CoreThink CLI

The CoreThink Workbench from your terminal. Build environments, train programs around small models, check them against the base model, release them and call them — everything the Workbench console does, scriptable.

pipx install corethink       # or: uv tool install corethink
corethink login              # opens your browser to approve this terminal
corethink env list

Python 3.10 or newer, on Linux, macOS and Windows. Docs: guide · every command

Install

error: externally-managed-environment from pip means your system Python (recent Ubuntu, Debian, Fedora, Homebrew) does not take packages: use pipx or uv as above. On Ubuntu: sudo apt install pipx && pipx ensurepath.

Quick start

corethink login                                     # sign in (approve the code in your browser)
corethink hub list                                  # starter programs that already passed their final check
corethink hub start employee-leave-requests         # copy one into your workspace; prints its environment id
corethink train start <environment> --limit 20      # train it, spending at most $20
corethink train watch <run>                         # follow it live: Data → Mine → Synthesise → Verify → Ship
corethink train show <run>                          # base model vs trained program on held-out tasks
corethink endpoint keys create <environment> --name my-app
corethink endpoint call <environment> "Can Ana take 3 days off next week?"

Every command has --help, and corethink --help lists them all.

What you can do

Your own data

corethink env import turns data in another shape into an environment's files: chat logs (OpenAI or Anthropic messages with tool calls, request and response logs, ShareGPT), tables of tasks with an answer column (CSV, TSV, .xlsx, JSON Lines, a JSON array) and folders of text files such as support tickets. It writes IMPORT.md beside them with what was guessed. When the shape is not clear, a small sample (at most 20 records, long text cut) is sent to an AI model to map it, after you agree; --no-ai never sends one.

corethink env import support-chats.jsonl --name "Support chat"   # writes support-chats-environment/
corethink env create support-chats-environment

Training modes and hyperparameters

corethink train start ENV uses Automatic mode: recommended settings, and you set the spending limit. For Advanced mode, pass any setting with --set, or a JSON file with --settings:

corethink train settings                       # every setting, its default and its allowed range
corethink train start payroll --method search \
  --set search_iterations=8 --set minibatch=30 --capacity spot_then_on_demand --limit 25

Values are checked against the same bounds the Workbench enforces before anything is submitted.

Scripts and CI

Make a token once (corethink token create github-actions, or in the Workbench: Settings › CLI sign-ins) and set it as CORETHINK_TOKEN. No login or config file is needed then:

- run: pipx install corethink
- run: corethink train start payroll --limit 10 --yes --watch   # fails the step if the run fails
  env:
    CORETHINK_TOKEN: ${{ secrets.CORETHINK_TOKEN }}

• --json on any command prints machine-readable output to stdout (messages go to stderr).
• --yes skips confirmations; without it, a non-interactive run refuses instead of guessing.
• Exit codes: 0 ok, 1 failed, 2 bad usage, 3 not signed in, 4 not allowed, 5 not found, 6 conflict, 7 busy or unavailable (try again), 130 interrupted.

Settings

corethink login stores a revocable token in ~/.config/corethink/cli.json (on Windows %APPDATA%\CoreThink), readable only by you. Revoke any sign-in in the Workbench under Settings › CLI sign-ins, or with corethink logout.

Shell completion: corethink --install-completion.

Help

Guide: https://corethink.ai/docs/workbench-cli · Every command: https://corethink.ai/docs/workbench-cli-reference · Workbench docs: https://workbench.corethink.ai/app/docs · Support: support@corethink.ai