Quickstart

Setting up is two moves: choose your skills at agentsinc.sh/editor, then run the command it hands you. The editor is where you choose; the CLI is what writes to disk. Nothing reaches your filesystem until that command runs, so the browser half is safe to poke at. This page walks the run and names every file it leaves behind. Read it end to end the first time; after that the only command you need daily is compile.

There may be a shorter route than this one

This page walks the default: open the editor, click a stack, run the command. Two situations have a better answer. Adding this to a codebase that already exists? The stack-detection skill reads your dependencies instead of asking you to retype them. Starting from nothing? The composer takes a sentence describing the project. Both end at the same command this page explains, so nothing below is wasted either way — Four ways in compares all four.

Before you start

You need Claude Code installed. Everything Agents Inc produces is a Claude Code sub-agent file or a Claude Code plugin.

Nothing needs to be installed globally. npx fetches the CLI for you.

Quick start

1. Open agentsinc.sh/editor and click a stack.
2. Press Install at the foot of the right-hand column, and copy the command out of the dialog.
3. Run it. This is the CLI, and it is the part that writes to disk.

npx agents-inc init --ui opens the same page from your shell if you would rather not leave it. It prints the address before it tries a browser, so it also works over SSH and on a machine with no desktop session — but it opens on the bare catalogue and carries nothing with it, so it is a link rather than a different route.

npx agents-inc init --from Ab3xY9_Q

Run it from your project root — the folder holding package.json. Anything you set to project scope is written relative to it; anything global lands in ~/.claude wherever you run from. Global is the recommended starting point — see Global-first setup.

Prefer the terminal?

npx agents-inc init runs the same selection as a wizard. Same catalogue, same output — it skips the id because it is already standing where the files go.

What you’re choosing

Editor

Three columns. The middle one holds the stack grid under choose your stack, and below it every skill in the catalogue, laid out in cells under its domain. The right-hand column is the roster: every sub-agent there is, with the skills each one carries listed underneath.

1. Click a stack. Its skills are selected and handed to the sub-agents that want them, so you are adjusting a starting point rather than filling a blank grid. The first cell is Start from scratch if you would rather not. See Stacks.
2. Click skill cells to add or drop them. The search field and the domain chips above the grid narrow what is on screen; they change nothing about what is selected.
3. Set the two words on a cell if the defaults are wrong: its install mode, plugin or eject, and its scope, project or global. The ••• button opens both as a panel, along with which sub-agents carry the skill. See Install modes.
4. Turn sub-agents on and off in the roster, and set each one’s model and reasoning effort by clicking the words on its row.
5. Press Install and copy the command.

The editor covers the whole screen, including the things it deliberately cannot do.

CLI

The wizard has six steps. You can go back at any point.

1. Stack — pick a pre-built stack, or “Start from scratch”. A stack is a named selection of skills for a common setup; picking one gives you a starting point rather than a blank grid. See Stacks.
2. Domains — which areas you are configuring: web, api, cli, mobile, shared, and so on. This decides which categories the next step shows you.
3. Skills — the skill grid, organised by category. This is the main step: one framework, one styling approach, one test runner, and so on. Categories marked exclusive take a single choice.
4. Sources — where each skill’s files should live: Plugin, registered with Claude Code, or Local, copied into your project. It is set per skill, so a mixed installation is normal. See Install modes.
5. Agents — which sub-agents to compile from what you just selected.
6. Confirm — the summary. Nothing is written until you accept it.

Picking a stack always lands you on Domains. From there, press A on the Skills step to accept the stack’s selections exactly as they stand and go straight to Confirm; Sources and Agents are skipped with it.

Model and reasoning effort are not wizard choices. They are config fields — see Models and effort.

What exists afterwards

Take the wording of the summary the CLI prints seriously — it names the exact directories it wrote to, and where those are depends on the scope you chose.

Your configuration, in .claude-src/:

File	What it is
config.ts	Plain TypeScript, meant to be hand-edited. Your skills, your sub-agents, and the skill-to-agent stack mapping.
config-types.ts	Generated. Type unions narrowed to what you actually installed, so a mistyped skill id is a type error.

At global scope these live at ~/.claude-src/.

Your compiled sub-agents, in .claude/agents/, or ~/.claude/agents/ at global scope — one markdown file per sub-agent. These are build outputs. Editing them directly works until the next compile overwrites them; to change a sub-agent for real, see Customizing subagents.

Your skills, in one of two places depending on install mode:

• Plugin (the default) — registered as Claude Code plugins, under ~/.claude/plugins/ at global scope and .claude/plugins/ at project scope. Nothing is copied into your project source.
• Eject — copied into .claude/skills/, or ~/.claude/skills/ at global scope. Yours to modify.

Mode is chosen per skill, so a real installation is often mixed. See Install modes.

Seeing this before you run anything

Preview generated code, above the editor’s Install button, draws the same file tree with each file’s contents beside it. It is rendered by the same package the CLI compiles with, so the paths are the ones you will get rather than an illustration of them.

The next command

Open .claude-src/config.ts, change something, and run:

npx agents-inc compile

compile is the command you will run most. It rebuilds every sub-agent from the current config and is non-interactive, so it is safe in scripts and CI. Editing your config covers what is worth changing in there — mainly the stack mapping, which decides which skills each sub-agent receives, and whether each one is preloaded or loaded on demand. Every field the file accepts is listed in the config reference.

When something looks wrong

npx agents-inc doctor

Health checks over the installation: config parses, skills resolve, sub-agents compiled, no orphans, marketplace reachable. It exits non-zero if any check fails.

npx agents-inc list shows what is installed. The full set is in the commands reference, and Troubleshooting covers what each refusal means.

Changing your mind

The editor cannot read what is on your disk, so going back to it is a round trip through two commands:

npx agents-inc edit --ui          # publishes this installation and opens it
npx agents-inc edit --from <id>   # applies what you changed there

The second command is the one to get right. The install dialog only ever hands out init --from, and that one refuses a directory that already has an installation — An installation already exists at …. edit --from is the one that accepts a setup already in place. It states a whole roster, so whatever it leaves out is removed, and every removal is named before anything is written.

Doing this from the terminal

npx agents-inc edit reopens the wizard with your current selections loaded, shows a diff at the confirm step, and on confirm reinstalls, rewrites the config and recompiles.

To pull newer skill content from the marketplace, npx agents-inc update. It refreshes every marketplace your config names through Claude Code’s own plugin registry, and that is all it does — it asks nothing, and it recompiles nothing, because a sub-agent references a plugin skill by pointer rather than by copy. Ejected skills are copies you own, so it never touches them.