Manage machine learning datasets and fine-tuning workflows through Adaption’s MCP server. Import data from Hugging Face, Kaggle, or Google Sheets; adapt, augment, translate, localize, and generate synthetic datasets, then export results and launch or monitor AutoScientist training runs with model and hyperparameter recommendations
Try asking
Encrypted at rest, isolated from the model
Resolved from an AES-256-GCM vault at the moment of the call and attached to the request — the model never sees the secrets.
List datasets visible to the authenticated Adaption organization. Returns concise status and pagination fields.
Get a concise public view of one dataset, including processing and column information.
Poll dataset processing status and progress. Use after an asynchronous dataset operation.
Get quality evaluation status and results for one dataset.
Poll saved external export URLs for a dataset after an export operation.
Poll the retrieval index of a dataset after index_dataset. Once status is succeeded, search the dataset by passing its id in file_search.dataset_ids on /responses, or by attaching it to an inference stack. A failed index carries an error; call index_dataset again to rebuild it.
Import a dataset from a HuggingFace, Kaggle, Google Sheets, or public S3 or Cloud Storage object URL. An object URL is one https .csv, .json, .jsonl, or .parquet file.
Estimate or launch adaptation for an imported dataset. Poll dataset status after launch. translate and localize add translated or localized rows, and expand adds generated or retrieved rows before adaptation (1 credit per 100 rows per block). They can be combined; passing any of them sets the whole selection for the run, turning off any mode saved in the dashboard that the call leaves out.
Estimate or launch creation of a new dataset containing the source rows plus curated domain/general rows, or generate similar rows based on the seed dataset.
Deprecated: use expand_dataset, which takes the same arguments and behaves identically. Estimate or launch creation of a new dataset containing the source rows plus curated domain/general rows, or generate similar rows based on the seed dataset.
Estimate or launch creation of a new dataset with sampled rows translated into target languages.
Estimate or launch creation of a new dataset with sampled rows localized for country/language pairs.
Combine compatible ready datasets into one new dataset. This is idempotent when the same idempotency key is reused.
Build a retrieval index so a dataset can be searched. Pick the text columns to search (for most uploaded datasets, the file’s column headers); they are combined into each row’s text in the order given. An unknown column is rejected with the list of columns that can be indexed. Indexing runs in the background: poll get_dataset_retrieval_index until status is succeeded, then pass the dataset id in file_search.dataset_ids on /responses, or attach the dataset to an inference stack. Calling again with the same text_columns returns the existing index; different columns are rejected unless the index failed.
List base models currently available for AutoScientist training.
List AutoScientist runs, optionally filtered to one dataset.
Poll one AutoScientist run, including iteration progress and any public failure message.
Resolve a suitable model and hyperparameters for a ready dataset without launching a run.
Launch a new asynchronous AutoScientist training run on a ready dataset. This spends credits; poll get_autoscientist_run afterward.
List valid domain and qualified subdomain codes for dataset generation.
Estimate or launch generation of a new invented dataset. Launches consume credits.
List your stacks in the authenticated organization, newest first. Pass a stack model_id as model to the Responses API to call it.
Get one stack: its model_id, status, model, tools and feature settings.
Create a stack: a hosted model with tools and feature settings, called through its model_id.
Change a stack. Only the fields sent change; a model or tool the caller cannot use is refused before anything is written.
Delete a stack for good. Requests to its model_id fail afterwards.
Pick your app for a step-by-step guide. Most sign in with OAuth, so there is no API key to copy.
Discovery, routing, credentials, tool scoping and execution logs all happen at the gateway→connections stay ACTIVE with no work from you
Adaption MCP runs through a gateway that holds the credentials, scopes the access and records every call.
Managed auth, hosted MCP servers, and every Adaption MCP tool your agent needs.
Free to start.
Add as preferred source on Google