Menu
Mattering Automations connector

A Recraft MCP for brand-consistent images and real SVGs inside a workflow

Connect one Recraft API token and 19 nodes land on the canvas: generate raster art or true editable vector files from a prompt, restyle an existing image, repaint one masked area, extend past the edges, cut out or replace a background, trace a PNG into an SVG, upscale, and hold a brand look steady with a custom style. Every call except the account read spends API units, so the connector is built to make that visible rather than surprising.

19 ready-made nodes One credential to connect Works with Claude, ChatGPT, and Matter

What the Recraft MCP connector does

The Recraft connector for Mattering Automations is 19 nodes on Recraft's public API: one read and eighteen actions, with zero destructive nodes. Generation covers a general entry point plus raster-only and vector-only variants, image-to-image, inpainting, outpainting, background replacement, background generation around a mask, and remixing an image with no prompt at all. Processing covers vectorizing to SVG, background removal, crisp upscaling, creative upscaling and region erasure. Two exploration nodes render a grid of interpretations and then a narrower set around one of them. Two more handle prompt enhancement and creating a reusable custom style.

Every node speaks MCP, the open protocol for connecting AI assistants to tools, so Claude, ChatGPT, and any MCP-capable assistant can use them, alongside Matter, the AI assistant built into the platform. The interesting version of that is not an agent chatting about images: it is an agent that reads a product record, writes the prompt, calls Recraft: Generate Vector Image or Recraft: Replace Background, and files the result against the record, with the whole thing running on a trigger rather than in a conversation. Recraft: Get Account And Credit Balance is the node that makes that safe to automate, because it is the only call that spends nothing and it reports the remaining balance.

Recraft's API is PAID, and stating that plainly is more useful than implying anyone can try it free. Recraft runs two separate balances: Studio credits, which are the web app's currency, and API units, which are bought separately. Recraft's getting-started documentation states that generating an API key requires an API-unit balance above zero, so there is no way to obtain a working token without buying units first. Units were sold at one US dollar per thousand at the time this connector was built, and per-operation costs start around two units for the cheapest edit, so a realistic first pass over the whole catalog is small money rather than no money. Confirm the current figures on Recraft's own pricing page before you budget against them.

That billing shape drove two design decisions worth knowing. The connector performs no automatic retries anywhere, because a retried generation is a second charge rather than a free redo, so a failure surfaces to the workflow to decide about. And the count of images per call multiplies the charge, which every generation node says on the field itself, because a canvas user reasonably assumes one call is one charge. The starkest example is upscaling: crisp upscaling costs about four units while creative upscaling costs about two hundred and fifty, so each node's description names the other one, which is what stops a media-library sweep becoming an expensive surprise.

One connection, three surfaces

  • Visual automations. Drag Recraft nodes onto the canvas and wire them to triggers, conditions, and other apps.
  • AI agents, yours or ours. These nodes speak MCP, the open protocol, so Claude, ChatGPT, and any MCP-capable assistant can use them, and Matter, the agent built into the platform, uses them natively.
  • The whole platform. Combine with native tasks, projects, and other connectors, no middleware in between.

Every Recraft node, ready to drop on the canvas

19 nodes, generated from the connector itself so this list always matches what ships. Expand a category to browse.

Read and query nodes 1

Recraft: Get Account And Credit Balance

Read the Recraft account behind the API token: its id name email and remaining API-unit balance. This is the only Recraft endpoint that spends nothing so it is both the credential check and the node to put in front of a batch to stop it when the balance runs low. API units are sold separately from Recraft Studio credits at 1000 units per USD 1.00.

Action nodes 18

Recraft: Create Custom Style

Learn a reusable custom style from 1 to 5 reference images and get back a style_id to pass to the generation nodes. This is how a workflow holds a brand look steady across hundreds of images instead of re-describing it in every prompt. SPENDS API UNITS at about 40 per style. Custom styles made through the API apply to recraftv3 and its vector variant only. Returns the new id at data.id. Recraft documents no API endpoint that lists or deletes styles so SAVE THE RETURNED ID somewhere durable in the same run; styles are otherwise managed in the Styles panel on recraft.ai.

Recraft: Creative Upscale

Raise resolution and REGENERATE detail so a small or soft source gains texture it never had. Use it to rescue a low resolution image but never on a logo or anything that has to match a reference exactly because the added detail is invented. EXPENSIVE at about 250 API units per call which is roughly USD 0.25 and around 60 times the cost of Recraft: Crisp Upscale, so gate it behind a check rather than running it across a whole library. Slow enough to need a longer timeout. NOTE THE RESPONSE SHAPE: ONE image object at data.image.url. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Crisp Upscale

Raise resolution and sharpen WITHOUT inventing new detail so the result still matches the approved original. This is the one to use on brand assets. The cheapest node on the connector at about 4 API units per call which is roughly 60 times cheaper than Recraft: Creative Upscale. NOTE THE RESPONSE SHAPE: ONE image object at data.image.url. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Enhance Prompt

Expand a short prompt into a richer one written the way Recraft's models respond to. Put it in front of a generation node when the prompt text comes from a form field or an agent rather than from a designer. Generates no image but still SPENDS API UNITS at about 10 per call. Returns the rewritten text at data.enhanced_prompt.

Recraft: Erase Region

Delete whatever the mask covers and fill the hole with plausible surroundings. This removes an object rather than replacing it so no prompt is needed. The cheapest edit on the connector at about 2 API units per call. This does NOT modify the source image: it returns a new one. NOTE THE RESPONSE SHAPE: ONE image object at data.image.url. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Explore Prompt Variations

Render a grid of interpretations of one prompt so a human can pick a direction before committing to a full generation. SPENDS API UNITS for every image in the grid so one explore call costs several generations. Returns data.data[] with a url and an image_id per result; feed one image_id to Recraft: Explore Similar Images to drill into that direction. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Explore Similar Images

Take one image from Recraft: Explore Prompt Variations and generate a new set clustered around it. This is the second half of the explore loop: broad grid first then narrow. SPENDS API UNITS for every image produced. Returns data.data[] with a url and an image_id per result. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Generate Background Around A Mask

Like Recraft: Replace Background but YOU supply the mask instead of letting Recraft find the subject. Reach for this when automatic subject detection picks the wrong thing or when the same cutout has to be reused exactly across a batch. RECRAFT V3 MODELS ONLY. SPENDS API UNITS per image (about 40 raster or 80 vector). Returns data.data[] with a url per result. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Generate Image

Generate images from a text prompt. This is the general entry point: it renders raster or vector output depending on the model you choose. SPENDS API UNITS on every successful call and is billed per image, roughly 35 units on recraftv4_1 through 250 on recraftv4 pro tier at 1000 units per USD 1.00. Returns an array at data.data[] with a url per image. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Generate Raster Image

Generate PNG-style raster images from a prompt. Same as Recraft: Generate Image but restricted to raster models so a misconfigured model cannot quietly return SVG into a pipeline expecting a bitmap. SPENDS API UNITS per image (about 22 to 250 units depending on model). Returns data.data[] with a url per image. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Generate Vector Image

Generate true SVG artwork from a prompt. This is the reason most workflows reach for Recraft rather than a raster-only image API: the output is a real editable vector file that scales without quality loss which is what logos icons and print assets need. Restricted to vector models. SPENDS API UNITS per image and vector generation costs MORE than raster (about 44 to 300 units). Returns data.data[] with an SVG url per image. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Image To Image

Redraw an existing image under a new prompt keeping as much or as little of the original as the strength setting says. Use it to restyle a product shot or push a rough sketch toward a finished illustration. The source image is read from a url and is never modified. SPENDS API UNITS per image (about 40 raster or 80 vector). Returns data.data[] with a url per variation. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Inpaint Masked Area

Repaint only the white area of a mask leaving everything else byte-identical. This is the surgical edit: swap a product on a shelf or fix one bad hand without re-rendering the whole frame. RECRAFT V3 MODELS ONLY. The source image is read from a url and is never modified. SPENDS API UNITS per image (about 40 raster or 80 vector). Returns data.data[] with a url per result. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Outpaint Beyond The Edges

Extend an image past its original borders inventing what would have been there. The practical use is reformatting: turn a square asset into a 16:9 banner without cropping the subject. Expand by pixels per side or give a target size. RECRAFT V3 MODELS ONLY. SPENDS API UNITS per image (about 40 raster or 80 vector). Returns data.data[] with a url per result. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Remix Image

Produce fresh variations of an image with no prompt at all. Recraft reads the composition and recomposes it which is the cheap way to get several usable takes on an approved layout. SPENDS API UNITS per image (about 40 units each). Returns data.data[] with a url per variation. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Remove Background

Cut the subject out and return it on a transparent background. The workhorse step in front of any catalogue or composite pipeline. Accepts SVG input as well as raster. Cheap at about 10 API units per call. NOTE THE RESPONSE SHAPE: ONE image object at data.image.url not a data.data[] array. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Replace Background

Keep the subject and generate an entirely new background from a prompt. No mask needed: Recraft works out the subject itself which is what makes this the practical node for putting one product photo into a dozen seasonal settings. RECRAFT V3 MODELS ONLY. SPENDS API UNITS per image (about 40 raster or 80 vector). Returns data.data[] with a url per result. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Recraft: Vectorize Image To SVG

Trace a raster image into a real SVG that scales without quality loss. The obvious use is rescuing a logo that only exists as a PNG. Cheap at about 10 API units per call. NOTE THE RESPONSE SHAPE: this returns ONE image object at data.image.url not the data.data[] array the generation nodes return. Returned image links live for about 24 hours and are publicly readable without a token, so download or re-host inside this run and do not treat one as private.

Connecting Recraft takes one credential

Log in to Recraft, open profile settings, and generate an API token under the API section. One field, one header, no account id, no region and no version header. The important operational fact is upstream of that button: Recraft documents that generating an API key requires an API-unit balance above zero, and API units are a separate purchasable balance from the Studio credits the web app spends. So the honest sequence is to buy units first and generate the token second, and nothing on this page should be read as a free API allowance, because Recraft's live documentation does not describe one. Cost is the gating factor on this connector rather than plan tiers, and it applies per node rather than per family. Recraft: Get Account And Credit Balance is the only call that spends nothing; every other node bills API units on success. At the time this connector was built, Recraft's pricing listed region erasure at about two units, crisp upscaling at about four, vectorizing and background removal and prompt enhancement at about ten, most raster generation between roughly twenty and forty, vector generation higher again, and creative upscaling at about two hundred and fifty. Confirm all of that on Recraft's pricing page rather than budgeting from this page, and note that the exploration nodes render a whole grid, so their cost is a multiple of a single generation.

1

Log in to Recraft and add API units to the account. This step is not optional: the button that generates a token requires a positive API-unit balance, so an account that has never bought units has no key to copy.

2

Open profile settings, go to the API section, and generate a token. Copy the value.

3

In Mattering Automations, open Connectors, click Add Connector, choose Recraft, and create a credential with that token.

4

Click Test Connection. Mattering reads the account endpoint, which is the only Recraft call that costs nothing, so testing and re-testing the credential is free. A rejected token fails before any generation work begins.

5

All 19 Recraft nodes appear in the canvas palette. Put Recraft: Get Account And Credit Balance in front of any batch and branch on the returned balance, so a run stops on a low balance rather than discovering it as a mid-batch failure.

Common uses for the Recraft MCP

Real SVGs, not a bitmap with a vector-sounding name

Recraft: Generate Vector Image produces true editable vector artwork from a prompt, which is the reason most workflows reach for Recraft rather than a raster-only image API: the output scales without quality loss and can be opened and edited rather than only displayed. Recraft: Vectorize Image To SVG comes at the same need from the other direction, tracing an existing raster image into a real SVG, which is the usual rescue for a logo that only exists as a PNG. Recraft: Generate Raster Image is the deliberate opposite, restricted to raster models so a misconfigured model cannot quietly return SVG into a pipeline expecting a bitmap.

Product shots that keep the product and change everything else

Three nodes do the work here and they differ by how much you want to specify. Recraft: Remove Background cuts the subject out onto transparency and is the workhorse in front of any catalogue or composite pipeline, accepting SVG input as well as raster. Recraft: Replace Background keeps the subject and generates a new setting from a prompt with no mask needed, because Recraft works out the subject itself, which is what makes it the practical node for putting one product photo into ten scenes. Recraft: Generate Background Around A Mask is the same idea when you supply the mask yourself, which is what you reach for when automatic subject detection picks the wrong thing or when the same cutout has to be reused exactly.

Surgical edits instead of a re-render

Recraft: Inpaint Masked Area repaints only the white area of a mask and leaves everything else byte-identical, which is how you swap one product on a shelf or fix one detail without re-rendering the frame and losing everything that was already approved. Recraft: Outpaint Beyond The Edges extends an image past its original borders, and its practical use is reformatting: turning a square asset into a wide banner without cropping the subject. Recraft: Erase Region deletes whatever a mask covers and fills the hole with plausible surroundings, needs no prompt, and is the cheapest edit on the connector. It also reads as destructive and is not: it takes a source image and returns a NEW one, so the original is untouched.

Hold one look across hundreds of images

Recraft: Create Custom Style learns a reusable style from one to five reference images and returns a style id, which the generation nodes accept. That is how a workflow keeps a brand look steady across a batch rather than hoping a prompt describes it the same way every time. Two practicalities decide whether this works for you. Persist the returned id in the same run that created it, because Recraft's published API documents no endpoint that lists styles back, so an id you lose is an id you cannot look up. And check the model you are pairing it with, because Recraft's own documentation states that styles are not supported on its newest model families, which makes a style set alongside the wrong model a silent no-op rather than an error.

Explore first, commit second

Recraft: Explore Prompt Variations renders a grid of interpretations of one prompt so a person can pick a direction before a full generation, and Recraft: Explore Similar Images takes one image from that grid and generates a new set clustered around it. Broad first, then narrow. This pairs naturally with an approval step: post the grid into a channel or onto a task, let a human choose, and let the chosen image drive the rest of the run. One cost note belongs with it, because the two nodes here are the ones people underestimate: a grid is several images, so an explore call costs a multiple of a single generation.

Upscale the right way for the job

The two upscalers are not interchangeable and the difference is worth reading before a batch. Recraft: Crisp Upscale raises resolution and sharpens WITHOUT inventing new detail, so the result still matches the approved original, which makes it the one for brand assets and the cheapest node on the connector after the account read. Recraft: Creative Upscale raises resolution and REGENERATES detail, so a small or soft source gains texture it never had, which rescues a poor original and is exactly wrong for a logo or anything that has to match a reference. The cost gap follows the same shape: creative upscaling was priced at roughly sixty times crisp upscaling when this connector was built, which is why running it across a whole media library is the mistake this connector works hardest to make visible.

Example: Turn a new product record into on-brand imagery, with a person in the loop

Built on the Automations canvas in a few minutes, or described in plain language to Matter and built for you.

1

Trigger: a product record is created or updated in a Mattering database, carrying the product name, a description and a photo URL.

2

Recraft: Get Account And Credit Balance runs first and a condition reads the returned balance. Below the threshold the run stops and opens a task, so a batch never discovers an empty balance halfway through.

3

Recraft: Enhance Prompt expands the short description into the kind of prompt Recraft's models respond to. One constraint decides where this can be used: enhancement makes prompts LONGER, and Recraft's older model families cap prompt length far below the newer ones, so an enhanced prompt fed to one of the older models can be rejected for length alone.

4

Recraft: Remove Background cuts out the supplied product photo, and Recraft: Replace Background generates the scene from the enhanced prompt. Note the response shapes differ between these two families: background removal returns a single image object while generation returns an array, so the two steps read their URLs from different paths.

5

Recraft: Explore Prompt Variations renders a small grid and a notification posts it to the channel that approves creative. A person picks one, which is the cheap place to put judgement, before anything expensive runs.

6

Recraft: Crisp Upscale sharpens the chosen image without inventing detail, and an HTTP or storage node downloads and re-hosts the result inside the same run. That last step is not optional: Recraft's output links live for about a day and are publicly reachable by anyone holding them, so a stored link breaks and an unstored asset was never private.

Recraft MCP questions, answered

What Recraft operations are supported?

19 nodes: one read and eighteen actions. Reading the account and its remaining API-unit balance. Generating images from a prompt, either through the general node or through raster-only and vector-only variants. Image-to-image restyling, inpainting a masked area, outpainting beyond the edges, replacing a background from a prompt, generating a background around a supplied mask, and remixing an image with no prompt at all. Vectorizing a raster image to SVG, removing a background, crisp upscaling, creative upscaling and erasing a masked region. Rendering a grid of prompt variations and then a narrower set around one image from it. Enhancing a prompt, and creating a reusable custom style from reference images. Deliberately not included: any style listing, retrieval or deletion, because Recraft's published API reference documents only style creation, and this connector does not build on undocumented paths. Video and the Studio design agent are likewise absent because Recraft documents them as web app features with no HTTP surface.

Does this work with Claude, ChatGPT, or other AI assistants?

Yes. Mattering exposes these nodes over MCP, an open protocol for connecting AI assistants to tools, so any MCP-capable assistant, including Claude and ChatGPT, can connect to your workspace and use them. Matter, the assistant built into the platform, uses the same nodes natively. Worth pairing with the cost note: because every generation node spends API units, an assistant with access to them is an assistant that can spend money, so put the balance check in front of batches and keep the expensive nodes behind a condition or an approval step.

Is there a free tier for the Recraft API?

Not one that Recraft's live documentation describes, and this page will not imply otherwise. Recraft runs two separate balances: Studio credits, which the web app spends and which the free plan describes in its own terms, and API units, which are purchased separately. Recraft's getting-started documentation states that generating an API key requires an API-unit balance above zero, which means the key itself is gated behind a purchase rather than the calls being gated afterwards. Nothing in the published documentation says free-plan Studio credits can be spent through the API. Units were sold at one US dollar per thousand when this connector was built, and the cheapest operations cost single-digit units, so trying the connector is inexpensive rather than free. Check Recraft's pricing page for the current position before committing to a budget.

Recraft publishes its own MCP server. Why use this connector?

It does, with a whole MCP reference section in its documentation including a remote server, and for some jobs a vendor server is the better answer. If what you want is Recraft available inside your own assistant's config for one-off image work, use Recraft's server. This connector is shaped for production instead: the same 19 nodes run as steps on an automation canvas as well as over MCP, so images get generated off a trigger or a schedule with nobody present, chained with every other connector in the catalog so the result lands on the record that caused it, with the token held once in a shared workspace store rather than in each person's local config, and with the balance check available as a gate in front of a batch. Neither replaces the other, and nothing here disparages Recraft's own tooling.

How does the connector handle cost, and can it run away with my balance?

Every node except the account read bills API units on success, and the connector is built to keep that visible. It performs no automatic retries anywhere, deliberately, because a retried generation is a second charge rather than a free redo, so failures surface to the workflow instead of quietly doubling a bill. The images-per-call field on the generation nodes states that it multiplies the charge, because a canvas user reasonably assumes one call is one charge. And the two upscalers name each other in their own descriptions, because the gap between them was roughly sixty to one at build time and running the expensive one across a media library looks like a one-line configuration change. The structural guard is Recraft: Get Account And Credit Balance, which costs nothing: put it in front of a batch, branch on the balance, and stop the run rather than discovering the problem mid-flight.

How long do the generated image links last?

About a day, and they are public. Recraft documents output retention of roughly twenty-four hours, so a workflow that stores the returned URL and reads it next week gets a dead link: download or re-host inside the same run. The second half matters just as much and is easier to miss: those links are accessible without authentication, so anyone holding one can fetch the image. Treat every Recraft output URL as a published asset rather than a private one, which is worth deciding about deliberately for any workflow putting client artwork through the connector. Both facts are stated on the affected nodes as well as here.

Why do some of my Recraft nodes return nothing when the run looks green?

Almost always because two response shapes are being read with one path. Recraft's generation family answers with an ARRAY, so the image URL sits under a data list, while the processing family answers with a SINGLE object, so the URL sits under an image key. A path written for one silently resolves to nothing on the other, with no error anywhere, which makes this the single most likely cause of a Recraft workflow that runs green and produces nothing. Every processing node's description names its own path for exactly this reason. Three nodes are singletons again: prompt enhancement returns the enhanced text, style creation returns an id, and the account read returns the balance and profile.

Can I pick a custom style from a dropdown?

No, and the reason is Recraft's rather than ours. The style id is a genuine identifier and it appears on nine nodes, but Recraft's published API documents style CREATION and nothing else under that resource: there is no endpoint that lists styles, fetches one, or deletes one. With nothing to populate a picker from, the field stays free text rather than pretending a lookup exists, because inventing a source would mean building on an undocumented path. The id comes from Recraft: Create Custom Style in the response, or from copying it out of the Styles panel in the Recraft web app. The practical consequence is a habit: persist the id in the same run that creates the style, because nothing in the API can list it back to you afterwards. The preset style names are a different thing and they ARE a closed list on the nodes, split so a vector endpoint cannot be handed a raster style.

Which nodes have model or size restrictions I should plan around?

Four editing nodes are restricted to one older model family by Recraft itself: inpainting, outpainting, background replacement and background generation around a mask. Their model fields default accordingly and their descriptions say so, because pointing them at a newer model does not work. Custom styles run the other way: Recraft documents styles as unsupported on its newest model families, so setting a style alongside one of those models is a silent no-op rather than an error. Prompt length is the third trap, and it interacts with the first: the older families cap prompts an order of magnitude shorter than the newer ones, and Recraft: Enhance Prompt makes prompts longer, so enhance-then-generate on an older model is a real failure path. Recraft also documents input caps of ten megabytes per file, sixteen megapixels and four thousand and ninety six pixels on a side, with a stricter five megabyte total across all reference images on style creation, and a request ceiling of five per second, which a fan-out loop over a list will exceed without trying.

19 nodes waiting on the canvas

Put Recraft to work inside your workspace

Connect once, then use every node in visual automations and AI agent workflows across the Mattering platform.

All product names, logos, and brands are property of their respective owners and are used for identification purposes only. Use of these names, logos, and brands does not imply endorsement or affiliation. Recraft is a product of its respective owner; Mattering is an independent platform.

Start Free Explore our Products