Get started

Connect Console to an agent, ask it to run a small analysis, then follow up using the same data. This walkthrough uses Python; you don’t need to install R or create a Console configuration file.

Install

Install uv to run Console with uvx, without a separate Console installation. Console prepares Python and its packages for you, so the first setup can take longer while they download.

If you also want to use R, install it separately. Install and connect covers supported platforms and runtime requirements, including the Windows sandbox prerequisite.

For a development checkout, see Build from source.

Connect a client

Codex and Claude Code

Run the command for your client:

uvx mcp-console configure-client codex
uvx mcp-console configure-client claude

This makes Console available across projects and allows the agent to call its send tool, subject to explicit client restrictions. Setup checks the runtimes and prepares caches in the background, printing a log path for progress and errors. Restart your client after setup so it loads the new tool. MCP clients covers project registration and other setup options.

Other MCP clients

In your client’s MCP settings, use command uvx and arguments ["mcp-console", "serve"]. See MCP client setup for a configuration example.

R and Python packages

Follow Use Console from R to add a Console tool to an ellmer chat, or Use Console from Python for direct calls and model framework integrations. Both guides include client package installation.

Once connected, the agent can run code in a persistent Console session.

Try an analysis

Ask your agent:

Use MCP Console to simulate rolling two dice 1,000 times in Python. Plot the distribution of their sums and keep the simulated rolls for follow-up.

The agent can run the simulation and inspect the plot in Console. The simulated rolls remain in the session after it responds to you. Follow up with:

Compare those results with the expected probabilities. Use the rolls already in memory, without simulating them again.

The agent can calculate another summary or revise the plot using the existing data. Sessions and languages shows how this also works across R, Python, and SQL.

This example works in memory and does not need to write files in your project. The built-in default sandbox restricts project writes and direct network access, but allows reading files your account can read, including readable secrets. Client setup and configuration can select different permissions; Codex setup defaults to inheriting the client’s permissions. To let the agent save files in a project, follow Give a session the access it needs.

Review the work

Ask the agent:

Show me the code you ran and the path to the recorded session.

Open transcript.md in that directory to review the calls, results, and links to plots. The recording stays on the machine where Console runs after the live session ends. It contains code, input, and output without redaction, so review it before sharing.

The recording does not preserve live objects for a later session. Restarting Console clears the simulated rolls and other objects in memory. Recordings and reports explains how to inspect saved output and reuse the code.

Continue with Sessions and languages to learn how the agent builds on earlier work and combines R, Python, and SQL.