Norcaster Companion

Route your AI coding assistant through the Norcaster AI Runtime Control Plane — keep the assistant you love, add the governance your organization needs.

Your engineers already use Claude Code, Codex, or Continue. Every prompt they send carries code, credentials, and sometimes personal data straight to a model provider — ungoverned and unrecorded. Norcaster Companion routes that traffic through your organization's Norcaster gateway: policy is enforced before anything reaches the provider, and every decision leaves an audit record.

This extension is deliberately not another chat. It does setup and visibility; your assistant stays exactly as it is.

What you get

Two-click setup. Sign in with your personal Norcaster API key, run one configure command, and your assistant is governed:

• Norcaster: Configure Claude Code
• Norcaster: Configure Codex CLI
• Norcaster: Configure Continue

Config writes always merge (never overwrite), always show you exactly what will be written first.

And you can turn it off. Norcaster: Remove Claude Code Configuration deletes exactly the three env keys the setup command wrote and leaves the rest of each file alone. It checks every settings file Claude Code reads env from — ~/.claude/settings.json plus .claude/settings.json and .claude/settings.local.json in each folder open in the window — and shows you the full list before touching anything. Removing from some files but not others is the one state worse than not reverting at all: a leftover ANTHROPIC_BASE_URL with no credentials makes every request fail with a 401. It works after Sign Out too — opting out should never depend on still being signed in. Codex and Continue are still manual removals.

Governed status, visibly. A status-bar indicator shows the boundary is active — with distinct warning states when your key is missing, rejected, or the gateway is unreachable. No silent failures, no wondering.

Blocks that explain themselves. When the gateway stops a request — say, an AWS key pasted into a debugging prompt — you get a notification with a direct link to the decision record in Norcaster Decision Audit. Eligible non-credential contextual blocks also offer Forward once when your owner has granted that permission; approve, then retry the exact request before expiry. Credentials and other terminal controls never offer it.

What the gateway does for your traffic

• Blocks access-granting credentials (private keys, cloud keys, tokens) before they reach any model provider
• Redacts structural secrets like connection strings — the model still helps, without seeing the password
• Records every decision with policy version, rule, and timestamp — per developer, per registered AI system
• Blocked requests do not leave your boundary unless a permitted developer explicitly approves an eligible one-time exception; that exception is consumed and audited

Requirements

• A Norcaster workspace with a gateway deployment (norcaster.com)
• A personal API key (one per developer — that's what makes audit attribution meaningful)
• The deployment ID to watch, set in norcaster.deploymentId

Settings

Security notes

Your API key is stored in VS Code SecretStorage (your OS keychain). The Claude Code and Continue configure commands necessarily write the key into those tools' own plaintext config files — that is how those tools consume credentials; the confirmation dialog says so explicitly and files are written with owner-only permissions. Codex reads its key from the environment, so the extension copies an export line to your clipboard instead of editing shell profiles.

Scope, honestly

Input-side enforcement only (output buffering would break streamed code generation). GitHub Copilot cannot be routed (fixed endpoints). Detector patterns are high-signal heuristics — pair with repository-side secret hygiene.

Questions or issues: github.com/norcaster/cloud