Summary

CodexSaver addresses the high cost of using powerful LLMs like Codex for routine coding tasks. It acts as an MCP tool, routing low-risk development work (e.g., tests, docs, search) to cheaper worker LLMs, while retaining high-risk judgment and final review for Codex. This pattern enables significant cost savings for developers without compromising critical decision-making quality.

How It Works

The system employs a Router to classify tasks by risk. Low-risk tasks are delegated via a Provider Client to a configured worker LLM (defaulting to DeepSeek), with context managed by a Context Packer. A Verifier validates outputs, and a Cost Estimator tracks savings. High-risk, ambiguous, or sensitive tasks (architecture, security, final review) remain with Codex. This approach ensures expensive models handle judgment, while cheaper models manage volume.

Quick Start & Requirements

Installation involves cloning the repository, setting up an API key for a chosen provider (e.g., DeepSeek), and running python cli.py install. A global MCP entry is created in ~/.codex/config.toml. Project-local installation is also supported via python cli.py install --project. Requires Python and Git. API keys are necessary for most hosted providers. Setup is rapid, with a "60-Second Demo" and doctor command for verification.

Highlighted Details

• Cost Savings: Benchmarks show an average estimated saving of 48.4%, reducing the cost index from 1.00 to 0.52 across typical low-risk development tasks.
• Transparent Routing: LLM responses include an interaction block detailing the routing decision (e.g., delegated_execution, route_label), making the tool's activity visible.
• Flexible Provider Support: Integrates with DeepSeek, OpenAI, Anthropic, Gemini, Qwen, Ollama, LM Studio, and custom OpenAI-compatible endpoints.
• End-to-End Verification: Features a robust verification flow and a doctor command to confirm setup readiness and provider configuration.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, or community channels (e.g., Discord, Slack).

Licensing & Compatibility

The repository's license is not specified in the README. This absence requires clarification for commercial use or integration into closed-source projects.

Limitations & Caveats

Current routing is primarily rule-based; advanced "cost-aware dynamic routing" and "cost-aware provider selection" are listed as future roadmap items. The system includes explicit keyword guards (e.g., "production logic") to retain tasks with Codex, indicating a deliberate, rather than fully adaptive, risk assessment. The lack of a stated license is a significant adoption blocker.