1 · A wrapper, not a fork

The agent itself is opencode, bundled into the binary and kept current. aixia-agent only supplies the parts that make it “Aixia”: the provider wiring, model, theme, and org rules ship inside the wrapper and are re-applied on every run, so a new release updates everyone at once. Your own additions (such as MCP servers) live in a separate config that is never overwritten.

2 · You sign in with your Aixia account

On first run the CLI starts an OAuth device flow against Aixia's authentik, which is backed by Entra and LDAP. You approve it in a browser on any device; MFA and conditional access apply automatically. The CLI never sees your password, only a short-lived session token. The session is valid for 24 hours, after which you sign in again.

3 · The key belongs to you, and only the cluster holds the master

After sign-in, a small key broker running in the cluster validates your token and mints a personal key for the LLM gateway. The broker is the only thing that holds the gateway's master key, it never leaves Aixia infrastructure and is never baked into the download. Because each person gets their own key, usage is attributable and a single user can be revoked without affecting anyone else.

4 · Prompts and code go only to Aixia's gateway

opencode talks to Aixia's LLM gateway (a LiteLLM proxy on Aixia infrastructure) and nowhere else. There is no path to a public AI provider, and opencode's own session-sharing feature is disabled in the shipped configuration. What you type and the code in your repository are processed within Aixia and are not handed to OpenAI, Anthropic, Google, or any other outside service.

5 · Terminal or desktop, the same engine

You can work straight from the terminal, or install OpenWork, an optional desktop app (an open-source Electron front-end). Either way it launches the same opencode on your machine, with the same per-user key and managed config, OpenWork is just a window onto the local agent, not a separate service, so the sign-in and the data path above are identical. The native Windows (.msi) and macOS (.pkg) installers bundle it; the aixia-agent wrapper launches it with the right environment so it inherits your Aixia configuration.

Your data stays inside Aixia

• Prompts, code, and responses are routed only to Aixia's own gateway on Aixia-operated infrastructure, never to a third-party AI provider.
• opencode's cloud session-sharing is turned off in the managed config.
• Credentials are stored locally with owner-only permissions (0600), and the session expires after 24 hours.

Identity is handled by authentik, not the binary

• Login goes through Aixia's authentik (Entra / LDAP), so MFA and conditional-access policies apply, the CLI never handles your password.
• The LLM gateway's master key lives only in the in-cluster broker; the download contains no shared secret to extract.
• Keys are per-user: revocable, budgeted, and attributable to a person rather than to a shared credential.

Downloads are verified end to end

• Every release ships a signed checksum file (minisign) plus a per-file SHA-256 sum; the install scripts verify both before running anything.
• The signing public key is embedded in the install script itself, fetched over TLS from this site, the trust anchor.
• The site and all services are deployed by GitOps on Aixia's Kubernetes cluster, the same way the rest of the platform is run.

One honest caveat: the native installers (Windows .msi, macOS .pkg) aren't code-signed with an OS vendor certificate yet, so Windows SmartScreen and macOS Gatekeeper warn on first open. The minisign signature and checksums still prove the download is genuine; OS-level signing is a planned addition.