Dev Agent Setup

A portable, harness-agnostic baseline for AI coding agents.

One engineering standard · one MCP list · one pinned skill set.

Built to run alongside Gentle AI, the harness layer I use — which is also where Context7 and Engram come from.

Works with Pi and Claude Code (the pair I recommend), OpenCode, Codex, Antigravity, Kilo Code, Cursor, Gemini CLI, Oh My Pi, and anything else that reads an AGENTS.md-style rules file and speaks MCP.

English · Español

---

What this is

Not a product and not a framework. It is my personal working setup, cleaned up and shared because it may save you an afternoon. Nothing here is load-bearing for anyone but me — fork it, delete half of it, keep the parts that fit.

Minimal on purpose: three documents, no installer, no vendored skills, no framework to learn. You can read the whole thing in ten minutes.

What you get

Piece	What it is	Where
Engineering standard	One short, enforceable standard — versions, gates, tests, structure, observability, docs — with adoption prompts and the config for every gate. Written for coding agents, so the rules are backed by gates that fail, not by prose.	standards/
MCP servers	The MCP servers I run and what each one is for: Tavily, Context7, Engram — with Exa and Firecrawl as drop-in alternatives to Tavily, and the Playwright MCP as an option next to the playwright-cli skill.	mcps/MCPS.md
Skills	25 agent skills, installed globally from their upstream repos with the skills CLI into ~/.agents/skills and linked into ~/.claude/skills for Claude Code.	skills/SKILLS.md

Used with Gentle AI

This repo is one half of the setup; Gentle AI is the other. Gentle AI is the harness layer over Pi: it plans, delegates to focused subagents, reviews its own changes, and ships its own skills on top of the twenty-five listed below. Into Pi it goes with pi install npm:gentle-pi.

It is also where Context7 and Engram come from — that is why mcps/MCPS.md lists them without asking you to configure them by hand.

The two are independent. Everything here is plain documents, so you can take the standard, the MCP list and the skills and skip Gentle AI entirely.

Layout

.
├── README.md               this file (en)
├── README.es.md            this file (es)
├── AGENTS.md               rules for working on this repo (not the standard you install)
├── LICENSE                 MIT
├── standards/
│   ├── README.md           the standard, index (en)
│   ├── README.es.md        the standard, index (es)
│   ├── en/
│   │   ├── Engineering_standards.md
│   │   ├── adopt.md
│   │   └── examples/       AGENTS.md, README.md, an ADR, gate config, CI
│   └── es/                 the same, in Spanish
├── mcps/
│   └── MCPS.md             the MCP servers I run
└── skills/
    └── SKILLS.md           install commands and inventory

How to use it

1. The engineering standard

Copy standards/en/Engineering_standards.md into your repo (root or docs/) and hand your coding agent the matching prompt from standards/en/adopt.md: it configures the gates for Python (uv) or Node/React (pnpm) and creates a short README.md and AGENTS.md. An existing repo gets a read-only audit first, then a retrofit with one commit per step.

Start at standards/README.md.

2. MCP servers

mcps/MCPS.md lists the servers and the canonical mcpServers block. Translate it to whatever shape your harness expects: every server is a plain npx stdio process, so only the JSON and TOML keys differ.

3. Skills

Skills are not vendored here. They are installed globally from their upstream repos with the skills CLI. It writes to ~/.agents/skills/, the directory Pi, OMP and most other harnesses read, and links each skill into ~/.claude/skills/, the only one Claude Code reads:

# run the commands in skills/SKILLS.md, then:
npx skills ls -g
npx skills update -g -y

See skills/SKILLS.md for the command set and the inventory.

Quick start (let an agent do it)

Open any coding agent — in any directory — and paste:

Set up this machine from https://github.com/umi008/dev-agent-setup (read its files over the network;
raw.githubusercontent.com works well). Do exactly this:

1. Install the global skills by running every command in skills/SKILLS.md verbatim, then run
   `npx skills ls -g` and show me the result.
2. Read mcps/MCPS.md and tell me, for each server, how to declare it in the harness I am using.
   Inspect the environment only: never print secret values, never write them to a file.
3. Report what you installed and what I still have to configure by hand.

Constraints: use the published files as-is — do not modify the repository, do not invent extra files.

Credentials

Nothing secret lives in this repo. Provide these yourself:

Credential	How
Your model provider	Whatever your harness expects — OAuth login (/login in OMP, claude login, codex login) or an API key such as ANTHROPIC_API_KEY / OPENROUTER_API_KEY exported from your shell rc.
Tavily (web search)	export TAVILY_API_KEY=... in your shell rc. The MCP entry reads it from the environment — no key is stored here. Skip the server entirely if your harness already talks to Tavily natively.
Exa or Firecrawl (only if you swap Tavily)	export EXA_API_KEY=... or export FIRECRAWL_API_KEY=... in your shell rc. Exa's hosted server also runs keyless, with no key at all.

Context7 and Playwright need no credentials.

Notes

• Editing this repo? AGENTS.md at the root holds the conventions. It is not the standard you install — that one is standards/en/Engineering_standards.md.
• Claude Code and AGENTS.md. Claude Code reads AGENTS.md since v2.1.277, but only when the project has no CLAUDE.md. If it has one, add the line @AGENTS.md to it.
• Skills stay upstream. npx skills update -g keeps them current instead of freezing a copy here.

License

MIT — see LICENSE. Take it, fork it, strip it down.