docs/zai/provider.md

z.ai provider (Anthropic-compatible passthrough)

Idea

Support z.ai (GLM) as an optional upstream for Anthropic-compatible requests (/v1/messages), without applying any Google/Gemini-specific transformations when z.ai is selected.

This keeps compatibility high (request/response shapes stay Anthropic-like) and avoids coupling z.ai traffic to the Google account pool.

Result

We added an optional “z.ai provider” that:

• Is configured in proxy settings (proxy.zai.*).
• Can be enabled/disabled and used via dispatch modes.
• Forwards /v1/messages and /v1/messages/count_tokens to a z.ai Anthropic-compatible base URL.
• Streams responses back without parsing SSE.

Configuration

Schema: src-tauri/src/proxy/config.rs

• ZaiConfig in src-tauri/src/proxy/config.rs
• ZaiDispatchMode in src-tauri/src/proxy/config.rs

Key fields:

• proxy.zai.enabled
• proxy.zai.base_url (default https://api.z.ai/api/anthropic)
• proxy.zai.api_key
• proxy.zai.dispatch_mode:
  • off
  • exclusive
  • pooled
  • fallback
• proxy.zai.models default mapping for claude-* request models:
  • opus, sonnet, haiku

Routing logic

Entry point: src-tauri/src/proxy/handlers/claude.rs

• handle_messages(...) decides whether to route the request to z.ai or to the existing Google-backed flow.
• pooled mode uses round-robin across (google_accounts + 1) slots, where slot 0 is z.ai.

Upstream implementation

Provider implementation: src-tauri/src/proxy/providers/zai_anthropic.rs

• Forwarding is conservative about headers (does not forward the proxy’s own auth key).
• Injects z.ai auth (Authorization / x-api-key) and forwards the request body as-is.
• Uses the global upstream proxy config when configured.

Validation

1. Enable z.ai in the UI (src/pages/ApiProxy.tsx) and set dispatch_mode=exclusive.
   • UI: src/pages/ApiProxy.tsx
2. Start the proxy.
3. Send a normal Anthropic request to POST /v1/messages.
4. Verify the request is served by z.ai (and Google accounts are not involved for this endpoint in exclusive mode).