CLI
The whole flow from a terminal, for scripts, and the MCP server for assistants.
One binary for macOS, Linux and Windows that does what the dashboard does. Open source: typestrict-lab/cli.
curl -fsSL https://typestrict.ai/install.sh | sh # macOS and Linux; Windows: the zip on the latest release
typestrict login # paste an API key, or set TYPESTRICT_API_KEYThe flow
typestrict train tickets.csv --goal "Which team handles the ticket?" --watch # upload, mask, read
typestrict structure <dataset> # the agent turns the data into training examples
typestrict review <dataset> # its proposal: the question, the answers, the warnings
typestrict structure <dataset> --feedback "..." # adjust it
typestrict label <dataset> # when asked, answer the test examples in the terminal
typestrict run <dataset> --watch # shows the cost and asks before spending the credit
typestrict result <model> # the verdict against Jev
typestrict pull <model> # decision_ai.json, tokenizer.json, onnx/model_8bit.onnx| Command | What it does |
|---|---|
train <files> [--goal] [--question] [--hf org/name:file] [--github org/name] [--no-anonymize] [--watch] | Sends data and starts preparing a training. |
status [<id>] [--watch] | Every training with its step and next action, or one of them. --watch follows it and exits with 1 if it failed. |
structure <dataset> [--question] [--feedback] | Runs the structuring agent, or again with the decisions you picked or your feedback. |
suggest <dataset> | Asks the agent for decisions this data can teach. |
review <dataset> | Shows the agent's proposal. |
label <dataset> | Answers the test examples one by one. |
encoders | The base models and which can be picked. |
credits | Your balance. |
run <dataset> [--base ts-moberto-150m] [--yes] [--watch] | Starts the training. Asks before spending unless --yes. |
result <model> | The verdict against Jev and until when the files can be downloaded. |
pull <model> [--dir] | Downloads the model package. |
mcp | The MCP server over stdio, for assistants. |
Every command takes --json for machine-readable output, so it fits scripts and CI.
For assistants
typestrict mcpThe same flow as MCP tools: list_trainings, get_training, wait_for_change, create_dataset,
structure_dataset, suggest_decisions, next_example_to_label, label_example, list_base_models, get_credits,
quote_training, start_training, get_model, get_audit, download_model. Starting a training always takes a
quote and its confirmation, so an assistant never spends a credit without asking.
The same tools are served remotely at https://api.typestrict.ai/mcp, with nothing to install. The local server is
for when the assistant should send files of any kind from your disk and save the model into your project. Setup for
each assistant: AI assistants.