What it actually does
This teaches harness engineering by building one. The framing in the README is the clearest statement of the idea I have seen: agency comes from model training, not from external code orchestration — but a working agent product needs both. The model is the driver; the harness is the vehicle. This repository teaches you to build the vehicle.
You construct a nano Claude Code from scratch, in Python, understanding each part: the agent loop, tool calling, context management, permissions. "Bash is all you need" is the subtitle, and it is making a point — a surprising amount of what feels like magic is a loop, a tool schema, and careful context handling.
The practical payoff is judgement. Once you have built one, you can evaluate agent tools on their actual architecture rather than their marketing, and you know which problems are hard and which are packaging.
MIT, 74,000 stars, English, Chinese, and Japanese, with only 68 open issues.
Who it's for
- 01
Engineers who want to understand agent tooling rather than just use it
- 02
Teams deciding whether to build or buy an agent harness
- 03
Anyone evaluating agent products who wants to see past the marketing
- 04
Developers moving into agent engineering as a specialism
Where it earns its keep
- Understanding what an agent harness actually does before choosing one
- Building internal tooling that needs a custom agent loop
- Training engineers into agent work with real depth
- Making a credible build-versus-buy decision
- Debugging agent behaviour by understanding the mechanism underneath
Use it, or skip it
Reach for it when
- You are about to make a significant bet on agent infrastructure
- Your team uses these tools daily and understands none of the internals
- You need to build something custom rather than adopt a harness
- You want to evaluate vendors on architecture rather than claims
Skip it when
- You need to ship something now — this is education, not a product
- Your team is happy using existing harnesses and has no build ambition
- You want current best practice rather than first principles
- Nobody has time to work through building a harness from scratch
10 automations
Ideas, not tutorials. Each one is work a team does by hand today.
- 01Founders
Build-versus-buy assessment
Work through it before committing to a harness, then write a grounded recommendation.
- 02Operations
Team depth training
Run it as an engineering learning track so agent tooling stops being a black box.
- 03Engineering
Custom internal harness
Use it as the foundation for a harness tailored to constraints no off-the-shelf tool meets.
- 04Operations
Vendor evaluation rubric
Build an architecture-based checklist for assessing agent products rather than trusting demos.
- 05Engineering
Debugging capability
Give your team enough understanding to diagnose agent misbehaviour instead of filing a ticket.
- 06Engineering
Context management study
Understand how context is assembled so your team can reason about cost and quality together.
- 07Operations
Permission model design
Use the permissions material to design what your agents may and may not do in production.
- 08Operations
Interview material
Base agent-engineering interview questions on real architecture rather than tool trivia.
- 09Operations
Internal architecture talk
Have an engineer present what they built so the whole team gains the mental model.
- 10Engineering
Migration confidence
Understanding the harness layer makes switching between agent tools a smaller decision.
Want one of these running by Friday?
LimeDock builds these as real workflows inside your stack — deployed to your cloud, wired into your Slack and CRM, with the code in your repo. You pay a build fee and your own API keys, nothing else.
Source
Repository stats were read from the GitHub API and reflect the last time we refreshed this entry. The editorial breakdown above is LimeDock’s own analysis — we are not affiliated with shareAI-lab.
https://github.com/shareAI-lab/learn-claude-code