atlas_chat
ChatGPTChat completion with an Atlas Cloud LLM (OpenAI-compatible), billed to the signed-in Atlas account. Use it when the user asks for an answer from a specific Atlas model (for example DeepSeek, Kimi, Qwen or GLM) or wants to compare models; the exact model ID is verified against the live Text catalog first. Billing follows the same spend gate as the generation tools. Call once without confirmation_token. The server bounds the worst-case cost of this exact request (input size, max_tokens and the model's highest per-token rates): under the auto-submit limit it runs straight away and reports the cost from actual token usage — pass that cost on to the user. At or above the limit, or when the model's rates are unknown, it returns the bound and a confirmation_token and spends nothing; show it and stop, and only after a new user message explicitly confirms may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. Do not add a confirmation step of your own on top of this. Args: - model (string, required): An exact current LLM model ID returned by atlas_list_models. - messages (array, required): Array of message objects with "role" and "content" fields. Roles: "system", "user", "assistant" - temperature (number, optional): Sampling temperature, 0-2. Default: 1 - max_tokens (number, optional): Maximum tokens in the response. Setting it also lowers the worst-case bound. - top_p (number, optional): Nucleus sampling parameter, 0-1. Default: 1 Returns: The LLM response including the generated message, token usage, finish reason and cost. Before calling, use atlas_list_models with type="Text" to obtain a current model ID.
atlas_generate_audio
ChatGPTGenerate audio — text-to-speech (TTS) AND music generation — using Atlas Cloud API. This covers both TTS/voice models and music/song models. The live catalog determines which models are currently available. This tool submits the generation request and returns immediately with a prediction ID. Use atlas_get_prediction to check the result later. Billing is gated by cost, not by ceremony. Call once, omitting confirmation_token. The server quotes the request against the live catalog: under the spend limit it submits immediately and the result carries the amount the platform reserved for the request. For an async job that figure is what the run is expected to cost, not a settled charge: report it as an estimate, and only call it spent once polling returns a successful result. A job that ends in status "failed" is not billed at all. At or above the limit, or whenever the platform will not give a firm quote, it returns a quote and an opaque confirmation_token without spending anything; show that exact quote and stop, and only after a new user message explicitly confirms it may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. Parameters are validated against the model's schema BEFORE the request is submitted: if a parameter is missing, has the wrong type, or is not accepted, the tool returns a precise error and does NOT spend credits. IMPORTANT: The "model" parameter requires an exact current model ID. If you don't know it, first call atlas_list_models with type="Audio". Do NOT guess model IDs. NOTE: For speech-to-text (transcription / ASR), use atlas_transcribe_audio instead. You should also use atlas_get_model_info to see the full parameter list and schema for your chosen audio model before calling this tool. Different models accept different params (voice/speaker IDs, format, sample rate, speed, lyrics, style tags, etc.). Args: - model (string, required): The exact audio model ID. Use atlas_list_models with type="Audio" to find valid IDs. - params (object, required): Model-specific parameters as a JSON object. For TTS the main field is usually "text" (the content to synthesize); for music models it is usually "prompt" and/or "lyrics". Use atlas_get_model_info to see available params. Returns: A prediction ID to check the result with atlas_get_prediction. The output is an audio file URL.
atlas_generate_image
ChatGPTGenerate an image using Atlas Cloud API. This also covers image-to-3D and text-to-3D models (they are Image-type models that return 3D files such as GLB/OBJ/USDZ). This tool submits the generation request and returns immediately with a prediction ID. Use atlas_get_prediction to check the result later. Billing is gated by cost, not by ceremony. Call once, omitting confirmation_token. The server quotes the request against the live catalog: under the spend limit it submits immediately and the result carries the amount the platform reserved for the request. For an async job that figure is what the run is expected to cost, not a settled charge: report it as an estimate, and only call it spent once polling returns a successful result. A job that ends in status "failed" is not billed at all. At or above the limit, or whenever the platform will not give a firm quote, it returns a quote and an opaque confirmation_token without spending anything; show that exact quote and stop, and only after a new user message explicitly confirms it may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. Parameters are validated against the model's schema BEFORE the request is submitted: if a parameter is missing, has the wrong type, or is not accepted, the tool returns a precise error and does NOT spend credits. IMPORTANT: The "model" parameter requires an exact current model ID. If you don't know it, first call atlas_list_models with type="Image". Do NOT guess model IDs. You should also use atlas_get_model_info to understand what parameters a specific image model accepts before calling this tool. Args: - model (string, required): The exact image model ID. Use atlas_list_models to find valid IDs. - params (object, required): Model-specific parameters as a JSON object. Each model has different parameters defined in its schema. Common params include "prompt", "image_size", "num_inference_steps", etc. For image-to-3D models, "image" is required. Use atlas_get_model_info to see the full parameter list for your chosen model. Returns: A prediction ID to check the result with atlas_get_prediction.
atlas_generate_video
ChatGPTGenerate a video using Atlas Cloud API. This tool submits the generation request and returns immediately with a prediction ID. Use atlas_get_prediction to check the result later. Billing is gated by cost, not by ceremony. Call once, omitting confirmation_token. The server quotes the request against the live catalog: under the spend limit it submits immediately and the result carries the amount the platform reserved for the request. For an async job that figure is what the run is expected to cost, not a settled charge: report it as an estimate, and only call it spent once polling returns a successful result. A job that ends in status "failed" is not billed at all. At or above the limit, or whenever the platform will not give a firm quote, it returns a quote and an opaque confirmation_token without spending anything; show that exact quote and stop, and only after a new user message explicitly confirms it may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. Parameters are validated against the model's schema BEFORE the request is submitted: if a parameter is missing, has the wrong type, or is not accepted, the tool returns a precise error and does NOT spend credits. IMPORTANT: The "model" parameter requires an exact current model ID. If you don't know it, first call atlas_list_models with type="Video". Do NOT guess model IDs. You should also use atlas_get_model_info to see the full parameter list and schema for your chosen video model before calling this tool. Args: - model (string, required): The exact video model ID. Use atlas_list_models to find valid IDs. - params (object, required): Model-specific parameters as a JSON object. Parameters vary by model - use atlas_get_model_info to see available params. Common ones include: - "prompt" (string): Text description of the video - "image_url" (string): Source image for image-to-video models - "duration" (number): Video duration in seconds - "aspect_ratio" (string): e.g., "16:9", "9:16" Returns: A prediction ID to check the result with atlas_get_prediction. Video generation typically takes 1-5 minutes.
atlas_get_balance
ChatGPTGet the current Atlas Cloud account balance and credit summary for the authenticated API key. Returns available balance, cash, bonus, frozen amount, and credit grant status. Args: None. Examples: - (no params) -> current account balance
atlas_get_model_costs
ChatGPTGet daily model cost buckets (spend amount) for the authenticated API key over a UTC date range. Omitting group_by returns total daily buckets. Use group_by to break costs down by model_type and/or model. API-key identifiers are not exposed by this plugin. Args: - start_date (string, required): Inclusive UTC start date, YYYY-MM-DD. - end_date (string, required): Exclusive UTC end date, YYYY-MM-DD (range <= 180 days). - scope (string, optional): "self" (default) or "account". - group_by (array, optional): Any of "model_type", "model". - model_types (array, optional): Filter by model type, e.g. ["text","image"]. - model_ids (array, optional): Filter by model ID (<= 100). - limit (number, optional): Max grouped rows (1-100, default 100). - page (string, optional): Pagination cursor from a previous next_page. Examples: - start_date="2026-06-01", end_date="2026-07-01" - start_date="2026-06-01", end_date="2026-07-01", group_by=["model"]
atlas_get_model_info
ChatGPTGet detailed information about a specific Atlas Cloud model, including API documentation, input/output schema, pricing, and usage examples. This tool fetches the model's OpenAPI schema and generates comprehensive API documentation with cURL examples. Args: - model (string): An exact current model ID returned by atlas_list_models. Returns: Markdown-formatted model details including: - Model metadata (type, provider, context length, etc.) - Pricing information - Full API input/output schema with parameter descriptions - Required and optional parameters with defaults - cURL usage examples - Playground link
atlas_get_model_usage
ChatGPTGet daily model usage buckets (requests, tokens, image/video counts) for the authenticated API key over a UTC date range. Omitting group_by returns total daily buckets. Use group_by to break usage down by model_type and/or model. API-key identifiers are not exposed by this plugin. Args: - start_date (string, required): Inclusive UTC start date, YYYY-MM-DD. - end_date (string, required): Exclusive UTC end date, YYYY-MM-DD (range <= 180 days). - scope (string, optional): "self" (default) or "account". - group_by (array, optional): Any of "model_type", "model". - model_types (array, optional): Filter by model type, e.g. ["text","image"]. - model_ids (array, optional): Filter by model ID (<= 100). - limit (number, optional): Max grouped rows (1-100, default 100). - page (string, optional): Pagination cursor from a previous next_page. Examples: - start_date="2026-06-01", end_date="2026-06-08" - start_date="2026-06-01", end_date="2026-07-01", group_by=["model"]
atlas_get_prediction
ChatGPTCheck the status and result of an image/video generation task. Use this after submitting a generation request to check if the result is ready. If the status is still "processing" or "starting", wait a moment and try again. When the result is ready (status is "completed" or "succeeded"), the output URLs are returned, and any image is also attached as an image block. You should then: 1. Present the image to the user. Any image is also attached as an image block; whether that block is displayed is up to the client, so do not assume the user has already seen it. 2. Give the URL, and save the file into the working directory without asking first (curl or wget send no Referer, so they work even on buckets that refuse a click from the chat). Then display the saved file inline in the conversation — with a player for video. Handing over a link instead does not satisfy the request; the server cannot attach a video block, so presenting it is your job. Args: - prediction_id (string, required): The prediction ID returned from a generation request Returns: The current status and output of the generation task. Examples: - prediction_id="pred_abc123" -> check generation status
atlas_get_upload_url
ChatGPTGet a short-lived URL that accepts a local file and returns a public Atlas media URL, for use as the image/audio URL parameter of generation tools. Use this when the user has a local file (product photo, reference image, audio to transcribe) and the target model needs a URL. The hosted MCP server cannot read files on the user's machine; this URL lets you push the bytes yourself. Steps: 1. Call this tool. It returns upload_url, expires_at and max_bytes. 2. POST the raw file bytes to upload_url with the file name (keep the extension) in the X-Atlas-Filename header. The curl_example in the result is ready to run. 3. The response JSON has "url". Pass that URL to atlas_generate_image, atlas_generate_video or atlas_quick_generate. Rules: - One URL serves several files until it expires (about 10 minutes); do not request a new one per file. - Uploads are free; nothing is billed until a generation is submitted with the returned URL. - Files larger than max_bytes are rejected with 413. Ask the user to host those elsewhere and give you a public link. - Only report a file as uploaded when the response contained "url". If the command fails, say so and offer the public-link fallback. - Uploaded files are temporary inputs for generation, not permanent hosting. Args: - filename_hint (string, optional): the file name you intend to upload; only used to fill in curl_example. Returns: upload_url, method, filename_header, expires_at, max_bytes, curl_example.
atlas_list_models
ChatGPTList available models on Atlas Cloud, optionally filtered by type, in bounded pages. Args: - type (string, optional): Filter by model type. Options: "Text", "Image", "Video", "Audio" - limit (integer, optional): Models per page, 1-100 (default 50) - offset (integer, optional): Zero-based result offset (default 0) Returns: Markdown-formatted list of models grouped by type, including model ID, description, provider, and pricing. Note: image-to-3D and text-to-3D models are Image-type models (filter with type="Image"). Audio-type models (filter with type="Audio") include text-to-speech / TTS, music generation (e.g., Suno, MiniMax Music), and speech-to-text / ASR (e.g., Seed ASR). Lipsync / talking-avatar models are Video-type models. Examples: - No params -> list all models - type="Image" -> list only image generation models (includes 3D) - type="Video" -> list only video generation models (includes lipsync / talking-avatar) - type="Text" -> list only LLM/text models - type="Audio" -> list only audio models (TTS, music generation, speech-to-text/ASR)
atlas_list_predictions
ChatGPTList the authenticated account's recent generation tasks (image, video, audio), newest first. Use this to find a task ID the user lost, to see what was generated earlier in a session, or to review which recent runs failed. It reads history only — nothing is submitted or billed. Each entry gives prediction_id, model, status, created_at, the output URLs (when finished), the error text (when failed) and a prompt snippet. For the full record of one task, call atlas_get_prediction with its prediction_id. Args: - page (integer, optional, default 1): page number, newest first - size (integer, optional, default 10, max 50): entries per page - state (string, optional): "running" | "succeeded" | "failed" — the reliable three-way filter - model (string, optional): exact model ID, e.g. "bytedance/seedream-v4.7/text-to-image" - model_prefix (string, optional): a whole family, e.g. "google/nano-banana-2" Returns: total, page, size, has_more and the predictions on this page. Notes: - Status values are the platform's own ("completed", "failed", "processing", ...); filter with "state" rather than matching status strings yourself. - A task in status "failed" was not billed.
atlas_quick_generate
ChatGPTTwo-step image, video, or audio generation - automatically finds exactly one model by keyword, fetches its schema, builds and validates parameters, then returns a quote before any billable submission. Covers all generation tasks: text-to-image, image editing, 3D (image/text-to-3D), text/image-to-video, lipsync & talking-avatar (video + audio_url), TTS, and music generation. Parameters are validated against the model's schema BEFORE submitting. If extra_params contains fields the model does not accept (or wrong values), the tool returns a precise error and does NOT spend credits. Billing is gated by cost, not by ceremony. Call once, omitting confirmation_token. The server quotes the request against the live catalog: under the spend limit it submits immediately and the result tells you what was charged — report that amount to the user. At or above the limit, or whenever the platform will not give a firm quote, it returns a quote and an opaque confirmation_token without spending anything; show that exact quote and stop, and only after a new user message explicitly confirms it may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. IMPORTANT: If this tool fails to find a model, call atlas_list_models first to get the exact model list, then use atlas_generate_image / atlas_generate_video / atlas_generate_audio with the exact model ID instead. Do NOT invent extra_params - only pass parameters you know the model accepts (check atlas_get_model_info). The tool searches for models by keyword matching against model ID, display name, and tags. After getting the prediction ID, use atlas_get_prediction to check the result. Args: - model_keyword (string, required): Prefer an exact current model ID returned by atlas_list_models; otherwise use a distinctive display-name keyword. - type (string, required): Generation type: "Image", "Video", or "Audio" - prompt (string, required): Text description of what to generate (for TTS, the text to synthesize; for music, the song description) - image_url (string, optional): Source image URL for image-to-video, image editing, image-to-3D, or talking-avatar models - audio_url (string, optional): Source audio URL for lipsync / talking-avatar video models (the speech the character should say) or speech-to-text models - extra_params (object, optional): Additional model-specific parameters to override defaults (e.g., {"duration": 10, "aspect_ratio": "16:9"}). Only include parameters the model's schema accepts. Returns: A prediction ID to check the result with atlas_get_prediction.
atlas_search_docs
ChatGPTSearch Atlas Cloud documentation, models, and API references by keyword in bounded pages. Returns matching models with descriptions, pricing, and links. For detailed API docs of a specific model, use atlas_get_model_info instead. Args: - query (string): Search keyword to match against model names, types, providers, tags, etc. - limit (integer, optional): Matches per page, 1-100 (default 50) - offset (integer, optional): Zero-based result offset (default 0) Returns: Markdown-formatted list of matching models with key information. Examples: - "video generation" -> finds all video generation models - "deepseek" -> finds all DeepSeek models - "image edit" -> finds image editing models - "qwen" -> finds all Qwen models
atlas_transcribe_audio
ChatGPTTranscribe speech audio to text (ASR / speech-to-text) using Atlas Cloud API. Use this for: transcribing recordings, meetings, interviews, podcasts, voice notes; getting subtitles/captions text from audio. This tool submits the transcription request and returns immediately with a prediction ID. Use atlas_get_prediction to check the result later — the output is the recognized text (some models also return timestamped utterances and speaker info). Parameters are validated against the model's schema BEFORE the request is submitted: if a parameter is missing, has the wrong type, or is not accepted, the tool returns a precise error and does NOT spend credits. IMPORTANT: The "model" parameter requires an exact current speech-to-text model ID. Speech-to-text models are Audio-type models — use atlas_list_models with type="Audio" to find them. Do NOT guess model IDs. The audio must already be reachable through an HTTPS URL. This remote OpenAI plugin does not expose a local-file upload tool. If the user only has a local file, ask them to upload it through an approved Atlas Cloud or customer-controlled HTTPS location before continuing; do not invent or call atlas_upload_media. Billing is gated by cost, not by ceremony. Call once, omitting confirmation_token. The server quotes the request against the live catalog: under the spend limit it submits immediately and the result carries the amount the platform reserved for the request. For an async job that figure is what the run is expected to cost, not a settled charge: report it as an estimate, and only call it spent once polling returns a successful result. A job that ends in status "failed" is not billed at all. At or above the limit, or whenever the platform will not give a firm quote, it returns a quote and an opaque confirmation_token without spending anything; show that exact quote and stop, and only after a new user message explicitly confirms it may you call again with the same idempotency_key, unchanged arguments, and confirmation_token. Args: - model (string, required): The exact speech-to-text model ID. - params (object, required): Model-specific parameters. The main field is usually "audio_url" (URL of the audio to transcribe). Other common params: "language", "format", "enable_punc", "enable_speaker_info". Use atlas_get_model_info to see available params. Returns: A prediction ID to check the result with atlas_get_prediction. The output is the transcribed text.
