Initialize a rig

cd path/to/agent-workspace
rig init

By default this is an interactive wizard. It scans the folder, infers most of the manifest, and asks you to confirm or fill in the rest.

What gets inferred

From	Into
Folder name	`[rig].name` (you can change it)
Your git config	`[rig].author` (asked, you can override)
`.mcp.json`	`[mcp].config` + `[mcp].servers`
`.env.example` placeholders	`[env].required` + `[env].template`
Shell scripts, `package.json`, etc.	`[tools].required`
`.claude/settings.json` permissions	`[permissions].allow` / `[permissions].deny`
`.claude/settings.json` plugin hints	`[plugins].required`
Folder shape (data/, scripts/, …)	`[package].include` / `[local].dirs`

You're shown each inference and asked whether to accept it. Accepting all defaults is fine; you can refine later by editing rig.toml directly.

What gets written

rig.toml at the workspace root.
.claude/skills/rig/SKILL.md — the skill that teaches Claude Code in this workspace how to interpret rig doctor and repair gaps.
.claude/CLAUDE.md — minimal hint pointing Claude at the skill (created if absent; otherwise left alone).
.rigignore placeholder (only if you ask for one) — additive excludes beyond the defaults.

Nothing else. rig init is read-mostly: it doesn't install MCP servers, doesn't write env values, doesn't create folders. Setup happens in rig install.

Flags

Flag	What
`--force`	Overwrite an existing `rig.toml`. Otherwise refuses.
`--author <name>`	Set author explicitly (skips the prompt).
`--json`	Machine-readable summary of what was written.
`--plain`	Non-interactive defaults.

After init

Read the generated rig.toml. Then:

Decide what gets shared and opt files in as needed.
Pack and inspect.
Publish.

Making it collaborative later

Add a [sync] section to rig.toml (see collab manifest) and run rig share --enable. You don't need to re-init the rig.