Skip to content
zedkit

Set up your coding agents

Install @zedtech/zedkit-ai, run init for your editor, validate the kit layout, and undo changes — independent from the UI package.

Every command on this page is run end-to-end in a scratch directory by automated checks.

This guide is optional: you can use @zedtech/zedkit-ui or the core, server, and nest packages without installing the AI kit.

Who this is for

You use Cursor, Claude Code, Codex, or similar tools and want a repo-local copy of skills, subagents, commands, and AGENTS.md without maintaining separate trees per vendor.

Prerequisites

  • Node.js 22+ and a git repository (any stack).
  • One provider for init: cursor, claude, codex, muse, or all (see zedkit-ai init --help in packages/ai/src/cli.ts).

Install

First-time onboarding is npx @zedtech/zedkit-ai setup (preset setup, adopt on). This tutorial uses init so you can pick a preset explicitly. If you already installed 1.1.x, re-run setup (adopt keeps your files).

In your repo root:

pnpm add -D @zedtech/zedkit-ai

Steps

1. See what the kit ships

npx @zedtech/zedkit-ai list

You should see preset names, subagents (for example explorer), and commands (for example review) printed on stdout — same checks as packages/ai/tests/unit/cli.test.ts (list).

2. Copy the kit into the repo

npx @zedtech/zedkit-ai init . --provider cursor --preset full

init never overwrites existing files; it skips paths that already exist (install.ts). This tutorial uses --preset full so every bundled command (for example audit) has its skill on disk — validate checks that link (kit.json). For Cursor it adds .cursor/rules/project.mdc and copies skills into .agents/.

Preview without writing:

npx @zedtech/zedkit-ai init . --provider cursor --dry-run

Stdout ends with dry run: nothing written and no new files on disk.

3. Validate

npx @zedtech/zedkit-ai validate

Expected result: exit code 0 and a single line on stdout:

ok — structural kit checks only (skills, agents, commands, wiring, tokenomics, atlas paths); browser rendering, focus, hosting and on-target behavior are NOT checked by validate

If skills, commands, or provider copies drift, you get path: message lines and a non-zero exit (see validate in cli.ts).

validate checks file shape and drift — it does not run your coding agent. After init, use your tool’s normal workflow to open the repo; the wired rules and skills are the files init created.

4. Optional project instructions

Ask your agent to run the bundled setup-ai skill (copied under .agents/skills/) so AGENTS.md reflects this repository. That step is interactive and is not part of CI.

5. Hand it a bigger task

When the work spans several steps or review lenses, use /orchestrate with the goal in one line (for example /orchestrate add auth and document the API). The orchestrator agent frames the work, dispatches specialists, and returns one aggregated report without editing files itself.

Scenario — repo that already has AGENTS.md

You already maintain team rules at the root:

# Team rules

Use our internal style guide.

Run the same init command. When AGENTS.md exists but lacks the kit ladder section, init appends the ladder instead of replacing your text (install.ts). Run validate again; stdout should still start with ok — structural kit checks only. Automated checks cover this flow.

What gets installed

After npx @zedtech/zedkit-ai init (the install command on Getting started), you get a project-owned tree the agent tools read:

AGENTS.md                 # root instructions (template or appended ladder)
CLAUDE.md                 # symlink to AGENTS.md
.agents/
  skills/                 # bundled skills (including setup-ai)
  agents/                 # subagent briefs (including orchestrator.md)
  commands/               # slash-command sources (including orchestrate.md)
  personas.md
.cursor/
  rules/project.mdc       # always-on Cursor rules
  commands/setup-ai.md    # runs the setup-ai skill
  skills/                 # symlink into .agents/skills
.claude/
  commands/setup-ai.md
  skills/                 # symlink into .agents/skills

init logs every path it creates. It never overwrites an existing file.

Undo

init only adds or appends files. To remove the kit:

  1. Delete .agents/ if you no longer want project-owned skills and commands.
  2. Remove provider wiring you do not need (for example .cursor/, .claude/, or symlinks listed in the init log).
  3. Revert changes to AGENTS.md with git if you only wanted the ladder append temporarily.

There is no uninstall command; keep a commit before init so you can roll back.

Troubleshooting

SymptomWhat to check
validate prints problemsRun from the repo root; paths are listed as relative/path: message. Fix or restore the file, or re-run init on a clean branch.
init skips filesExpected when paths already exist — remove or rename the conflicting file, or use --dry-run to preview.
Unknown providerUse cursor, claude, codex, muse, or all only (cli.test.ts rejects unknown providers).
Agent ignores instructionsOut of scope for validate — confirm your tool loads .cursor/rules or the provider copy init created.

Next steps

On this page