Import your existing AI config into a hub

You already have AI config. A CLAUDE.md that took months to get right, a .cursorrules file someone tuned last spring, Copilot instructions, a handful of skills. Import brings that content into your hub without asking you to rewrite anything and without touching the originals.

The flow is scan-first: AgentBoot finds candidate content, classifies it, shows you the full plan, and writes nothing until you approve it.

What the scanner detects

agentboot import scans a repo (or a whole directory of repos) for:

Location	Detected as
.claude/ directory contents	skills, agents, rules, settings.json, .mcp.json
Root CLAUDE.md	project instructions
AGENTS.md — root and nested	project/package instructions (the industry-standard file; nested files are the spec's monorepo pattern, so the scanner walks the tree for them, skipping node_modules/build/vendored dirs)
.cursorrules	Cursor rules
.github/copilot-instructions.md	Copilot instructions
.github/prompts/*.prompt.md	Copilot prompt files
skills/<name>/SKILL.md	Agent Skills (agentskills.io convention)
Any .md with paths: frontmatter	path-scoped rule (gotcha candidate)

Binary files and symlinked directories are skipped — the scanner never follows a symlink out of the repo.

The CLI flow

# Scan the current repo
agentboot import

# Scan a specific repo, or every repo under a parent directory
agentboot import --path ~/work/auth-service
agentboot import --parent ~/work/ --hub-path ~/work/personas

# Review the generated plan, then apply it
agentboot import --apply

Import is a two-step commit: the scan produces a plan file (.agentboot-import-plan.json) describing what would be created — types, names, target paths. Nothing lands in the hub until you run --apply.

Useful flags:

Flag	What it does
--non-interactive	Auto-apply items the classifier marks high confidence (categorical); medium/low are left for review
--overlap	Heuristic overlap analysis — flags imports that look like near-duplicates of existing hub content
--retry-failed	Retry files that timed out on a previous run
--isolated	Classify using a temporary Claude config, leaving your personal settings untouched

Classification is LLM-powered (it runs claude -p under the hood), so import requires an active Claude Code login. The deterministic parts — scanning, planning, applying — don't.

Import from a GitHub URL

agentboot import --url https://github.com/some-org/some-repo

--url accepts a repo URL (shallow-cloned) or a raw/blob file URL (fetched, with a size cap). Either way the content lands in a temp directory, and AgentBoot prints the agentboot import --path <tempdir> command to classify it. Download and classification are deliberately separate steps — you can look at what was fetched before any LLM reads it.

The /ab flow

Inside Claude Code, /ab drives the same pipeline conversationally:

1. Scan — you name the repos; the agentboot_scan_for_import MCP tool returns two buckets: highConfidence (recognized file types — CLAUDE.md, skills, rules, Cursor/Copilot files) and items that need a human look.
2. Review the uncertain items first, one at a time, with an excerpt and a proposed type. Accept, reclassify, rename, or skip.
3. Confirm the confident batch as a table — proceed with all, or review individually.
4. Duplicate check — each item is compared against existing hub content; near duplicates are flagged with the choice to import as new, update the existing artifact, or skip.
5. Pattern promotion — if essentially the same content shows up in three or more repos, /ab suggests promoting it to core scope so there's one maintained version.
6. Apply as pull requests — every write goes through a PR, so the import is reviewed like any other change. Each imported artifact carries attribution frontmatter (contributor and source repo) so credit and provenance survive the move.

Limits, honestly

• Classification is a draft, not a verdict. The LLM proposes persona / trait / gotcha / instruction; you confirm. Expect to reclassify some items — that's what the review step is for.
• Originals are never modified or deleted. Import copies content into the hub; cleaning up the source repos afterward is your call (and usually a follow-up PR).
• It imports what exists, as it exists. Exact-duplicate content across repos is detected and converged onto one promoted org artifact with every source attributed (see cross-repo promotion in CLI Reference § agentboot import). What import won't do is merge similar-but-different wording into one rule or improve prose. --overlap flags likely near-duplicates; deciding is on you.
• AGENTS.md needs the explicit route described above.

Import and the first sync

Import is also the recommended first step before syncing to a repo that already has hand-written agent config. A first agentboot sync onto a repo with pre-existing instruction files (CLAUDE.md, AGENTS.md, .cursorrules, .github/copilot-instructions.md) stops with an error rather than replace them — import first so the content lives on in the hub, then sync. To replace without importing, pass sync --adopt-existing: everything the sync overwrites, including root-level artifacts, is archived to the target directory's .agentboot-archive/ first, and agentboot uninstall restores it. See CLI Reference § agentboot sync.

When not to use this

If you have one repo with one CLAUDE.md that works, leave it alone — a checked-in file needs no hub. Import earns its keep when the same hard-won rules live in several repos or several tools' formats and you want one maintained source instead of N drifting copies.