Introduction

Halos (halo) is a control plane for AI developer tools in organizations. It treats configuration of Claude Code, Codex, Gemini CLI, Copilot CLI, Cursor CLI, OpenCode and similar harnesses as a signed, ring-deployed release, gates promotion on evidence, and routes model traffic through a plane you control so the most common rollback needs no release: a signed kill reaches gateways at their next poll (about 10 s).

The problem

AI coding CLIs have become production infrastructure, but they are still managed like personal tools. A vendor release or a model change reaches the whole engineering organization at once, unmeasured. When behaviour regresses behind a corporate gateway, platform teams lack three answers: who is running which version, whether the change made outcomes worse, and how to reverse it without asking every engineer to act.

• Settings sprawl. Each harness has its own config format and enforcement knobs.
• No central channel on gateways. Claude Code’s server-managed settings are not fetched with Bedrock or a custom ANTHROPIC_BASE_URL, so config must be delivered as files, MDM profiles or environment definitions.
• All-at-once upgrades. No way to A/B a CLI or model upgrade.
• No eval gate, no rollback.

What you get

Need	Halos feature
One source of truth	Policy repo validated by strict typed decoding and Go guardrails (JSON Schemas are for editors and CI)
Safe rollout	Rings and releases with signed pointers
Prove a model or CLI change	Experiments, shadow, evals
Enforce on every machine	Delivery: dev containers first, halod and MDM for laptops
Developers help themselves	Self-service portal
Any gateway, any IdP	Stack-agnostic traffic plane
Trust	Signed bundles and pointers, verified identity, stripped headers
Let an agent drive it, safely	Agentic operations

Mental model

Where to go next

1. Quickstart.
2. Architecture.
3. The guide matching your setup: Bedrock via Kong or direct Anthropic.
4. Going live: production deployment and the threat model.