You are setting up Tilebox for a coding agent that can run terminal commands and edit code, such as Amp, Codex, Claude Code, or Cursor.
If you are using ChatGPT or claude.ai, stop. Tell the user to switch to Codex or Claude Code. Do not suggest MCP.
Get the agent ready to use Tilebox from the terminal:
- Install the Tilebox CLI and Agent Skills.
- Help the user create an API key.
- Give the user the exact shell command to set
TILEBOX_API_KEYand reload their current environment. - Verify read-only setup. Do not change Tilebox resources.
Run:
curl -fsSL https://install.tilebox.com/wizard.sh | shThen verify:
tilebox --help
tilebox agent-contextIf tilebox is not on PATH, use the default installed path directly:
$HOME/.local/bin/tilebox --helpConfirm the Tilebox skills are present in the user's skills directory.
Send the user here:
https://console.tilebox.com/settings/api-keys
Ask them to create a key with the smallest permissions needed for their task.
Do not ask them to paste the key into chat unless there is no safer option. Prefer that they run a command in their own terminal.
Detect the user's shell with:
basename "$SHELL"Then give them one of these commands, replacing YOUR_TILEBOX_API_KEY with their key.
echo 'export TILEBOX_API_KEY="YOUR_TILEBOX_API_KEY"' >> ~/.zshrc && source ~/.zshrcecho 'export TILEBOX_API_KEY="YOUR_TILEBOX_API_KEY"' >> ~/.bashrc && source ~/.bashrcIf they use macOS bash login shells, use:
echo 'export TILEBOX_API_KEY="YOUR_TILEBOX_API_KEY"' >> ~/.bash_profile && source ~/.bash_profileset -Ux TILEBOX_API_KEY "YOUR_TILEBOX_API_KEY"Fish universal variables apply to new shells automatically. For the current shell, also run:
set -gx TILEBOX_API_KEY "YOUR_TILEBOX_API_KEY"Run read-only checks only:
tilebox agent-context
tilebox dataset list --jsonConfirm:
- CLI installed
- skills installed
TILEBOX_API_KEYis available without printing ittilebox agent-contextworks- datasets can be listed
Do not create, update, delete, submit, retry, cancel, deploy, or undeploy anything unless the user explicitly asks.
Keep the response short. Tell the user what is done, what they need to do next, and the one command they should run.