Building Claude Code from Scratch

I'm learning agentic engineering by building a small Claude Code clone through the CodeCrafters "Build Your Own Claude Code" challenge.

The basic idea: send a prompt to an LLM, let the LLM call tools like reading a file or running a shell command, run those tools, send the results back, and keep looping until the model gives a final answer.

What it can do right now

• Read a file
• Write a file
• Run a bash command
• Load skills from .claude/skills, either by typing /name or by letting the model pick one
• Run a skill in its own subagent when the skill sets context: fork
• Loop through tool calls until the model stops asking for them

The tool names the model sees are Read, Write, RunBashCommand, and Skill (only when there is at least one skill).

Running it

To be honest, I mostly run this through the CodeCrafters tests and have not spent much time running it locally. If you want to try it locally:

1. Install Go 1.25.
2. Get an API key from OpenRouter.
3. Set OPENROUTER_API_KEY. You can also set OPENROUTER_BASE_URL if you want to point at a different OpenAI compatible endpoint. It defaults to https://openrouter.ai/api/v1.
4. Run:

./run.sh -p "your prompt here"

I have not added a .env loader, so the key needs to be in your environment. Don't commit your key.

Be a little careful with this: the agent can write files and run bash commands. Try it in a scratch directory and don't put secrets in your prompts.

Tests

go test ./...
go vet ./...