Gives Claude Code's "UltraCode" mode (high effort, deep reasoning) to any LLM backend. This project targets users of Anthropic's Claude Code who wish to leverage its advanced workflow and reasoning capabilities with other models they already access via API. It provides a significant benefit by extending sophisticated AI interaction patterns across a broader, user-chosen model ecosystem without requiring new model subscriptions.

How It Works

At its core, UltraCode-Shim acts as a local HTTP proxy. It intercepts requests destined for LLM APIs and injects specific parameters—effort=xhigh, adaptive thinking flags, increased max_tokens, and a tailored system prompt—to emulate Claude's high-effort "UltraCode" mode. This allows any configured backend model to operate under these enhanced reasoning conditions. The proxy also supports an "Orchestrator + Worker" model, enabling users to designate separate models for the main interactive loop and background sub-agent tasks, optimizing for cost or specialized capabilities. An optional "Auto Router" further refines this by dynamically selecting the most cost-effective yet capable model for each specific task based on user-defined criteria.

Quick Start & Requirements

• Primary Install: A single bash (curl) or PowerShell (irm) script automates setup, including running self-tests, creating config.json, and installing a system PATH launcher. Manual installation via git clone is also supported.
• Prerequisites:
  • Claude Code CLI (npm i -g @anthropic-ai/claude-code).
  • Python 3.8+ (standard library only; no pip installs required).
  • Credentials for at least one backend model (e.g., API keys for MiMo, OpenRouter, Ollama, or a Codex login for GPT-5.5).
• Links:
  • Setup Scripts: https://raw.githubusercontent.com/OnlyTerp/UltraCode-Shim/main/install.sh (macOS/Linux/WSL), https://raw.githubusercontent.com/OnlyTerp/UltraCode-Shim/main/install.ps1 (Windows).
  • Documentation: AGENTS.md (AI setup runbook), docs/SETUP.md, docs/HOW_IT_WORKS.md, docs/AUTO_ROUTER.md, docs/ADD_A_MODEL.md.

Highlighted Details

• Dual-Model Workflows: Select distinct models for the primary orchestrator and background workers, or use a single model for end-to-end execution.
• Intelligent Auto Routing: Dynamically assigns tasks to the cheapest capable model based on configurable "capability cards" and a quality threshold, optimizing cost-performance.
• Production-Hardened Reliability: Includes automatic retries for empty turns, mitigation for stalled API streams, and repair mechanisms for rejected tool calls to prevent workflow interruptions.
• Flexible Backend Integration: Supports various model types via openai_compat, codex_oauth, and direct Anthropic endpoints, configurable through config.json.

Maintenance & Community

This is an unofficial, community-driven project, not affiliated with Anthropic, OpenAI, or other model providers. The README highlights AGENTS.md as a detailed runbook for AI assistants to perform setup, configuration, and testing, suggesting a focus on automated deployment. No specific community channels (like Discord or Slack) are listed.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Users are responsible for adhering to the terms of service for any third-party APIs they route through the proxy. The project itself does not impose additional restrictions beyond the MIT license.

Limitations & Caveats

Requires the Claude Code CLI and valid credentials for desired LLM backends. Cursor's Composer integration necessitates the separate cursor-agent CLI and is not HTTP-based. The project is unofficial and relies on the stability and API compatibility of the configured third-party models.