AI image tools — eight image APIs behind one connection
Generation, editing, background removal, upscaling, vector output, transparent cutouts, compression and format conversion — eight operations, one API key, one credit balance. Each is a REST endpoint you can call from your own code and an MCP tool that appears in your agent’s tool list after a single config line.
Most image MCP servers generate a picture and stop. The table below is the whole surface, including the parts that behave differently from what their names suggest.
Create
make something newgenerate_image8–30 crImage generation
Text-to-image and image-to-image with six aspect ratios and reference-image steering.
text to image · AI image generator API
generate_transparent_image16–38 crTransparent PNG generation
Generate the subject and cut it out in one call, on a real alpha channel.
transparent PNG generator · sticker API
text_to_svg5 crText to SVG
Genuine vector paths from a prompt — the description is the only control you get.
AI SVG generator · AI icon generator
Change
work on what existsedit_image8–30 crImage editing
Inpaint, retouch, restyle, remove objects and re-frame — all from a sentence, no mask.
inpainting API · AI photo editor API
remove_background8 crBackground removal
Matting rather than a hard mask, so hair and soft edges stay usable.
background removal API · cutout API
upscale_image15 crUpscaling
Crisp super-resolution that sharpens real detail instead of inventing it.
AI upscaler API · crisp super resolution
Ship
get it production-readyAll eight compared
Cost, what actually comes back, and the one thing per tool that is not obvious from its name. Several of these differ from what the tool schema implies — where they do, this table follows the code rather than the schema.
| Tool | Credits | Returns | Worth knowing |
|---|---|---|---|
generate_image | 8–30 | Hosted URL | Cost tracks the model you pick. The style field is recorded and echoed back, but never reaches the model. |
edit_image | 8–30 | Hosted URL | Priced like generation, not at a flat rate — an edit runs the same model path. |
generate_transparent_image | 16–38 | Hosted URL | Model price plus 8. Read result.isTransparent: a failed cutout still returns 200 and bills in full. |
text_to_svg | 5 | SVG source, plus a vendor URL or data: URI | The only tool whose file is not hosted on our CDN. The prompt is the sole control — there are no style presets. |
remove_background | 8 | Hosted URL | The model field is accepted but ignored; every cutout runs on the same matting model. |
upscale_image | 15 | Hosted URL | No scale factor — the model decides the size. Transparency comes back white, so cut out afterwards. |
compress_image | 1 | Base64 bytes — no hosted URL | The one tool that hands back the file itself, in result.image. Lossy in every mode it offers. |
convert_format | 1 | Hosted URL | Encoder settings are fixed and only PNG output is lossless. Animation is not carried across. |
Seven of eight return a URL — not all eight
It is tempting to say every call returns a hosted image URL, because that convention is what lets the tools chain. But compress_image returns base64 bytes with no hosted copy, and text_to_svg returns the SVG source alongside a vendor URL or a data URI rather than a file on our CDN. Both are worth knowing before you write the glue code that assumes otherwise.
Three more tools that are not about pixels
The MCP server exposes eleven tools in total. These three are what let an agent behave responsibly — check the balance before it spends, read the model list before it guesses, and batch work instead of serialising it.
The tools compose, because most of them speak in URLs
Every tool accepts a hosted image URL as input, and seven of the eight return one. That convention is what turns separate endpoints into a pipeline an agent can run end to end without ever holding bytes in memory — until the final step, where compression deliberately hands you the file instead.
The order matters more than it looks
Upscale before you remove the background, never after. The upscaling model does not support transparency and returns a white background, so cutting out first destroys the alpha channel you just paid 8 credits to create. The cutout belongs near the end.
generate_imageprompt → draft.pngedit_imagedraft.png → fixed.pngupscale_imagefixed.png → large.pngremove_backgroundlarge.png → cutout.pngcompress_imagecutout.png → bytesone prompt, one production asset, no human in the middle
What they cost
One balance covers all eleven tools. The fixed-price tools are genuinely fixed — background removal is always 8, upscaling always 15, vector output always 5, and compression and conversion always 1, whatever the file size.
The two that vary are generation and editing, and they vary for the same reason: both price at the rate of the model you choose, from 8 for a fast model to 30 for a premium one. Transparent generation is that model price plus 8 for the cutout stage. Costs are deducted before the work starts and refunded automatically if it fails.
Plans and credit packs →Per call
- Compression, conversion1 credit
- Text to SVG5 credits
- Background removal8 credits
- Upscaling15 credits
- Generation, editing8–30 credits
- Transparent generation16–38 credits
- Balance and model lookupsfree
Questions about the toolset
Which MCP server should I use for image generation?
Most of them generate and stop there — the MCP directories are full of single-purpose servers wrapping DALL·E, Flux or Together AI. That is fine until your agent has a picture and needs to do something with it. The argument for this one is coverage: generation, editing, background removal, upscaling, vector output, compression and format conversion arrive together, on one key and one credit balance, so the agent can finish a job rather than hand you a draft.
What tools does my agent see after connecting?
Eleven. The eight image tools listed above, plus get_user_info, list_models and multicall. They appear in the client’s tool list after one config entry, with no per-tool setup.
Can these tools analyse or describe an image?
No. There is no vision, OCR, classification or object detection here — that is a different category of API, and your agent’s own model can usually already look at an image. These eight tools produce and transform image files. If you need to read what is in a picture, use a vision model; if you need to make or change one, use these.
Do I need a separate API key for each tool?
No. One key covers all eleven tools, one credit balance is drawn down by all of them, and every call lands in the same request log. That is most of the point — the alternative is a background-removal vendor, an upscaling vendor and a generation vendor with three billing relationships between them.
Can I use these without MCP, from ordinary code?
Yes. Every tool is a plain REST endpoint under /v1 with bearer-token auth, and the MCP server is a thin wrapper over exactly those endpoints. Nothing is exclusive to the agent path, so you can prototype in an agent and ship in Python or Node without changing behaviour.
Is there a free tier?
30 credits a month with no card. That is enough for roughly two to three generations, or thirty conversions, which is deliberately sized for trying the pipeline rather than running one.
What order should I call them in?
Generate or edit first, then upscale, then remove the background, then compress or convert last. The one that catches people is upscaling: the model does not support transparency and returns a white background, so a cutout has to come after the upscale, never before it.
Can several tools run at once?
Yes, through multicall. It takes an array of tool-and-arguments objects, runs them concurrently, and returns one result per item in the order you sent them. Every item is priced by its own tool and the total is deducted once, up front.
What happens to my credits if a call fails?
They are refunded automatically. Costs are validated and deducted before the work starts, and a failure refunds the deduction. The exception worth knowing is a partial success — a transparent-PNG call whose generation half works but whose cutout half does not returns 200 and bills in full, which is why that tool returns an isTransparent flag to check.
Reference
Credit costs and response shapes on this page were read from the request handlers rather than from the tool schemas, because on several tools the two disagree. Re-checked 8 September 2026.
Connect once. All eleven tools appear.
30 free credits every month, no card. Pick your client and the setup takes about two minutes.