11 Repos That Changed How I Ship Code With AI Agents

A map of the eleven open-source repos behind the way I work with coding agents now — grouped into the three things they actually fix: how the agent works, what it knows, and how much of the loop you have to crank yourself.

Over the last few months a handful of open-source repos quietly changed how I work with coding agents — less typing prompts into a chat box, more building a system that does real work. I went deep on each in a three-part hands-on series; this is the map, with a one-line quick-start for every repo. The deep dives have the full install-and-use walkthrough.

Eleven repos, three problems. Not “eleven cool tools” — three things a coding agent gets wrong, and the repos that fix each.

1. How the agent works → skills

A skill is a reusable instruction file that teaches the agent how to do a recurring task — your way, once, instead of re-explaining it in every prompt.

mattpocock/skills — Matt Pocock’s own skills; TypeScript taste from someone who’s argued it in public for years. → npx skills@latest add mattpocock/skills
addyosmani/agent-skills — a production-grade skillset focused on process: plan, verify, stay surgical. → /plugin marketplace add addyosmani/agent-skills
forrestchang/andrej-karpathy-skills — first-principles, build-the-simple-version-first thinking, as a CLAUDE.md. (Community-made, named after Karpathy — not by him.) → /plugin marketplace add forrestchang/andrej-karpathy-skills
nextlevelbuilder/ui-ux-pro-max-skill — makes a backend-heavy agent reason about real user needs, not just pretty screens. → npm i -g ui-ux-pro-max-cli && uipro init --ai claude

Deep dive: Skills Worth Stealing →

2. What the agent knows → context engineering

An agent forgets what it learned last session and over-reads to answer small questions. Both are on your token bill. These four index it, remember it, and compress it.

DeusData/codebase-memory-mcp — a local, auto-syncing knowledge graph of your codebase the agent queries instead of re-scanning. → curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash
safishamsi/graphify — the same idea, widened: code, SQL schemas, docs, even media in one queryable graph. → uv tool install graphifyy && graphify install
thedotmack/claude-mem — persistent memory so the next session starts warm, not cold. → npx claude-mem install
headroom — compresses what the agent reads before it hits the model; 60–95% fewer tokens (verify on your workload). → pip install "headroom-ai[all]" && headroom wrap claude

Deep dive: Context Engineering →

3. How much you crank yourself → loop engineering

Stop being the runtime. Instead of prompting the agent, design the loop that prompts it for you — with gates and handoffs so you can trust it unattended.

cobusgreyling/loop-engineering — ready-made loops (triage, PR babysitter, CI sweeper) plus CLIs for cost estimation and readiness audit. → npx @cobusgreyling/loop-init . --pattern daily-triage --tool claude
Panniantong/Agent-Reach — cross-platform search (Twitter, Reddit, YouTube, GitHub) from one CLI, no API fees — the ingestion step that feeds a loop. → pip install agent-reach && agent-reach install --env=auto
bytedance/deer-flow — a heavy autonomous engine (self-hosted, Docker, 8 GB RAM) a loop can dispatch and collect from; runs multi-step work for minutes to hours. → git clone … && make setup

Deep dive: Loop Engineering →

The one idea under all eleven

If you only take one thing from the whole list, take this: let the pain pick the tool.

That’s the whole thing I’d want a younger version of myself to hear. The tools are good. But the leverage isn’t in adopting all eleven — it’s in noticing, honestly, where your own workflow hurts, and reaching for exactly the layer that fixes it. The same instinct that makes a good data platform makes a good agent setup.

Start with the skills post, and pick the one chore you’re tired of doing by hand.