A package manager and sync engine for AI-agent context. Keep rules, skills, and
workflows in a single .agents/ tree — pull them from upstream with a lockfile,
scan the untrusted ones, and sync everything to Claude, Windsurf, Cursor, and Copilot.
brew install brickhouse-tech/tap/sync-agents · go install github.com/brickhouse-tech/sync-agents@latest
// live session
Remote agent context is a supply chain. sync-agents treats it like one — every fetch is SHA-verified, statically scanned, and quarantined until you approve it.
// capabilities
Lockfiles, provenance, integrity hashes, editable installs — the tooling discipline your dependencies get, applied to your agents' brains.
Declare upstream skills, rules, workflows — or whole .agents/ trees — in
sources.yaml. pull installs them reproducibly and records every
resolution in sources.lock. Commit both; clones rebuild byte-for-byte.
Every ref resolves to a commit SHA via the GitHub API, and a deterministic sha256 content hash is verified before anything touches disk. A tampered tarball or corrupted cache aborts the install — no partial writes, ever.
Remote content is treated as a hostile supply chain. Fetches land in
.quarantine/ after a scan for curl | bash patterns, credential
access + network calls, obfuscation, and prompt injection. Critical findings block
approve; overrides are recorded in the lock.
The npm link of agent context. source add --link=../foo-skill
symlinks a live local checkout instead of a snapshot — edits flow both ways, no
commit→push→update loop. Recorded declaratively, always as relative paths.
Drop artifacts into macos/, linux/, unix/, or
windows/ subdirectories and they sync only where the host matches. Your
brew rules stop polluting the Linux homelab, and vice versa.
One generated index of everything in .agents/, symlinked to
CLAUDE.md so Claude reads it natively. watch keeps it fresh,
inheritance links chain project → team → org → global.
Rules, skills, workflows, subagents, plans, specs, hooks, and ADRs — each with its own
lifecycle. Optional buckets activate on demand, and ADR status lives in the directory
tree: proposed/, accepted/, denied/.
A Claude skill is invocable; a Windsurf workflow is a slash command. sync-agents routes on behavioral semantics, not directory names — so nothing gets dumped into another tool's always-on context by accident.
// how it works
A single Go binary (shipped via npm, Homebrew, or go install) with
no runtime dependencies — it doesn't even need git to pull.
Write what your agents should know in sources.yaml — a skill, a rule, or an entire upstream tree.
sources: - skill:anthropic/skill-pack@v1.2.0 - rule:my-org/norms@main/rules/sec.md - tree:my-org/team-agents@v2.0.0
Fetch everything, verify hashes against the lockfile, and review whatever the scanner flags before it can reach an agent.
$ sync-agents pull ✔ 2 installed · 1 quarantined $ sync-agents approve code-review
Regenerate AGENTS.md and symlink every artifact into each tool's directory. Change once, updated everywhere.
$ sync-agents sync ✔ 4 targets linked ✔ AGENTS.md → CLAUDE.md