Anthropic's ant apply turns Claude agents into version-controlled code
The ant CLI's new apply command syncs agents, skills, memory stores and cron deployments from repo files — plan-and-approve, with a lockfile.
Copy markdown
Your whole agent stack, declared as files
`ant apply` reconciles five resource types — agents (Markdown), environments, skills (SKILL.md bundles), memory stores and scheduled deployments — against the Claude API. Point it at a directory and it creates them in dependency order, swapping real resource IDs in for file-path references so a coordinator agent can wire up its sub-agents and skills by relative path.
Plan, approve, lock — then prune
Every run prints an interactive plan of create/update diffs, and on approval writes `claude-lock.json` (resource IDs plus local and remote hashes) so a second run never duplicates and an out-of-band Console edit blocks the apply until you pass `--force`. `--dry-run` surfaces the plan in a pull request, `--yes` auto-approves in CI, and `--prune` deletes resources you've removed from the repo.
Ship this weekend: a self-updating agent fleet
Commit an agents/ + skills/ + deployments/ tree, run `ant apply --dry-run` on PRs and `--yes` on main, and a deployments/*.md with a cron block becomes a scheduled run — e.g. "review any open PRs at 3am, oldest first." It's GitOps for Claude agents: every prompt, skill and schedule reviewed like normal code.
The catch: it only manages what it made
`ant apply` can't adopt agents you built in the Console or CLI — only what's in the lockfile is tracked — and renaming a file spins up a new resource, leaving the old one until `--prune`. There's no lock on the lockfile, so run one apply at a time, and Anthropic recommends Workload Identity Federation over stored API keys in CI.