This project rebuilds popular coding-agents (Claude Code, etc.) harness from scratch: that explores, edits, tests, and reports on a real bug-fix task, with planning, memory, checkpointing, a sandbox, and human-in-the-loop approval built in one layer at a time.

E2B is used for sandboxed shell and Python execution.
CrewAI to build the hierarchical Agentic workflow.
OpenRouter, as the underlying LLM provider.

Setup and installations

Get an E2B API Key:

Go to E2B and sign up for an account.
Create a new API key from your dashboard.
Store it in the .env file (after renaming .env.example to .env).

E2B_API_KEY="..."

Get an OpenRouter API Key (or any LiteLLM-supported provider key):

Go to OpenRouter and sign up for an account.
Create a new API key.

OPENROUTER_API_KEY="..."
MODEL="openrouter/anthropic/claude-sonnet-4-6"

Get an OpenAI API Key (for memory only, not for the agents or the planner):

memory=True is on for this crew, and CrewAI's memory system needs an embedding model to turn text into vectors before it can save or recall anything.
By default that embedder is OpenAI's text-embedding-3-large, regardless of which provider the agents themselves run on, so this key is required even though every LLM call elsewhere in the project goes through OpenRouter.
Go to OpenAI and create a key.
If you'd rather not add a second provider, point the crew at a different embedder instead (embedder={"provider": "ollama", ...} is a documented CrewAI option in memory), or turn memory=True off.

OPENAI_API_KEY="..."

Install Dependencies: Ensure you have Python 3.11 or later and uv installed.

uv sync

Run the project

Finally, head over to this folder:

cd build-code-harness

and run the project by running the following command:

uv run python code_harness.py

The task has human_input=True, so the run will pause once it has an answer and wait for your approval on the terminal before finishing.

checkpoint=True writes progress to ./.checkpoints/ after each completed task; safe to delete between runs, and worth adding to .gitignore.

Sample Test

The demo repo in workspace/ ships with two real bugs (an overdraft check missing from withdraw, and transfer crediting the wrong account) and a pytest suite that starts at 3 failing / 2 passing. A full run explores the code, fixes the implementation only, and drives the suite to 5 passing.

Get a FREE Data Science eBook 📖 with 150+ essential lessons in Data Science when you subscribe to our newsletter! Stay in the loop with the latest tutorials, insights, and exclusive resources. Subscribe now!

Contribution

Contributions are welcome! Please fork the repository and submit a pull request with your improvements.

README.md

Build Claude Code Harness using CrewAI

Setup and installations

Run the project

Sample Test

📬 Stay Updated with Our Newsletter!

Contribution