Skip to content

Claude Code Skill

The Tilde Agent Skill for Claude Code teaches Claude how to work with your Tilde repositories directly from the command line — running sandboxed commands, launching interactive shells, and managing sandbox lifecycle through natural conversation.

The Agent Skill works by giving Claude knowledge of the tilde CLI commands. Claude runs CLI commands on your behalf, so you see exactly what it does at every step.

Source: github.com/tilderun/tilde-skill

Prerequisites

Install the Tilde CLI

Download and install the tilde CLI — see CLI for platform-specific instructions.

Create an Agent Key

The CLI requires an agent API key (prefixed tak-). Create an agent and key in the Tilde web UI or via the API.

Set the key as an environment variable before launching Claude Code:

export TILDE_API_KEY="tak-your-agent-key-here"

Installation

Clone the skill repository and run the install script:

git clone https://github.com/tilderun/tilde-skill.git
cd tilde-skill
./install.sh

This installs the skill files to ~/.claude/skills/tilde/, making them available in all Claude Code sessions.

To install for a single project instead:

./install.sh --project

This writes to ./.claude/skills/tilde/ in the current project directory.

Manual install

You can also copy the skill files directly:

mkdir -p ~/.claude/skills/tilde
cp SKILL.md SESSIONS.md DATA-OPS.md WORKFLOWS.md ~/.claude/skills/tilde/

How It Works

Once installed, Claude automatically uses the skill when you ask about Tilde sandboxes. Claude can run commands in sandboxed environments, start interactive shells, and manage sandbox lifecycle on your behalf.

Example: Running a Data Pipeline with Claude

Here is a real-world scenario showing how the skill works in practice. Suppose you want Claude to run a Python analysis script inside a sandbox.

You say:

Run my analysis script in a Python 3.12 sandbox on my-team/my-repo

Claude will:

tilde exec my-team/my-repo --image python:3.12 -- python /sandbox/analyze.py

Claude streams the output in real time and reports the results.

For longer-running jobs, Claude can use detached mode:

# Start the job in the background
SANDBOX_ID=$(tilde sandbox run -r my-team/my-repo --image python:3.12 -d -- python /sandbox/train.py)

# Check on it later
tilde sandbox info -r my-team/my-repo $SANDBOX_ID

# Follow the logs
tilde sandbox logs -r my-team/my-repo -f $SANDBOX_ID

Use Cases

Run Commands in a Sandbox

Ask Claude to execute commands in an isolated environment:

  • "Run ls -la in a sandbox on my-team/my-repo"
  • "Execute my Python script in a python:3.12 sandbox with FOO=bar"
  • "Run the test suite in an alpine container on my-team/my-repo"

Claude uses tilde exec to run the command and streams the output.

Interactive Debugging

Ask Claude to start an interactive session:

  • "Open a shell in my-team/my-repo so I can debug"
  • "Start an interactive Python session in a sandbox"

Claude uses tilde shell to open a fully interactive terminal.

Monitor Running Sandboxes

Ask Claude to check on sandbox status:

  • "What's the status of sandbox abc123 on my-team/my-repo?"
  • "Show me the logs for the last sandbox run"

Claude uses tilde sandbox info and tilde sandbox logs to retrieve details.

List Repositories

Ask Claude what's available:

  • "What repositories do I have access to?"
  • "List the repositories in my-team"

Claude uses tilde repository ls to list accessible repositories.

Environment Variables

Variable Required Default Description
TILDE_API_KEY Yes — Agent API key (must start with tak-)
TILDE_ENDPOINT_URL No https://tilde.run Tilde API endpoint