AI Control Plane Public — AI skill for Claude Code

AI Control Plane

**A governance layer for agent work.** Agents enter through one router, disperse into cost tiers and isolated scopes, and ship only through verified, human-gated exits. Built for the case where you need to hand real work to agents and still be able to say what they did, why, and on whose behalf.

Runs on Docker + n8n. Firewalled runners shell out to vendor CLIs (`claude -p`, `gemini -p`); the router decides which one, at what cost tier, under what policy.

**AGPL-3.0-or-later · brand- and vendor-neutral · [QUICKSTART](QUICKSTART.md) · [Generated docs ↗](https://jgobuilds.github.io/ai-control-plane-public/)**

🚧 Under construction — read this before evaluating it

This is **actively being built**, with features tested and rolled out iteratively. Some lanes run daily on real work; others are loaded but have never executed once. Treat it as a system being assembled in the open, not a finished product.

**Nothing here is dressed up as more complete than it is.** The status table below marks each lane's real state, the known gaps are filed as public issues rather than buried in prose, and where something is a stated intent rather than a working control — the retention sweep deletes nothing today — it says so at the point you would otherwise assume otherwise.

If you are deciding whether to depend on this: don't yet. If you are deciding whether the approach is sound, that is what it is here to show.

flowchart LR
  subgraph HOST["Host — both ports bound to 127.0.0.1"]
    n8n["n8n :5678
schedules · gates · logs"]
    view["status.html (generated)"]
  end
  subgraph NET["agentnet — internal, token-gated"]
    router["router
deterministic rules first
then cheapest capable tier
under a maxTier ceiling"]
    claude["claude-runner"]
    gemini["gemini-runner"]
    scrubber["scrubber
PII vault"]
    lanes["lanes
scheduled Python"]
  end
  anth["api.anthropic.com"]
  goog["googleapis.com"]
  l