MCP App Store
by KeenEthics
Abyssale icon

Abyssale MCP Server for ChatGPT

By ABYSSALE

Overview

Connect your Abyssale workspace to create and automate marketing visuals from a conversation. Create new editable designs from a brief, a brand URL or a reference. Browse designs and projects, inspect formats and editable layers, find fonts, and check your generation and AI credits. Generate images, video, GIFs, HTML5 banners, and print PDFs from existing designs, including personalized and multi-format outputs. Import new editable designs from JSON, rename, move and delete designs, and create projects. Sign in with your Abyssale account through OAuth. Access depends on your workspace role and granted permissions. Generation consumes Abyssale credits. Subscription and credit purchases take place on Abyssale, outside this plugin.

How to connect

  1. ChatGPT: Open the Abyssale listing in the ChatGPT app directory and choose Connect. ChatGPT listing Abyssale exposes 25 tools here.

Tools

check_asset_job_status

ChatGPT
Check an asset job started by images_remove_background, images_text_to_image or images_upscale. PENDING or PROCESSING means still working — call this again later rather than in a tight loop, and never start the same action again meanwhile. COMPLETED gives the result: outputs, each with its type (image today), url and mime_type. FAILED is final: tell the user it didn't work and why (error_message), and do not retry unless the user asks.

check_design_import_status

ChatGPT
Check the status of a design import started by import_design_from_json. is_finalized: false means still going — call this again later rather than immediately in a tight loop. WAITING_FOR_VALIDATION means at least one image upload command has not been run yet; run them and check again, and the import will proceed on its own. This status only ever arises from an import whose image src was a bare filename. If you cannot execute shell commands, don't hand the user the curl commands as a manual chore — instead, ask the user for a hosted https URL of the image, or inline it as a 'data:image/...;base64,' src if it is 100 KB or less, and call import_design_from_json again; both of those resolve during the import itself and never reach this state. The upload commands stay valid for one hour from when the import was created; after that the import is dead and has to be started over. Once is_finalized: true, status is DONE (result.uuid is the created design's id) or FAILED (error has the reason).

check_generation_status

ChatGPT
Check the status of a generation started by generate_banners_async or generate_multipage_pdf. is_finalized: false means still rendering — call this again later rather than immediately in a tight loop. Once is_finalized: true, every requested format is accounted for: succeeded (banners) or failed (errors, with a reason) — nothing is silently still pending. A format present in NEITHER list while is_finalized is false is still in progress. Results are kept for 7 days; after that this answers generation_request_gone.

create_project

ChatGPT
Create a new project (folder for designs) in the user's Abyssale workspace. Fails if a project with this name already exists. Name must be 2-100 characters.

delete_design

ChatGPT
Permanently delete a design from the user's Abyssale workspace. This cannot be undone through this server — before calling it, confirm with the user, naming the design. Banners already generated from it are not deleted. Requires a role that can manage designs (owner, admin, manager or designer) and access to the design's project; a design the user cannot open, or one already deleted, is reported as not found. A design attached to a product that is not closed cannot be deleted (design_in_open_product).

generate_banners_async

ChatGPT
Generate one or more formats of a static, animated, or printer design asynchronously — video, HTML5, GIF, multi-format in one call, and AI image properties (text_to_image, remove_bg) that generate_static_banner doesn't support. printer_multipage designs aren't supported by this tool — use generate_multipage_pdf instead. By default this waits internally for up to about 60 seconds and returns the finished result directly if it lands within that window — most single/simple-format jobs do. If it's still not done after that, you get back generation_request_id with is_finalized: false instead — tell the user it's taking longer than usual and use check_generation_status with that id to check again later. Pass wait_for_result: false to skip the internal wait entirely and get the id back immediately — do this when you already know the job will be slow (many formats requested) or the user said not to wait. Call get_design/get_design_format first to find valid template_format_names and elements keys — an unknown name is silently accepted and does nothing, with no error.

generate_multipage_pdf

ChatGPT
Generate a print-ready multi-page PDF from a printer_multipage design — the only generation tool that works on that design type; generate_static_banner and generate_banners_async both reject it. Call get_design_as_import first: pages have no id/name field of their own, so address the Nth entry of its payload.pages[] array here as pages.page_N (1-indexed — page_1, page_2, ...), and use that same page's layers[].name for the layer names inside it. Unknown page/layer/property names are silently accepted and do nothing, with no error — get the ids right up front rather than debugging a no-op result. By default this waits internally for up to about 60 seconds and returns the finished result directly if it lands within that window. If it's still not done after that, you get back generation_request_id with is_finalized: false instead — tell the user it's taking longer than usual and use check_generation_status with that id to check again later. Pass wait_for_result: false to skip the internal wait entirely and get the id back immediately — do this when you already know the job will be slow (many pages) or the user said not to wait.

generate_static_banner

ChatGPT
Generate a single image from a static Abyssale design, synchronously — you get the file URL back immediately, no polling. Only works on static designs; animated, printer and printer_multipage designs fail with template_not_static (there is no equivalent synchronous call for them — use generate_banners_async for animated and printer designs, generate_multipage_pdf for printer_multipage). design_id must come from list_designs or search_designs. All fields besides design_id are optional — omitting elements renders the design's saved default content. Rendering is capped at 10 seconds server-side; in practice most renders finish in a few hundred milliseconds. Call get_design (and get_design_format for one specific format) first to find valid template_format_name and elements keys — an unknown layer or attribute name here is silently accepted and does nothing, with no error, so guessing wrong gives no signal that anything went wrong.

get_design

ChatGPT
Look up a design's layers (elements), their types, and which attributes can be overridden — call this before generating rather than guessing at element/attribute names. Generation elements payloads are lenient: an unknown layer or property name is silently accepted and simply does nothing, with no error. This tool does NOT show current values for a specific format — use get_design_format for that (it also gives the exact shape a generation elements payload should take). Works on any design type; type says which tool generates it: static -> generate_static_banner (or generate_banners_async for several formats at once or AI image properties); animated (video, GIF, HTML5) and printer (PDF) -> generate_banners_async; printer_multipage -> generate_multipage_pdf.

get_design_as_import

ChatGPT
Return a design's ENTIRE structure — every format, layer, animation tween, printer bleed/safe/dpi, or every page for printer_multipage designs — in the exact shape a design-import payload takes. Different from get_design/get_design_format: those trim to what generate_static_banner can override; this is the full, unfiltered structure, useful for explaining how a design is built or as a worked example to hand to the user for manual re-import/duplication elsewhere, or as the payload import_design_from_json accepts to recreate/duplicate the design. Works on every design type, including printer_multipage.

get_design_format

ChatGPT
Look up the current values, layout and overridable attributes for ONE format of a design, flattened (not a map keyed by format). Use get_design first to see which formats and layers exist; use this tool once you've picked a format_specifier to see its concrete current state before overriding anything with generate_static_banner. printer_multipage designs have no formats — expect format_not_found for any format_specifier on one of those.

get_design_import_example

ChatGPT
Returns one full worked payload for import_design_from_json's payload argument, downloaded verbatim from developers.abyssale.com's own example gallery — not hand-written. Use static-multi-format when the design needs more than one format: it's the only example that demonstrates format-keyed layout/properties, the structure a single-format example cannot teach. The same examples are also published as MCP resources under abyssale://design-import-examples/*; call this tool instead if your client doesn't surface those.

get_me

ChatGPT
Look up who you're signed in as and which Abyssale workspace you're working in.

get_remaining_credits

ChatGPT
Shows how many generation credits (image/video renders) and AI credits (text-to-image, remove-background, etc.) this workspace has left this billing period.

images_remove_background

ChatGPT
Remove the background of one image, given by URL, and get back a transparent webp. Spends the workspace's AI credits; the result says what it cost. Call list_ai_models to see which models can be used. For an image layer inside a design, use the remove_bg property of generate_banners_async instead. Waits internally for up to about 45 seconds and returns the finished result directly, its URL in outputs, when it lands in that window — most do. Otherwise you get a job_id with status PROCESSING: call check_asset_job_status with it later, and do not call this tool again for the same image meanwhile. If it fails, the failure is final: tell the user it didn't work and why, and do not retry unless the user asks.

images_text_to_image

ChatGPT
Generate one image from a text prompt, optionally guided by reference images (a product, a style, a person…), and get back its URL in outputs. Spends the workspace's AI credits; the result says what it cost. Call list_ai_models (operation images_text_to_image) to see the models, and each model's accepted aspect_ratio and quality values and how many reference images it takes — they differ per model. For an image layer inside a design, use the text_to_image property of generate_banners_async instead. Waits internally for up to about 45 seconds and returns the finished result directly, its URL in outputs, when it lands in that window. Otherwise you get a job_id with status PROCESSING: call check_asset_job_status with it later, and do not call this tool again for the same prompt meanwhile. If it fails, the failure is final: tell the user it didn't work and why, and do not retry unless the user asks.

images_upscale

ChatGPT
Upscale one image, given by URL, to a larger and sharper version (2x by default), and get back its URL in outputs. Spends the workspace's AI credits: the price depends on the output size (width × height × factor²), so it is only known once the job is done — the result says what it cost. Call list_ai_models (operation images_upscale) to see the models and the upscale_factor values each takes. The image must be at least 128 px on each side, and each model caps the output size: a too small image or a too large output fails the job, at no cost. Waits internally for up to about 45 seconds and returns the finished result directly when it lands in that window. Otherwise you get a job_id with status PROCESSING: call check_asset_job_status with it later, and do not call this tool again for the same image meanwhile. If it fails, the failure is final: tell the user it didn't work and why, and do not retry unless the user asks.

import_design_from_json

ChatGPT
Create a design (or workspace template) from a JSON structure — the same shape get_design_as_import returns and POST /designs/import/json accepts: {name, type: static|printer|animated|printer_multipage, target?: {kind: design|workspace_template, project_uuid}, formats/layers (or pages for printer_multipage), ...}. HOW TO SUPPLY AN IMAGE — four options, in order of preference. Use the first one you can actually reach, and do not settle for a later one without ruling out the earlier ones. (1) An already-hosted https URL — atomic and free: Abyssale fetches it, there is no upload step and no second round-trip. Prefer this whenever the image is already online or the user can give you a link to it. (2) An inline data URI, 'data:image/png;base64,<payload>' — also atomic: it is decoded and stored during this same call, so again no upload step. Keep inlined images to 100 KB or less. Base64 costs roughly bytes/2.2 in tokens, so 100 KB is already ~45k; larger payloads are accepted but are far past what is worth spending on one asset. Over 100 KB, use (1) or (3) instead — or, when neither is reachable, recompress the image to fit. Never re-encode, downscale or recompress silently: tell the user what you changed, with the before and after (e.g. '320 KB source, recompressed to ~17 KB at quality 60'). The design keeps whatever you send, so a quality trade outlives the conversation and the user has to know it happened. Shrinking is often fine, especially for a photo shown far smaller than its source; when the original's quality matters more, (1) or (3) keep it exactly. (3) A bare filename, for a file on the user's machine — this is the only option that is NOT atomic: it comes back as WAITING_FOR_VALIDATION with one ready-to-run curl command per layer. Run each one yourself, substituting the real local path, then call check_design_import_status; the import proceeds by itself once the files land. Pick this only if you can execute shell commands on the machine holding the file. (4) If you cannot run a shell, don't hand the user the curl commands as a manual chore. Instead, ask the user for a hosted https URL of the image (1), or use (2), recompressing the image to fit 100 KB and telling them so. One upload is planned per LAYER, so every format of a layer must name the same file — multiple distinct filename srcs on one layer are rejected, so use https URLs for per-format src variation. Image and logo layers accept any image/* type; the stored file is extensionless because the pipeline sniffs magic bytes, so the data URI's declared mimetype does not have to be exact. IMPORTANT — flat vs format-keyed layout/properties: a single-format design may send either shape (flat, or keyed by its one format name or 'default'), but a MULTI-format design MUST key every layer's layout/properties by format name — flat becomes invalid the moment a second format is added. A single-format worked example cannot teach this rule, since flat happens to validate there either way — check get_design_as_import's own payload shape for the design's actual format count before assuming flat is safe. One real layer, trimmed to 2 of its 12 formats, showing the keyed shape: {"name":"image_bg","type":"image","layout":{"linkedin-square":{"x":0,"y":0,"width":1200,"height":1200},"facebook-post":{"x":0,"y":0,"width":1200,"height":1200}},"properties":{"default":{"filter_name":"duotone","filter_properties":{"name":"deep_purple_beige"}},"linkedin-square":{"src":"https://.../image.jpeg","fitting_type":"fill","alignment":"top left"},"facebook-post":{"src":"https://.../image.jpeg","fitting_type":"fill","alignment":"top left"}}}. Full real worked examples (one per design type, downloaded from Abyssale's own docs, not hand-written) are available as resources at abyssale://design-import-examples/{static, static-multi-format, animated, animated-video, printer, printer-multipage} if this client supports reading resources, and at https://developers.abyssale.com/rest-api/designs/import/examples either way. By default this waits internally for up to about 60 seconds and returns the finished re…

list_ai_models

ChatGPT
List the AI models this workspace can use for an AI action (images_remove_background, images_text_to_image, images_upscale), and which one is the default. For images_text_to_image each model also lists the aspect_ratio and quality values it takes, and for images_upscale the upscale_factor values, with their defaults — they differ per model. Use it when the user wants to choose a model, a ratio/quality or a factor, or asks which ones exist. Not needed otherwise: an action called without them uses the defaults. Prices are not listed — each job's result says what it cost.

list_designs

ChatGPT
List the designs in the user's Abyssale workspace, most recently edited first unless sort says otherwise. Use this to browse what exists, or to obtain a design id — the generation tools need an id, not a name. If the user already described which design they want by name, prefer search_designs. Only designs the signed-in user can open are listed: designs in private projects they are not a member of are excluded. Shared workspace templates are NOT designs and are not listed — they must be copied into the workspace before they can be generated. A design that is not in a project, or whose project was deleted, is not listed. Drafts are not listed; only finished designs are.

list_fonts

ChatGPT
List fonts available for text and button elements. There are roughly 1149 Google Fonts (always available, every workspace) plus this company's own custom uploads (few, workspace-specific) — omitting type returns both kinds mixed together. Prefer filtering over browsing: paging through ~1149 entries to find one font is slow and wastes context. Filter by name (exact names rank first), by category for a style ("an elegant serif", "something handwritten"), by weight for a weight and by style for italic — "bold" is weight 700 and "italic" is style italic, not part of a name. Render italic with text markup in the payload: <w=italic>…</w>, or <w=700-italic>…</w> for bold italic. The id this returns is the value to set under elements.<layer_name>.font when calling generate_static_banner or generate_banners_async on a text or button layer — get_design/get_design_format show which layers are text/button but do not enumerate font choices; this tool is where those come from.

list_projects

ChatGPT
List the projects (campaigns/folders that group designs) in the user's Abyssale workspace, in alphabetical order by name. Use this to find a project's id, then pass it as project_id to search_designs or list_designs, or just to answer 'what projects do I have'. Only projects the signed-in user can open are listed — projects they are not a member of are excluded, same as list_designs' own visibility rule.

move_design

ChatGPT
Move one design to another project in the user's Abyssale workspace. The design keeps its id, content and generated banners; its review/approval status is reset. Moving it to the project it is already in changes nothing. Requires a role that can manage designs (owner, admin, manager or designer) and access to both the design's current project and the destination; a destination the user cannot open is reported as project_not_found.

rename_design

ChatGPT
Rename a design in the user's Abyssale workspace. Only the name changes — layers, formats and generated banners are untouched. Name must be 1-100 characters (surrounding spaces are removed). Requires a role that can manage designs (owner, admin, manager or designer) and access to the design's project; a design the user cannot open is reported as not found.

search_designs

ChatGPT
Find designs when the user describes which one they want — 'the black friday banner', 'my instagram design', 'the summer sale story'. Use query first: its words are matched in any order, each in the design name or the project (campaign/folder) name, ignoring case and the hyphens or underscores names are written with. name and project are strict filters for one exact phrase in that one field. Results are most recently edited first unless sort says otherwise. Returns design ids, which the generation tools require. Use list_designs instead to browse without a search term.

Capabilities

Writes

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/abyssale"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/abyssale)