Build Claude Code Harness using CrewAI

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.

📬 Stay Updated with Our Newsletter!

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