Per‑agent API key hygiene lands

Tailscale published the Aperture announcement as a private alpha on January 27, 2026, positioning it as an “AI gateway” that leverages Tailscale identity to centralize AI access for coding agents and developer tools. (tailscale.com) Tailscale’s documentation shows Aperture remained in alpha through March 2026 and is available at no additional cost during the alpha period, with default preview limits that mirror the Personal plan (three free users) and a request-access workflow at aperture.tailscale.com. (tailscale.com) The Aperture proxy explicitly injects provider credentials and forwards calls to upstream LLM endpoints while attaching the originating Tailscale user and device identity, and the product captures session telemetry, token usage, and tool-call metadata for audit and cost analysis. (tailscale.com) Documentation and early reviews list first-class compatibility with major LLM providers and endpoints — OpenAI and Anthropic (with model names surfaced in defaults), plus cloud endpoints including Google/Vertex and AWS Bedrock — and the UI can pre-populate available models for each provider. (tailscale.com) Tailscale published a companion CLI project (aperture-cli) that launches preconfigured coding agents (Claude Code, Gemini CLI, OpenCode, Codex) against an Aperture endpoint, and the core blog post shows simple per-agent configuration examples (e.g., ~/.claude/settings.json) to route agents to the gateway. (github.com) Partner integrations for enforcement are already public: Cerbos documents a workflow where Aperture supplies identity-linked agent telemetry and Cerbos enforces per-request policies (least-privilege tool-call allow/deny), enabling policy-as-code updates that don’t require agent redeploys. (cerbos.dev)