backdraw
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Generates images with the user's own ChatGPT account (already logged in in Google Chrome), with no OpenAI API key and no extra cost.
When the user asks for an image "with ChatGPT" (in any language), call generate_image with a detailed, self-contained prompt.
generate_image and fetch_images return within about 45 seconds. Generating usually takes 1-3 minutes, so the first result is often STILL RUNNING or QUEUED with a job id: keep calling job_status(job_id, wait_seconds=45) until it reports a final status. Do not call generate_image again for the same request while its job is still running.
To EDIT an image or keep a character consistent, pass the image file(s) in reference_images (e.g. a file saved by an earlier call). A fresh chat with reference_images is the most reliable way to edit; long revision chains via conversation_url make ChatGPT lose track of earlier images.
For a specific shape (banner, avatar, phone wallpaper) pass aspect_ratio instead of describing the size in the prompt; for stickers/logos/icons that need a see-through background pass transparent_background=true and read the transparency line in the result.
When the user asks for an animation, GIF or moving image, pass animated=true. Use one subject, one clear motion, a still camera and no small text in the scene; aspect_ratio is not applied to the GIF. For a common motion (idle, walk, run, jump, wave, dance, bounce, wiggle, spin) pass motion; describe the subject and style in the prompt.
For local motion of the same character in a seamless sticker/emoji loop, use animate_image on a transparent PNG; for changing poses (walking, jumping, waving), use generate_sprite; for a ready sprite sheet, use animate_sheet.
For a sprite sheet PNG with all poses in one image, use animate_sheet: the character stays unchanged and frames are aligned; tell the user about quality warnings.
When the user mentions one of their Backdraw collections (reference images they saved from the web in Chrome), call list_collections if needed and pass reference_collection to generate_image.
The saved file is named after the start of the prompt. When the prompt opens with something other than the subject (e.g. instructions about the reference images), pass filename_hint with a few words naming what the user asked for.
Read every result line: NEEDS REFERENCE IMAGE means ChatGPT could not reach the image to edit; CHATGPT REPLIED WITHOUT AN IMAGE shows ChatGPT's own text. Never resend the identical prompt in a loop; follow the next step given in the result.
An EDIT CHECK warning means ChatGPT drew a new picture instead of editing; tell the user and retry with reference_images in a fresh chat if an exact edit is needed.
The work happens in a background Chrome tab; never tell the user to switch tabs unless a result says NEEDS USER ACTION.
NEEDS USER ACTION means ChatGPT shows a login or human-verification (CAPTCHA) page: ask the user to open the ChatGPT tab in Chrome and finish it themselves. Never try to get around it. The job continues automatically afterwards (check with job_status).
If no project is selected, call list_projects, then select_project with the user's images project.
If a job timed out but has a conversation link, fetch_images can still download the result later; pass the job's aspect_ratio again so it is cropped the same way.
After SUCCESS, show the saved image to the user in the chat: if you can display or share a local file with the user, do that with the saved path (the preview in the result is a small copy for you to check); otherwise give the saved path clearly.
Report the saved file path(s) and the ChatGPT conversation link to the user.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| generate_imageA | Generate or edit an image with ChatGPT through the user's own logged-in ChatGPT account in Google Chrome (no OpenAI API key, no extra cost). A new chat is started inside the user's selected ChatGPT Project, in a background tab: the user's active tab, window focus, mouse and keyboard are never touched. Waits up to wait_seconds (default 45) and returns: SUCCESS with the saved file path(s) (always kept in the images folder), a preview and the conversation link; NEEDS REFERENCE IMAGE / CHATGPT REPLIED WITHOUT AN IMAGE together with ChatGPT's own reply text; or, while ChatGPT is still working, STILL RUNNING / QUEUED with a job id: then keep calling job_status with that job_id. Use it whenever the user asks for an image, visual, illustration or concept art with ChatGPT, in any language. Write a detailed, self-contained prompt (subject, style, composition, colors, mood); give the shape in aspect_ratio rather than in the prompt. To edit an image or keep a character consistent, pass the image file(s) in reference_images: a fresh chat with reference_images is the most reliable way to edit. To use one of the user's saved reference collections (see list_collections), pass reference_collection. |
| generate_spriteA | Ask ChatGPT to draw poses in one sprite sheet; Backdraw cuts and aligns frames locally, saving a transparent GIF, atlas and manifest. Retries once in a new chat if count, empty-frame or identity checks fail. Returns a job_id; follow with job_status. For another motion of the same character, pass a frame from the previous GIF folder or a character PNG in reference_images. |
| job_statusA | Check a ChatGPT image job by job_id, waiting up to wait_seconds (default 45) for it to finish; call it again while the result says STILL RUNNING or QUEUED. Without job_id it reports the bridge itself: whether Chrome and the extension are connected, the selected ChatGPT project, the images folder, the queue and recent jobs. |
| cancel_jobA | Cancel a queued or running ChatGPT image job. If ChatGPT is already generating, it presses Stop in the background tab and closes that tab. |
| fetch_imagesA | Download the generated images of the latest answer in an existing ChatGPT conversation (by its link) without sending anything. Use it when a job timed out or failed after ChatGPT had started, but the chat finished later, or to save images from an earlier chat. Like generate_image it returns within wait_seconds (default 45); if it says STILL RUNNING or QUEUED, keep calling job_status with its job_id. |
| list_projectsA | List the Projects in the user's ChatGPT account (read from ChatGPT's sidebar in a background tab) and show which one is selected for images. Use it during setup. |
| select_projectA | Select and remember the ChatGPT Project where new image chats are created. Pass the project name as shown in ChatGPT (matched loosely) or its https://chatgpt.com/g/g-p-…/project link. |
| list_collectionsA | List the reference collections the user saved in Chrome (right-click an image, "Add to Backdraw"): each with its image count, folder and files (newest first, with the site they came from). Pass a collection's name as generate_image's reference_collection to attach its images (up to 16 references; reference_files picks specific ones). |
| animate_sheetA | Split a local sprite sheet PNG with all poses in one image into aligned frames and a transparent looping GIF, atlas PNG, manifest, frame PNGs and contact preview in a new sibling folder. No ChatGPT or Chrome; takes seconds. The source PNG is never changed. Frames share a ground line and horizontal center of mass; jump adds an upward arc. Report quality warnings to the user. For light motion from one image, use animate_image. |
| animate_imageA | Move a local PNG (e.g. a sticker saved with transparent_background=true) with code and save a new looping GIF beside it as _.gif, without overwriting files. No ChatGPT, no Chrome; takes seconds. The character stays unchanged: it only bounces, sways, squashes or breathes. For pose changes such as walking or waving, use generate_sprite. The source PNG is never changed. Returns a numbered preview of every frame. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 10 tools
Most tools are clearly separable by resource and action, and the descriptions explicitly draw the boundaries (e.g. animate_image vs generate_sprite: pose changes require the latter). The only likely confusion is within the animation family — generate_sprite, animate_sheet and animate_image all end in frame/atlas/GIF outputs and differ mainly by input source, which an agent must read carefully to pick correctly. job_status is also dual-purpose (job lookup plus bridge health), but the description spells out both modes.
Nine of ten names follow a clean snake_case verb_noun pattern (generate_image, animate_sheet, list_collections, fetch_images, cancel_job, etc.). The exception is job_status, a noun phrase that also doubles as a health check, which is a minor deviation rather than a break in convention.
Ten tools is squarely in the well-scoped range: four for the ChatGPT image pipeline (generate_image, generate_sprite, fetch_images, job_status/cancel_job), two for setup (list_projects, select_project), one for references (list_collections), and two for local animation post-processing. Each tool covers a distinct stage and none feels redundant.
The core lifecycle is covered end-to-end: configure project, generate or edit, poll or cancel, recover timed-out results, and post-process into animations. Minor gaps remain in collection management (list only, no add/remove from disk or delete of generated files) and in configuration such as setting the images folder. These are workable around, mostly via the extension itself.