ZCode Switcher

Desktop manager (Linux / Windows / macOS) for your z.ai coding-plan tokens, built with Tauri 2 + Vue 3 + TypeScript.

Store one or more tokens, check your plan's quota usage at a glance, and hot-swap the active token straight into zcode's config.

Main view	After "Cambiar Token"

Features

• Multi-token: add as many z.ai tokens as you need; the list persists in the app's own config (never inside the agents' configs).
• Multi-agent: choose where the switch applies — ZCode, OpenCode, or both. Your choice persists.
• Cambiar Token: one click rotates round-robin to the next configured token and applies it to the enabled agents.
• Quota card: live usage from GET https://api.z.ai/api/monitor/usage/quota/limit (Authorization: Bearer $TOKEN), fetched Rust-side (reqwest + rustls) to avoid webview CORS. Unknown response shapes fall back to a raw JSON view.
• Safe config writes: reads each config → modifies only the token fields → writes atomically (temp file + rename), keeping a one-shot .bak backup. Everything else is preserved.
• Native View Transitions on supported webviews (Windows / macOS), instant updates elsewhere.

How it works

The active token is written into the credential store of each enabled agent, resolved from the user home directory on every platform:

Agent	Credential store	What changes
ZCode	~/.zcode/v2/config.json (C:\Users\$USER\…, /Users/$USER/…)	provider.builtin:zai.options.apiKey + provider.builtin:zai-coding-plan.options.apiKey
OpenCode	~/.local/share/opencode/auth.json (same on all OSes — it's what opencode auth login writes)	zai-coding-plan.key + zai.key

The key is written to both provider entries on each agent (and stale systemDisabledReason flags are cleared on zcode), so the switch takes effect no matter which provider the agent resolves. Agents rewrite these files on their own, so the app never caches them — every switch is a fresh read → modify → write cycle.

Token list lives in the app config dir (tokens.json, mode 0600 on Unix). Tokens are never logged and never leave your machine except to the official z.ai quota endpoint.

Development

Bun is required — src-tauri/tauri.conf.json runs the frontend through bun run ...:

bun install
bun run tauri dev      # full desktop app
bun run tauri build    # release bundle, all targets

bunx vue-tsc --noEmit  # typecheck (the only automated gate)
cargo check            # inside src-tauri/, for the Rust side

bun run dev serves the frontend alone in a browser — invoke() calls will fail there. For UI iteration without the Tauri shell there's a demo mode with a mocked IPC bridge and fabricated data:

bun run dev
# then open http://localhost:1420/demo.html

The README screenshots above were captured from that demo (?autoSwitch=1 clicks the real "Cambiar Token" button after mount).

Cutting a release

bun run release 0.1.2          # bump, validate, commit, tag, push
bun run release 0.1.2 --dry-run

bun run release <version> bumps every manifest (tauri.conf.json, Cargo.toml, package.json, Cargo.lock), validates that the tag matches the new version, commits, tags and pushes — pushing the tag is what triggers CI to build and publish the installers for all platforms.

Platform notes

• Linux + NVIDIA/VM: WebKitGTK's DMABUF renderer can produce a black window (Failed to create GBM buffer). The app sets WEBKIT_DISABLE_DMABUF_RENDERER=1 at startup on Linux — a pre-set env var always wins if you need to override it.
• Linux + View Transitions: WebKitGTK exposes startViewTransition but can hang the webview on the forced legacy renderer; transitions are therefore disabled on Linux by design (progressive enhancement, never a hard dependency).
• Building for all OSes: Tauri can't cross-compile — Windows and macOS installers are built by GitHub Actions (.github/workflows/build.yml, runs on v* tags) and attached to a draft release. When bundling the AppImage locally, use APPIMAGE_EXTRACT_AND_RUN=1 NO_STRIP=true bun run tauri build.
• Vite dev server uses fixed port 1420 (strictPort: true) with HMR on 1421.