Skip to content
Open Agents Builder

Using the Code Execution tool tutorial

Why code execution?

Open Agents Builder can now write, run, and iterate on code in a Docker sandbox. Shell, Node.js, and Python are pre-installed—but you can add Ruby (or anything else) in minutes.

A code-capable agent lets you:

  • Query live APIs (GitHub, Stripe, …) without exposing keys to the user.
  • Transform user data—convert, clean, visualise.
  • Iterate automatically: When the script fails, the LLM inspects the traceback, patches the code, and re-runs it.

Prerequisites

RequirementNotes
Open Agents Builder accountSign up for the free trial; takes ~3 min
A GitHub repo URLAny public repo works. We’ll use CatchTheTornado/open-agents-builder in this tutorial.
(Optional) Local installClone the project & run it locally if you want to watch container logs in real-time

Creating a new agent

  1. Dashboard → “New Agent…”

  2. Name it → Project condition checker.

  3. Keep Agent type = Smart assistant [Chat].

Writing the system prompt

In AI Prompt tab:

You're a smart, open-source-savvy agent that can help users check their GitHub projects.
You can use the code execution tool to fulfil user requests.
Welcome the user and ask them for the GitHub project URL.
Generate a document with a summary of the last two issues and a general overview.
Generate a PNG diagram with the stargazers' history.

Screenshot for reference:

Tip Keep prompts declarative: describe the workflow & deliverables—let the LLM decide how.

Enabling the Code Execution tool

  1. Go to Tools → Add Tool → choose Code Execution.

  2. No extra config is needed—hit Save.

Under the hood the agent now spins up a disposable Docker container and uses the open-source code-interpreter-tool library to run scripts, capture output, and return any generated files.

Previewing & test-driving your agent

  1. Click Preview.

  2. The agent greets you and asks for a repo URL. Paste https://github.com/CatchTheTornado/open-agents-builder.

  3. Watch the magic ✨

    • The LLM writes a Python script to hit the GitHub API.

    • If a run fails, it parses the traceback, patches the code, and retries (subtitles 00:03:37-00:03:59).

    • When succeeded, it uploads two files:

      FilePurpose
      project_summary.mdIssues overview & repo stats
      stargazers_history.pngAuto-generated Matplotlib chart

    Example chat view:

Heads-up You’ll see Code executed with errors… lines if the first attempt fails—this is expected. The agent will fix & re-run automatically.

Handling generated files

Every chat session stores artefacts in Sessions → select a session → file list appears on the right.

Click ⬇️ to download, or reference them in follow-up prompts, e.g.:

Can you create a .docx that combines the summary and the chart?

The agent will:

  • Import python-docx (or install it on the fly).
  • Embed the PNG and markdown content.
  • Return github_project_report.docx.

Going further

  • Add other languages: extend the Docker image or mount a custom one.
  • Chain tools: enrich with a Browser tool to fetch web pages, or Automations to send scheduled reports.
  • Safety: define rate limits and guardrails in the Safety Rules tab.
  • Template it: when your agent is stable, click “Save agent as template” to share.

Conclusion

You now have a fully-featured GitHub analyst powered by live code execution—no backend coding required. Experiment with different prompts, outputs (PDF, CSV, PowerPoint…), and data sources. The only limit is your imagination.

Happy hacking! 🛠️🧑‍💻