The Forward-Deployed Agent
Command-driven, not hook-driven — instructions live as reviewable files in your repo.
We don't ship a wizard and we don't rely on hooks. We ship an agent that deploys into your repo like an embedded engineer and levels it up with you over weeks. Instruction delivery is command-driven, not hook-driven — more reliable and less intrusive: no per-machine settings.json wiring, and everything lands as durable, reviewable files in your repo.
The model
The agent is the product, not a settings panel. It deploys into your repo, learns it, and gets more capable session over session. Because nothing is wired into a hidden per-machine hook, every instruction it follows is a file you can read, diff, and revert. Two parts make this work: a one-time bootstrap prompt you paste to the agent, and an npm client the agent drives via commands.
The four commands
The agent runs these for you, as commands — not a hidden hook:
cloudclawer-warp install # install the client into the repo
cloudclawer-warp add-roles # write role files (requires an active subscription)
cloudclawer-warp wire-claude-md # insert a managed section into CLAUDE.md
cloudclawer-warp update # pull latest roles + instructions, brief you on changesinstall — installs the client package into the repo.add-roles — writes the role files into the repo (fetched from CloudClawer). Requires an active subscription — this is the billing gate.wire-claude-md — inserts a managed section into CLAUDE.md (between markers) referencing the role files and the behaviors and workflows the agent should follow.update — pulls the latest role files and instruction section, updates the repo idempotently, and briefs you on what changed.Why no hook is needed
Instructions live in CLAUDE.md (read every session) and roles live as files in the repo, so the agent continues day after day with no hook required. The managed CLAUDE.md section itself tells the agent to run update at session start and brief you — that is the auto-update loop, driven by the agent, not a hook.
Auto-update on session start
The session-start flow runs auto-update checks and emits success/error messages; it may ask permission before applying an update. Today it fetches and injects your latest instructions; a version-check plus permission prompt is on the roadmap.
update is idempotent and writes to reviewable files, you can always see exactly what changed in a diff before it lands — the agent briefs you, and the markers keep its edits scoped to the managed section.Cadence & KPIs
The CoA leaves notes about your repo and what it still needs to climb warp. Climbing warp takes weeks, not days — you and the agent agree a cadence (a couple of items a day) and hit KPIs as you go: ideas fed into the ticket system and automation coverage expanded.