Le Bac à Sable

Add a sandbox to Le Chat to supercharge its capabilities. Iterate on your files and datasets directly in your sandbox and deploy your product—all from Le Chat.

Le Bac à Sable attaches a Linux container sandbox to every chat, enabling the model to create, modify, and execute files during the conversation. While Mistral Medium excels at coding tasks, the native code interpreter tool has limitations. This is especially useful when working with medium to large files or transferring them between tools.

Examples

• Ask Le Chat to download and clean a dataset. Le Bac à Sable’s MCP provides a tool to run commands for that.
• Build a full-stack app, iterate on the generated code, and deploy it—all from Le Chat.
• Convert video or audio files to your desired format.
• Expose your workspace to the internet with ngrok for quick sharing or testing.
• Push your sandbox files directly to a new GitHub repository.
• …and more 🚀

---

Installation

uv python install
uv sync --locked

Usage

Start the server on port 3000:

uv run main.py

---

Available Tools

Command & File Tools

1. run_command
   Execute arbitrary shell commands with structured output inside the sandbox container.

   • Parameters: command, stdin, timeout, shell, max_output_bytes.
   • Response: segments, exit_code, truncated, timeout, is_error.

   Example:

   run_command("ls -1")

2. write_to_file
   Create or overwrite a text file with provided content. Parent directories created automatically.

   • Parameters: path, content.
   • Response: path, bytes_written, created, timestamp.

3. replace_in_file
   Perform multiple literal (non-regex) search/replace operations in a file.

   • Parameters: path, replacements.
   • Response: changed, replacements, timestamp.

4. read_file
   Read full textual contents of a file inside the sandbox.

   • Parameters: path.
   • Response: content, size, timestamp, truncated.

5. list_file
   List (non-recursive) directory entries.

   • Parameters: path (default ".").
   • Response: entries, count, timestamp.

---

Sandbox Management

6. spawn_sandbox
   Ensures a long-lived detached docker container exists.

   • Parameters: name, image, recreate.
   • Response: container_id, created, message.

7. list_files
   List files in /workspace inside the sandbox container.

---

Collaboration & Sharing

8. push_files (experimental, may be disabled)
   Push files in the sandbox to a new GitHub repository.

   • Parameters: repo_name.
   • Response: repo_url, status, stdout or error fields.

9. get_workspace_public_url
   Start http.server + ngrok inside the sandbox to expose /workspace over the internet.

   • Response: url if successful.

---

Prompt

command_help – concise guidance for using run_command including parameters and error semantics.

---

Running the Inspector (Optional)

Optional Requirement (Only for Inspector UI)

If you want to use the MCP Inspector UI for debugging/introspection you need Node.js (tested with >=22). The Python MCP server itself does NOT depend on Node.js.

Quick Start (UI mode)

npx @modelcontextprotocol/inspector

The inspector server will start up and the UI will be accessible at http://localhost:6274.

Configure test connection:

• Transport Type: Streamable HTTP
• URL: http://127.0.0.1:3000/mcp

---

Development

Adding New Tools

To add a new tool, modify main.py:

@mcp.tool(
    title="Your Tool Name",
    description="Tool Description for the LLM",
)
async def new_tool(
    tool_param1: str = Field(description="The description of param1"),
    tool_param2: float = Field(description="The description of param2")
) -> str:
    # The new tool logic
    result = await some_api_call(tool_param1, tool_param2)
    return result

Tests

Run minimal async tests:

uv run test_run_command.py

Adding New Resources

@mcp.resource(
    uri="your-scheme://{param1}/{param2}",
    description="Description of the resource",
    name="Your Resource Name",
)
def your_resource(param1: str, param2: str) -> str:
    return f"Resource content for {param1} and {param2}"

Adding New Prompts

@mcp.prompt("Helpful Prompt")
async def your_prompt(
    prompt_param: str = Field(description="The description of the param for the user")
) -> str:
    return f"You are a friendly assistant, help the user and don't forget to {prompt_param}."

---

Docker Environment

Le Bac à Sable uses a multi-stage Dockerfile:

1. Builder Stage

   • Based on python:3.12-slim-bullseye
   • Installs build dependencies and compiles Python wheels.

2. Runtime Stage

   • Lightweight Python 3.12 slim image.
   • Includes runtime tools:
     • GitHub CLI (gh)
     • ngrok
     • ffmpeg
     • curl, wget, jq, unzip, git
   • Optional Node.js + npm (commented out).
   • Creates a non-root sandboxuser.

Exposes ports:

• 8000 (http.server)
• 4040 (ngrok API)

Default CMD: bash

---

Environment Variables

• NGROK_AUTHTOKEN – Required to use get_workspace_public_url.
• gh-api-token – GitHub API token (injected via headers) for push_files.

---