a clip down to one stretch of it. |
| videos_add_audio | Lays a sound file over a clip. |
| videos_render_timeline | Cuts up to a minute of video from a list of shots, with transitions, a colour grade and lyrics. |
| videos_join | Puts rendered parts end to end with a song under them, for a whole music video. |
| videos_get_edit | Reports on an editing job and gives you the download link. |
| files_upload | Gives your agent a link to upload a photo, clip or song from your computer, free. |
| credits_balance | How many credits are left, so an agent can price a run before it starts. |
Images and clips both come back as a job id your agent polls, because a picture takes seconds and a clip takes minutes. Download links are good for an hour.
Chaining clips together, without leaving the agent
A video model gives you one shot at a time, so a thirty second video is five generations that have to cut together. The editing tools are the other half of that. Your agent takes the last frame of one clip with videos_extract_frame, hands it to videos_generate as the first frame of the next one, and the two clips carry on from each other instead of jumping. Then videos_compose joins the run into one file, with stills held for as long as you say between them.
For a music video your agent goes a step further. It lists every shot with how long it stays on screen, where it starts in its clip and how it comes in, and videos_render_timeline cuts up to 60 seconds of that with a colour grade and the lyrics drawn in. Your agent renders a song in parts of that length, and videos_join puts the parts together with the song under them.
No model runs for any of the editing, so it costs 5 to 50 credits a call rather than a few hundred. Your agent passes the link a finished job gave it straight back in, so a clip it just made never has to come down and go back up again.
What your agent spends, and how you cap it
Every call comes out of your credit balance at the price you pay in the browser. A 5,000 credit top-up costs €5.00, so a thousand credits are one euro and you can read any credit price as money by moving the decimal point three places. The cheapest image model here is Z-Image Turbo at 8credits, about €0.005 a picture. Video starts at a few hundred credits a second, and videos_list_models gives your agent the figure before it generates anything.
An agent that runs away can spend what you have loaded and nothing beyond it, so your balance is the cap. You can revoke a key you regret on the API docs page. The price of every model is on the models page, covering all 28 image models and 22 video models.
Skills that stop your agent spending more than a job needs
An agent with the tools connected will generate a ten second clip when a two second test would have answered the question. In the 6 skills below we tell it to read the price list first, and to make cheap stills before it spends a video budget.
Product photos from one picture
A photo you took on a desk becomes a set of clean shots for a shop listing.
Ad variants in every placement size
Your agent composes each crop separately, with room left for the headline.
A clip from a picture you already have
A still becomes frame one and the model carries on from there.
The same prompt on several models at once
Your agent runs one prompt on four models and lays the results out with what each cost.
A video longer than one generation
Each clip starts on the frame the last one ended on, so the run carries on instead of jumping.
A script into a run of clips that cut together
Every shot gets a still first, because a still costs a fraction of a clip.
Questions people ask before connecting
Do I need a paid plan?
No. Confirm your email and you get free credits every day for your first 3 days, and your agent can spend those. On a subscription the monthly allowance is spent first and the daily free grant sits untouched behind it, so an agent should read the total rather than the free figure.
Can my agent see the pictures I made in the browser?
No. The media tools generate and edit, and they hand back what they just made. Your browser history stays in the browser.
Which models can it use?
All 50 of them, the same list the picker shows. Your agent reads the list with images_list_models and videos_list_models, so it knows about a new model the day we add one.
What if a call times out but the job was made?
It happens: the request reaches us, the job runs and is charged, and the answer is lost on the way back. Every submit tool takes an idempotencyKey. Give each request its own, and if your agent has to resend it gets the same job back and is not charged again. Polling a job is safe to repeat as it is.
What happens if a generation fails?
A job that comes back failed costs you nothing, because we put the credits back.
Is there a REST API as well?
Yes, on the same key. The API docs cover every endpoint, and the MCP tools call those same routes underneath.