A new open-source proxy tool enables developers to access Claude Code's advanced UltraCode mode on any AI model they already subscribe to. UltraCode-Shim, created by developer OnlyTerp, is a local HTTP proxy that intercepts Claude Code API requests and adds the UltraCode envelope before routing to configured backends, breaking the lock-in to specific models.
Technical Architecture Enables Multi-Model Flexibility
The proxy implements three core capabilities to extend UltraCode functionality across different AI models:
-
Orchestrator/Worker Routing: Splits Claude Code's single model slot into two tiers—one for interactive planning and another for parallel sub-agents. The system classifies each request by structural signals and routes orchestrator tasks to premium models while distributing worker tasks across cheaper alternatives.
-
Auto Router: A classifier model scores each backend's suitability for specific tasks, directing trivial requests to cost-effective models and complex operations to premium backends.
-
Reliability Hardening: Handles empty turns with retries, prevents stream stalls with timeouts, repairs rejected tool calls, and manages reasoning model processing time without freezing the interface.
How UltraCode Works at the API Level
UltraCode mode operates through a specific API envelope: effort=xhigh combined with adaptive thinking, increased token limits, and system reminders. UltraCode-Shim adds this envelope to every request, enabling any backend to receive the UltraCode treatment. The proxy translates Anthropic request formats to OpenAI Chat Completions, supporting models from MiMo, DeepSeek, StepFun, Ollama, OpenRouter, and local llama.cpp or LM Studio instances.
Supported Models and Setup Requirements
Pre-configured backends include GPT-5.5 (via Codex login), MiniMax-M3, MiMo v2.5 Pro, DeepSeek V4 Pro/Flash, Step Flash, Ollama Cloud, OpenCode Go, and OpenRouter. Real Claude models (Opus/Sonnet/Haiku) appear automatically, and any OpenAI-compatible endpoint can be added through configuration.
Setup requires Claude Code CLI with UltraCode access, Python 3.8+ with no external dependencies, at least one backend credential, and runs on Windows 11, macOS, Linux, or WSL. Installation uses a single command with offline self-tests.
Community Adoption and Developer Background
The project has garnered 205 GitHub stars and 16 forks since release, indicating moderate but growing community interest. The repository includes AGENTS.md, a step-by-step runbook designed for AI systems to follow for self-setup. Developer OnlyTerp emphasized flexibility in the announcement: "You deserve the flexibility to use LOCAL models & cost efficient models. So I made that happen for you."
Key Takeaways
- UltraCode-Shim is a local proxy that enables Claude Code's advanced UltraCode mode on any AI model through API envelope injection
- The tool implements orchestrator/worker routing to split tasks between premium and cost-effective models, plus auto-routing for intelligent backend selection
- Supported models include GPT-5.5, MiniMax-M3, DeepSeek V4, and any OpenAI-compatible endpoint with stdlib-only Python requirements
- The project has achieved 205 GitHub stars and includes an AI-readable setup guide for automated deployment
- Installation requires Claude Code CLI with UltraCode access and at least one backend API credential