codexsplit

AITabby/codexsplit is ranked 6,379 among these 8 peers; sylearn/AIUsage is ranked 6,377.

It manages third-party providers, model metadata, the official GPT account pool, GPT-Live handoffs, voice settings, Agent routing, and local session import while keeping native Codex behavior available. Modern Dual-Routing Architecture CodexSplit utilizes a **Native Gateway + SQLite Trigger Dual-Routing Architecture**: third-party models communicate directly with the local gateway via native HTTP, cleanly separated from native official GPT. | Scenario | Actual Path | Bridge Involved? **This is the primary remaining purpose of Desktop Bridge**: dynamic credential injection during 5

-hour rate limits and quota failover | | **Third-Party Model Main Session** | Native Direct HTTP → CodexSplit Gateway ( ) | ❌ No | Routed to provider automatically via SQLite triggers. Full conversation context and live Token usage are 100% faithfully preserved | | ** / Task Delegation** | Routed by Agent Capability Profile | ❌ No | Main conversation keeps its path; delegated child tasks route directly to configured targets | Gateway vs Desktop Bridge Roles **CodexSplit Gateway (Core)**: Listens on . Manages third-party models (Claude, Gemini, DeepSeek, Qwen, MiniMax, etc. ), API Key credential pools

, local OAuth subscriptions (Antigravity, Grok, Cursor, etc. ), protocol transformations (Responses ↔ Chat), and token usage normalization. **If you do not use official multi-account rotation, daily third-party model usage requires only the gateway running. ** **Desktop Bridge (Only for Account Pool)**: An optional stdio interceptor. **Its sole purpose is rotating multiple official ChatGPT accounts on native GPT** (quota-weighted round-robin and 429 failover), plus occasional desktop UI refresh triggers. **CodexSplit App**: Local control dashboard for managing models, credentials, routing policies

, and sessions. **Context Preservation Guarantee**: Both official and third-party models inherit 100% complete session history without synthetic truncation or fake system notices. Upper limits are natively managed by Codex's auto-compaction. Configure an API Key Provider or import a local OAuth subscription. Save a model and test it under **Pending Models**. After the test succeeds, click **Restart Codex (apply model menu)** so the model appears in Codex Desktop. If you only use official GPT, you do not need to

add a third-party model or manually restart Codex just to open CodexSplit. Third-party model configuration, Desktop Bridge switching, and Codex Desktop restart are explicit actions. Windows Windows users download and run the Windows EXE from the v1. Windows source is not distributed; do not clone or build the repository for the Windows installation workflow. The Windows EXE provides the features included in that release package. macOS DMG notes The macOS DMG includes its own Node. The DMG is

not notarized; if macOS blocks the first launch, allow it under **System Settings → Privacy & Security**. Add an API Key Provider Open **Gateway → API Key Access**: 1. Choose a preset such as DeepSeek, MiniMax, Qwen, Z. A custom OpenAI-compatible Provider also needs an Endpoint / Base URL. The returned model list from the current Endpoint is the source of truth. Choose a model and select or according to the upstream API. The model first appears under **Pending Models** and does not

automatically restart Codex. On failure, check the Endpoint, Key, backend model ID, and protocol. When you want Desktop to use