STAYCHARTED AI MODEL TRAINER

Release notes

What’s new in StayCharted AI Model Trainer, newest first.

AI assistants guide →

App: every plan · Access codes: Essentials and Business

Ask your published models from an Android phone: take a photo, choose one, or type a question, and get the answer with the model's confidence. On Essentials and Business, unsure answers go to the Review Queue on their own. Sign in with your AMT account, or use an access code: a Builder makes one for a model under Use → Access codes, with limits on days, phones and answers a day, and can pause or revoke it at any time. Codes and every phone that joins are in the Activity log. The iPhone app follows. Get it on Google Play → · Mobile app →

Essentials and Business

Connect Snowflake or Databricks in Settings → Warehouses, then train a model from a table and fill new rows from a table. Every answer, with its confidence and whether it needs review, goes to one results table in your warehouse, and "only new rows" makes a regular fill read just what was added since. StayCharted only reads the tables you list. Snowflake signs in with a key pair StayCharted makes; Databricks with a service principal. Guide →

The category from each picture's own metadata. Photos your photo library, DAM or inspection app has already keyworded can train a picture model as they are. On Where is each picture's category?, choose In the picture's own metadata, and the field: Keywords, Title or Description. The screen shows how many photos have each one before anything is imported. If a photo has several keywords, a pattern picks the one that is the category, such as ^category:(.+)$. It works for a ZIP, a spreadsheet of links and a folder in your storage. The model still learns from the picture itself, so a new photo with no metadata is answered the same way.

New pictures come back with their category written in. When you categorize a ZIP of new pictures, What you get back now offers Add the category to each picture's metadata, as a keyword, title or description. You get the photos back as a ZIP with the category written in, as well as the spreadsheet.

  • Your own keywords stay. The category is added beside them, once.
  • The image data isn't re-encoded, and every other field is kept: camera details, author, copyright, location.
  • Answers the model is unsure of go in a _needs-review/ folder with nothing written, and to the Review Queue.
  • JPEG and PNG are written. Other formats come back as they were, and are listed.
  • From your storage, the category is written into the sorted copies. Originals are never touched.
  • If you prefer just the spreadsheet, leave the box unchecked.

Your files are kept as you uploaded them. The ZIPs you upload to train or to fill are now kept exactly as uploaded, with their metadata, instead of a cleaned copy. AMT's own small copy of each picture, the one your model reads, is still made from the pixels alone. Original training uploads are deleted after 90 days unless deleted sooner. Fill uploads are deleted when the fill finishes, or after seven days if never run. Returned files are deleted 24 hours after first download, or after seven days if never downloaded. See the Privacy Policy for derived copies and other retention periods.

"Categorize new pictures." On picture models, "Fill a spreadsheet" is now Categorize new pictures, because a ZIP or a storage folder is photos, not a spreadsheet. The screen says what you'll get back before you start.

Smaller fixes.

  • Coming back to a ZIP you uploaded but hadn't finished setting up now keeps the metadata choice.
  • "Pictures we couldn't use" says why: no category in the chosen metadata field, or several values and no pattern.

Pictures straight from your storage. Picture models can now train on, and categorize, the photos already in your Amazon S3, Cloudflare R2, Backblaze B2, Wasabi, DigitalOcean Spaces or Google Cloud Storage, or any S3-compatible storage at a public HTTPS address. There's no ZIP to build and no spreadsheet of links. Available with Vision.

  • Connect in Settings → Storage (the workspace Owner). For Amazon S3, a CloudFormation template makes a role in your account, with no key to paste. You list the folders AMT may read, and, if you want sorted copies, the folders it may copy into. Test reads each one.
  • Train from a folder: New model → Pictures → From your storage. The category comes from the folder name or the file name, with a preview before anything is read. Add new photos from this folder reads only the new ones.
  • Categorize a folder: only photos the model hasn't answered before, by default. You get the spreadsheet, and optionally a sorted copy of each photo in a folder for its category (unsure ones in _needs-review/), with tags on each copy (category, confidence, review) for your own workflow to route by.
  • Your originals are never written to: not moved, re-encoded, tagged or deleted. The only write is a copy into a folder you listed.
  • The model remembers its folders, so next time it's one button: Fill new photos, with Open in storage links to your provider's console.

Guide: Connect your storage.

Train in your own Microsoft Foundry and Google Cloud. Dedicated AI Models could already be trained in your own Amazon Bedrock. Now they can also be trained in your own Microsoft Foundry (GPT-4.1 mini, GPT-4.1 nano, GPT-4.1 and GPT-4o mini) and your own Google Cloud (Gemini 3.5 Flash). On Business.

  • Microsoft Foundry: connect with Sign in with Microsoft, then Deploy to Azure gives StayCharted's app two roles on that one resource. There's no key to paste. Each run is tested on a Developer deployment with no hourly hosting fee (inference remains billable); publishing makes the deployment that answers, in your resource.
  • Google Cloud: a Cloud Storage bucket, a service account with Agent Platform User (formerly Vertex AI User) and Storage Object Admin on the bucket, and its key. Gemini is tuned in your project and answers per token.
  • Either way: the training file is deleted from your account when training ends, the trained model stays in your account and answers there, and every model is tested on rows it never saw before you use it.

Train from Google Sheets in your own Google Cloud. The Google Sheets add-on now offers Where it trains, so a sheet's rows can train a Gemini model in your own project and fill the empty rows with it.

The training page says what your platform is doing. While a model trains in your own platform, the page now says when the platform is checking the file, when the job is waiting in the platform's queue, and when it's learning, and names the platform.

Clearer model colors. Dedicated AI stands out from Pre-trained AI wherever the "Use of AI" is shown. Your AI Models now have their own muted blue, apart from StayCharted's purple and the green of "Ready", and a Dedicated AI Model trained in your own cloud keeps StayCharted's purple.

StayCharted is in Claude's connector directory. Find "StayCharted AI Model Trainer" under Settings → Connectors → Browse connectors, or open the listing. Connect, sign in and pick a workspace. Claude can then list your models, show how they score, categorize new items with your trained model, help fix mislabeled rows and, with your yes, retrain and publish. AI assistants guide →

StayCharted is on Microsoft Marketplace. Subscribe to Free, Essentials or Business and pay on your Microsoft bill. Vision is available as an add-on on Essentials and Business. After subscribing, you sign in to StayCharted and choose the workspace. Plan changes and cancellation are in the Azure portal or the Microsoft 365 admin center. View the listing →

Each model's own Review and Filled files. A model's page now has a Review tab, with the answers it was unsure about, and a Filled files tab, with the files it has filled across all its versions. Links to its waiting answers open the tab instead of leaving the page. The workspace's Review and Filled files screens are unchanged. AI assistants (assistant access and the labelling guide) moved to the end of the Use tab.

Your AI Models: use the AI your company already runs. Connect Amazon Bedrock, Microsoft Foundry, Google Cloud, OpenAI, Anthropic or any OpenAI-compatible API in Settings → AI platforms, then choose Use your own AI, without training when you create a model. With every question, your AI is given your categories, your labelling guide and the most similar rows your team already categorized. Nothing is trained. It is tested on rows it hasn't seen, with the same report as every model, so you can compare it and publish the better one. Setting it up and testing it are included on Essentials and Business; publishing it needs the Your AI Models add-on ($49 a month on Essentials, $99 on Business).

Train in your own Amazon Bedrock (Business). Tick Also train models here on an Amazon Bedrock connection, and Dedicated AI Models can be trained with Bedrock's own fine-tuning in your AWS account, on Amazon Nova. The training file goes to a private bucket in your account and is deleted when training ends. The trained model stays in your account and answers there.

Choose your model, more clearly. The first choice is now explicit: Train a model on your examples, or Use your own AI, without training. Training then asks which kind of model, and where it trains: on StayCharted servers or on your own.

Models and Filled files list every kind of model as buttons above a full-width list.

Choose how your model is tested. Before you use a model, AI Model Trainer tests it on rows it never trained on. A new section, How it's tested, on the screen where you choose the columns, lets you pick the test that matches how you'll use the model:

  • Random rows: 15% of each category's rows, as before. Right when every row stands on its own.
  • Keep rows that share a value together: choose a column, such as customer, search query, thread or product. Every row with the same value is either tested or trained on, never both, so the test uses ones the model never saw.
  • Test on the newest rows: choose a date column. The model learns from older rows and is tested on the newest of each category.

When many rows share a customer, a query or a thread, a random test can make a model look far better than it will be on new ones. When a column's name says rows belong together, we suggest keeping them together, and the validation report says which test was used. The Google Sheets add-on and AI assistants use the suggestion too.

Dedicated AI Model reports match real use. The validation report now reads the model's answers the same way your fills, the Try box and the Prediction API do: choosing one of your categories. Earlier reports could show a lower accuracy than the model actually delivers. Runs trained from today get the new report; a version trained before keeps its report until you train again or test it with a file whose answers you know.

Fewer false alarms in the personal-information check. Product codes, part and model numbers, and year ranges ("010-02028-00", "2016-2020") are no longer reported as phone numbers, bank account numbers (IBAN) or card numbers. Real ones are still found.

Emails, in one place. Profile has one Emails card with three switches: the Review Queue summary, getting started emails, and product tips and news.

Getting started emails start when you sign up, whether or not you've confirmed your address yet. Product tips and news still start only once you confirm it, and only if you ticked that box.

Try a picture from your computer. On a picture model, Try a picture now has Choose picture next to the link box, so you can test your model on a photo from your desktop without putting it online first. Phone photos are made smaller in your browser to fit the 1.5 MB limit. The photo isn't kept: the prediction is saved under the file name.

See the picture behind each answer. Recent predictions show a thumbnail of the photo, beside the category and confidence. Marking an uploaded photo's answer right or wrong still teaches the next version.

Click a thumbnail to see the picture larger, with the model's answer under it. Thumbnails of uploaded photos stay when you leave the page and come back (in the same browser).

Fix: a request that's too large now says so, and suggests sending large pictures as links, instead of a general error.

Buy through Microsoft Marketplace. AI Model Trainer is listed on Microsoft Marketplace, with Free, Essentials and Business plans, and Vision billed by Microsoft on your Azure invoice.

  • One Microsoft subscription pays for one workspace: a new one, or one you already have. Moving an existing workspace refunds the unused days of its card payment.
  • Change or cancel the plan in Microsoft; the workspace follows.

The Prediction API describes your models.

  • GET /api/v1/models now says what each model reads (its input columns, or pictures) and how it answers (one category, a list, or several fields), and which workspace the API key belongs to.
  • New: GET /api/v1/models/{id}, with the model's input and answer as JSON Schema, for tools that build a form from it.
  • Both are in the API reference.

"Starting up", not "failed". When a Dedicated AI Model has no GPU ready to answer on, Try says the model may be starting up, and the Prediction API answers 503 api.model_unavailable instead of a general error. The call isn't counted against your allowance.

Getting started emails, with their own switch. A few emails to help you get started after you confirm your address, such as after your first model or your first filled file. Turn them off any time under Profile. Product tips and news are still sent only if you tick that box.

Start here says what to bring: a spreadsheet of past decisions, photos in folders, or a SharePoint permission export. AI Model Trainer learns from how your team already does the work and sends unsure answers to a person.

SharePoint permission exports. Upload your SharePoint permission report as it is, train, and get suggested access for new files. It suggests; you decide.

Answers with several parts. An answer can be a list (the groups that can open a file) or several columns (GL account and cost center).

Pictures with text. Picture models can also read text, such as a customer's message with their photo.

Download your trained model from Manage.

More to start from. New Start here cards and samples, including SharePoint and file permissions.

JSON fills keep your file's shape.

Start here. Your first sign-in says what AI Model Trainer does, what to bring, and the four steps, with a sample to try.

See your rows first on the upload screens, before choosing columns.

Help in the Review Queue: the model's top three categories and similar items your team labeled, beside each waiting answer.

Find what's waiting. Models, fills and Try link to their answers in Review, and a daily summary email is optional under Profile.

See a picture larger. Click any picture thumbnail to open it larger, with its category, in Check your pictures, Look through the pictures, the Review Queue, cold start and the validation report. Press Esc or click outside to close it.

Clearer "Where the model got it wrong".

  • The list now says plainly what it shows: the rows of yours the model was tested on but never trained on, how many it got wrong, and that the rate is an estimate for new rows similar to those tested.
  • The button says which label it means: The file's label is wrong. Use it only when your file had the wrong category for a row; when the model is wrong, there's nothing to correct.
  • Rows show as you'd read them, without the column name in front. A model that reads several columns shows them joined with " · ".

Undo a label correction. A label you corrected on the validation report can be put back with Undo, as can "Relabel all" for similar rows. A corrected row keeps saying "Corrected to … — will apply on retrain" after you reload the page.

Leave while pictures are read. You can leave a picture model while its pictures are being read. The Models list shows it as "Reading pictures", and opening it again takes you to where you left off.

Calmer colors. Work in progress is shown in StayCharted's purple, and amber is kept for things that need your attention. Errors are red. The "Confirm your email address" reminder is now a quiet strip.

Cold start fixes.

  • Confirm and train starts on the model that read your rows, rather than a Dedicated AI Model.
  • "Accept suggestions at or above" starts at 75%.

Product tips and news, only if you ask. An unticked box at sign-up, and a switch under Profile.

Fix: the Use tab of a model that is training again shows the training progress instead of an empty page.

Start from rows that have no categories yet. Choose New model → "Not yet — help me categorize them".

  • Upload your rows or pictures and pick the columns to read. We email you when they're ready to label.
  • Name your categories, or choose Find them for me and StayCharted groups similar rows for you to name. You can split a group in two, merge groups by giving them the same name, or group again with a different number of groups.
  • Review the suggestions, least sure first, and accept every suggestion above a confidence you choose.
  • Then train on the reviewed rows straight away, or download them as CSV or Excel.
  • If some rows already have a category, those count as reviewed and the rest are suggested from them.
  • On Essentials and Business. On Business you can also ask a Dedicated AI Model for the suggestions. Pictures need the Vision add-on.
  • Reading the rows counts toward your monthly rows filled (or pictures read). Grouping, suggesting again and reviewing don't.

The Review Queue. A new Review tab collects every answer a model was unsure about, below the review cutoff you set: from filled files, Try, the API and AI assistants, for text and pictures.

  • Answer one at a time: confirm, pick the right category, skip, or mark "not enough information". Keyboard shortcuts: 1–9 to pick, Enter to confirm, S to skip, U for not enough information, Z to undo.
  • Reviewed answers join the model's next training run. A Reviewer can start that retrain; publishing the new version stays with a Builder.
  • Download a filled file with the reviewed answers in it.
  • On Essentials and Business. Up to 2,000 answers wait per model; answers nobody reviews are removed after 30 days.

Files without a header row. When the first row of an upload looks like data, StayCharted asks whether it's column names or data, and reads it as a row in one click.

Accents and dashes read correctly. CSV files saved as UTF-8 without a byte-order mark (from Google Sheets, Numbers and most tools) showed characters like "—" instead of "—". Every upload and every filled file now reads them correctly. Older Excel CSVs are read correctly too.

AI assistants can start from unlabeled rows. Claude and ChatGPT can start a cold start, group rows, suggest categories and record your reviews. Anything that uses your plan's allowance asks for your confirmation first.

Test a training run on a file whose answers you know. On a model's validation report, choose any finished run, published or not, and upload a file that already has the right category in a column. The run fills a new column beside yours and the two are compared row by row: accuracy, how many rows would be flagged at each cutoff and how many mistakes they'd catch, and the categories it mixes up. Download the file to see every row. It counts as predictions, like any fill.

Choose how much of a fill is flagged for checking. Every training run shows, for each confidence cutoff, the share of rows that would be flagged, the share of the model's mistakes among them, and the projected accuracy if every flagged mistake were corrected. Pick the cutoff your fills use to flag rows for that model.

Compare runs, publish any of them. Every run trained on the same data is listed side by side on the Versions tab: accuracy, mistakes, rows flagged, training time, and its test-file result when you have one. Open any run's report and publish the one you choose. A model is the kind of the run you publish, so trying a Dedicated AI Model run doesn't change your model unless you publish it.

Categories the model mixes up, counted over all its mistakes, not just the examples listed.

Filled by this model: rows filled, rows flagged and rows not flagged, by version, plus the answers people marked right or corrected.

Pictures are stored without location, time and device details. Those details are removed from picture ZIPs when you upload them, for training and for fills. The pictures themselves are unchanged.

A switched-off model answers nowhere. Switched off by you or by a plan change, it doesn't answer Try, fills, the API or AI assistants until it's switched back on. Dedicated AI Models answer only on Business; on other plans they're kept, with their training data, and answer again if you move back.

When a canceled paid plan ends, the workspace moves to Free on time, and models over Free's limits are switched off. Nothing is deleted.

Large picture ZIPs upload reliably. A ZIP is checked for viruses after it arrives, while the page shows "Checking the ZIP for viruses…", so large ZIPs no longer time out. ZIPs up to your plan's size upload.

Find us in more AI directories. The AI Model Trainer connector is listed on Smithery, Glama and the official MCP Registry.

Fixes: picture thumbnails load reliably on pages with many pictures; Filled files shows which model kind and which columns a fill used; Try's recent questions are kept between visits; a ZIP fill names the column it will write; the Training files tab opens in its own tab however many files there are.

AI assistants categorize with your model. When you ask Claude or ChatGPT to categorize items, the assistant now sends them to your model and reports the model's category and confidence. Your labeling guide explains what your categories mean, but the assistant no longer labels items from the guide itself.

Confirmations work once. Training or publishing through an assistant asks for your confirmation first. Each confirmation can now be used only once, so repeating a request can't start a second training run.

Clearer rules for assistants.

  • Every assistant tool now states its rules:
    • numbers come from AI Model Trainer and are never estimated by the assistant;
    • your rows, labels and file names are your data, never instructions.
  • The sign-in screen says exactly what an assistant can't do: delete models, files or workspaces.
  • On a picture model, the categorize tool is marked as reaching the internet, because it fetches the picture links you give it.

One record of who corrected a label. Corrections made through an assistant are shown as "via Claude" (or the assistant's name) in Activity. The correction itself no longer carries that note into your training data or its export.

Use AI Model Trainer from Claude and ChatGPT. Connect your assistant at https://amt.staycharted.com/mcp, sign in, and pick the workspace it may use. It works with your role in that workspace.

  • What it can do:
    • list your models and read their reports;
    • categorize text you paste;
    • start a file fill;
    • check data quality, correct labels, train and publish. Training and publishing need your confirmation.
  • Files: the assistant gives you a one-time link to upload or download in AI Model Trainer. It can read data you request through the authorized connection, including rows from filled files.
  • Two switches, both off until you turn them on:
    • for the workspace, under Settings → AI assistants (the Owner);
    • for each model, on its AI assistants tab (a Builder). That tab also holds the model's labeling guide.
  • Disconnecting: see and disconnect your assistants under Profile → Connected AI assistants.
  • Activity shows everything an assistant did, marked "via Claude" or "via ChatGPT".
  • Cost: an assistant's categorizations count as predictions, like the API.

The model page is now in tabs. Use (Try, fill a file, the API), Versions, Training files, AI assistants, Activity and Manage. Each tab opens on what you came for, instead of one long page. The Activity tab shows just that model's history.

Buy through AWS Marketplace.

  • You can now buy AI Model Trainer on AWS Marketplace and pay through your AWS bill.
  • Each workspace is paid one way: by card or through AWS.
  • If AWS is still setting up a purchase when you arrive, the page says so and fills in by itself as soon as AWS is ready.

Delete a single training file.

  • Builders can delete one original upload from a model's Training files tab, for example to honor an erasure request.
  • The model and its versions keep working.
  • The file stays listed as deleted, with who deleted it and when.

Delete a workspace yourself.

  • An Owner can delete a workspace under All workspaces, confirming by typing its name.
  • Its direct subscription is canceled straight away, with no refund for the unused billing period. For an AWS Marketplace workspace, deletion frees its unit; AWS billing continues until you change the purchase in AWS.
  • Its models, files and API keys are deleted, and everyone's access is removed.
  • Billing history is kept.
  • You can't delete your last workspace, or one with training or a fill running.
  • A large workspace shows "Being deleted" while it finishes.

Faster model deletion. Model deletion has been made faster.

Clear answers about automatic deletion.

  • The Training files tab says how long original uploads are kept: 90 days, for text files and picture ZIPs alike.
  • Filled files says when each file will be deleted.
  • When a file is gone, it says why:
    • deleted by someone in the workspace;
    • deleted a set time after download;
    • not downloaded in time;
    • reached the end of its retention period.
  • The periods are the ones stated in our Privacy Policy.

StayCharted AI Model Trainer is open. Upload a spreadsheet of examples you've already labeled, train a model on your own categories, and use it to fill in the missing column in new files, or call it from your own systems.

Models for text

  • Classifier Model: learns your categories from your words. On every plan, Free included.
  • AI Classifier Model: also reads meaning, using a pre-trained AI, so it copes with different wording. From Essentials up; Free includes one.
  • Dedicated AI Model: an AI fine-tuned on your own examples. On Business.

Models for pictures, with the Vision add-on

  • AI Image Classifier Model: sorts pictures into your categories, from a ZIP of folders or a spreadsheet of picture links.
  • Dedicated AI Image Model: an AI fine-tuned on your own pictures. On Business with Vision.

Before training

  • Data quality checks: duplicate rows, category names that look like the same category, categories with too few examples, and, for pictures, pictures that may be in the wrong category.
  • Personal information: email addresses, phone and card numbers, IBANs, US Social Security numbers, IP addresses, passwords and API keys are checked for before training. Pattern-based checks may not find every sensitive value. Mask them in one click, leave the rows out, or keep them. A masked model also masks what it reads later.
  • A validation report for every training run: accuracy on examples the model never saw, accuracy per category, and the examples it got wrong. You can fix a label from there.

Using a model

  • Try a single example in the browser.
  • Fill a file: CSV, Excel or JSON. Rows the model is unsure about are marked for checking. You publish each new version when you're ready, and the published version keeps answering while a new one trains.
  • The Prediction API, with keys per workspace. From Essentials up.

Working together

  • Workspaces, each with its own plan, models and people.
  • Roles: Owner, Builder, Reviewer and Viewer.
  • Activity: a record of who did what, with CSV export.

Plans: Free, Essentials and Business, with Vision as an add-on. Enterprise by arrangement.