Loading posts
5 Repos Seniors Run Daily: Zero Agent Reads, Steal the Setup, and What Nobody Teaches
The five open-source repos our senior software engineers actually have installed — what each one fixes, what each one costs you, how they cover the whole vibe coding loop, and one setup guide for all five, in order.
12 min read
Every few weeks someone publishes a list of a hundred AI coding tools. This is not that list. These are the five GitHub repositories the senior software engineers on our team have installed right now, what each one changed, what each one costs, and the order to install them in.
They were not picked as a set. Our senior engineers found them separately, over about four months, each one solving something specific that kept going wrong. What makes them worth writing up together is what happened afterwards: between them they cover almost every point where vibe coding falls apart, and they barely overlap.
All five are free and open source — four MIT, one Apache-2.0. Star counts below were read from the GitHub API on 13 August 2026. They measure attention rather than fit, and they are quoted here only to establish that these are maintained projects rather than weekend experiments.
The four things a coding agent does not bring with it
A coding agent arrives fluent and empty. It writes better TypeScript than most people our senior engineers have worked with, and it knows nothing about four things that decide whether the output is any good:
- Context — what is already in your repository. Out of the box it rediscovers your codebase by
grepand file reads, every session, from zero. - Taste — what good looks like. Left alone it produces the statistical mean of its training data, which is the purple-gradient SaaS landing page you have seen four hundred times.
- Depth — the specific judgement of a security reviewer, a Terraform engineer, an accessibility auditor. It is a generalist everywhere and a specialist nowhere.
- Process — a repeatable method. Ask twice, get two different approaches, neither of them the one that worked last time.
One repo per gap, and design splits in two — inside the code and outside it. That is the whole set.
1. CodeGraph — the agent stops reading your files
CodeGraph (66.2k stars, MIT) builds a knowledge graph of every symbol, call edge and dependency in a repository, then serves it to your agent over MCP. A Rust kernel parses 30-odd languages with tree-sitter into a SQLite database, and native OS file watchers keep it in step as you type.
The effect is the one our senior engineers did not expect from anything in this category: the agent stops opening files. Instead of grep, glob and eleven reads to work out how locale negotiation reaches the Cloudflare Worker behind this site, it asks the graph once and answers. Its own benchmark — seven open-source codebases, seven languages, median of four runs per arm, with the CodeGraph CLI blocked in both arms so the control is a real control:
- 88%
- fewer tool calls to answer one architecture question CodeGraph benchmark, re-measured 2026-08-05
- 62%
- fewer tokens processed to reach the same answer CodeGraph benchmark
- 0files
- read by the agent — on all seven benchmark repositories CodeGraph benchmark
2. UI/UX Pro Max — taste as a lookup table
UI UX Pro Max (116.2k stars, MIT) is a skill, not a component library. It carries 84 UI styles, 192 colour palettes, 74 font pairings, 25 chart types, 98 UX guidelines and 161 reasoning rules. Its v2 flagship reads a brief and returns a complete design system — layout pattern, section order, palette, type pairing, and the anti-patterns to avoid for that particular industry.
What it changes is the starting point. The agent no longer reaches for Inter and a violet gradient because that is the average of everything it has seen; it picks from a catalogue with reasons attached, and it can tell you why. Our senior engineers run it at the start of any interface work, before a line of JSX exists.
3. Open Design — the design work that is not code yet
Open Design (85.4k stars, Apache-2.0) is the open-source answer to Claude Design: a local-first desktop app and daemon that turns whichever coding CLI you already have into a design engine. It ships 151 brand-grade design systems, 100+ functional skills and 277 plugins, runs on 25 different agent CLIs, and exports to HTML, PDF, PPTX and MP4.
Two things earn its place. The first is DESIGN.md — a brand contract as a Markdown file that lives in the repository, gets reviewed like code, and that the agent reads before rendering anything. The second is everything that is not the product: the pitch deck, the social card, the client-facing prototype. Those are the jobs that used to pull a senior engineer out of the editor for a whole afternoon.
4. VoltAgent's subagents — depth, in its own context window
awesome-claude-code-subagents (24.3k stars, MIT) is 158+ role-specific subagents across ten categories — language specialists, infrastructure, quality and security, data and AI, meta-orchestration. Each is a Markdown file whose front matter names its tools and its model tier, so a documentation agent gets haiku and read-only tools while a security audit gets opus.
The mechanism that matters is the isolated context window. A penetration-test pass or a dependency audit generates an enormous amount of intermediate noise, and running it as a subagent means that noise never lands in the conversation you are actually trying to hold. You get the conclusion instead of the search.
5. claude-skills — the method, written down
claude-skills (24.4k stars, MIT) is 362 skills across 18 domains, plus 102 agents, 116 commands, 741 reference templates, and 644 Python CLI scripts that use the standard library only — no pip install, which is the detail that makes them safe to run anywhere.
A skill is a procedure: a SKILL.md carrying the steps, the decision points and the checks. The value is not that the agent cannot do an SEO audit or write a Playwright suite unaided — it is that it does it the same way every time, and the way is a file on disk you can edit. This is the layer that turns a good session into a repeatable one.
How the five cover the loop
Laid against the real stages of building something with an agent, the coverage is close to complete and the overlap is small. That is the argument our senior engineers make for running them together rather than picking a favourite.
Where each repo lands in the build loop
| Stage | What goes wrong unassisted | Repo | What changes |
|---|---|---|---|
| Understanding existing code | grep, glob and a dozen file reads per question — and dynamic dispatch it cannot follow at all | CodeGraph | One graph query; zero file reads across all seven benchmark repos |
| Deciding how it should look | The statistical average of every landing page on the internet | UI UX Pro Max | 84 styles and 161 reasoning rules, selected against your brief |
| Design that is not code | Decks and prototypes by hand; brand rules living in somebody's head | Open Design | DESIGN.md as a reviewable contract — decks, prototypes and video from it |
| Specialist work | One generalist prompt covering security, Terraform and accessibility alike | VoltAgent subagents | 158+ role prompts, each in an isolated context window |
| Doing it the same way twice | A freshly improvised method every session | claude-skills | 362 written procedures and 644 dependency-free scripts |
The gap this set does not close is review. Nothing here tells you whether what you shipped is safe to put in front of customers — that is a different problem, and the one our first post is about.
Installing all five, in order
Order matters for exactly one reason: CodeGraph changes the shape of every interaction that follows it, so it goes first. The rest are additive and can be installed in any order you like.
bash
npm i -g @colbymchenry/codegraph
codegraph install # wires the MCP server into every agent it detects
cd your-project
codegraph init # creates .codegraph/ and builds the graph in one step
codegraph status # confirm it indexed what you expected1 — code intelligence. Once per machine, then once per project.
powershell
irm https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.ps1 | iexNo Node on the machine? The bundled installer brings its own runtime.
bash
npm install -g ui-ux-pro-max-cli
uipro init --ai claude --global # installs to ~/.claude/skills/2 — design intelligence, global so every project inherits it.
/plugin marketplace add alirezarezvani/claude-skills
/plugin install engineering-skills@claude-code-skills
/plugin marketplace add VoltAgent/awesome-claude-code-subagents
/plugin install voltagent-core-dev@voltagent-subagents3 and 4 — skills and subagents, typed inside Claude Code.
bash
od mcp install claude # --print for a dry run, --uninstall to remove5 — design artifacts. Install the desktop app first, then wire the server.
- Open a new terminal after step 1. The installer puts
codegraphon yourPATHbut does not modify the shell you ran it from. - Restart your agent after anything that touches MCP configuration — that is steps 1 and 5.
- Ask a structural question — "what would break if I changed this function" — and watch the transcript for
codegraph_calls. If you seeGrepinstead, the graph is not wired in. - Run
codegraph initin every repository, not just the first. The CLI is global; the index is per project.
What running all five actually costs
Nothing in money. Three things in other currencies, and they are worth knowing before you install rather than after.
- Context budget. Every skill, subagent and MCP server adds description text the model weighs before it starts working. Five well-chosen categories beat five hundred available ones. The catalogue counts above are a menu to shop from, not a target to hit.
- Trust. These repositories inject instructions your agent will follow and, in two cases, scripts it will run. Popular and MIT-licensed is not the same as audited. Read the
SKILL.mdfiles you install, and treat a marketplace update like any other dependency bump. - Drift. Marketplaces update underneath you and skills change behaviour between versions. When an agent that worked last week starts producing something different, check what updated before you rewrite your prompt.
Common questions
Do I need all five to get value?
No. If you take one, take CodeGraph — it improves every other tool in the list by making the agent's picture of your repository correct. The design pair only pays off if you build interfaces; the skills and subagents only pay off once you have workflows worth repeating.
Do these only work with Claude Code?
No. CodeGraph wires into Cursor, Codex, opencode, Gemini CLI, Copilot and others from a single codegraph install. UI UX Pro Max ships an installer for around twenty assistants. claude-skills converts to nine more formats via ./scripts/convert.sh. Open Design supports 25 agent CLIs plus any OpenAI-compatible endpoint. Only VoltAgent's subagents are genuinely Claude Code-shaped — though the files are plain Markdown, so porting one is a copy rather than a rewrite.
Won't 362 skills and 158 subagents slow the agent down?
Yes, if you install them all — and that is the single most common mistake with both repositories. Both are organised into per-category plugins precisely so you do not have to. Install the domains you work in and add more when a task actually demands one.
Does CodeGraph send my code anywhere?
The index is a SQLite file in .codegraph/ inside your project, and parsing happens locally. The project does collect anonymous usage telemetry — which commands and languages, and explicitly no code, paths, symbol names or queries — and the installer asks about it up front. codegraph telemetry off, DO_NOT_TRACK=1 or CODEGRAPH_TELEMETRY=0 each disable it.
We already pay for Figma. Is Open Design a replacement?
Not for a design team working in a canvas. It replaces the part where a developer needs a deck, a social card or a branded prototype today and no designer is free. DESIGN.md is the piece worth adopting even if you keep Figma — a brand contract that lives in the repository, gets reviewed in pull requests, and that the agent actually reads before it renders.
How current is this?
Every count, command and licence above was checked against the repositories on 13 August 2026, and the star counts came from the GitHub API on the same day rather than from the READMEs — two of which quote figures their projects have long since outgrown. All five had commits within that same week. This space moves fast enough that you should treat the numbers as a snapshot.
This set makes building faster. It does not tell you whether what you built is safe to launch — if that is the question you are sitting with, that is the review our senior engineers do by hand.