Mixdog: Open-Source AI Coding Harness Reduces Costs and Context Usage vs Native CLIs

Overview

Mixdog is an open-source AI coding harness designed to optimize token usage and cost efficiency for LLM-based coding agents. It provides an integrated workspace for managing sessions, agents, and workflows across Desktop, CLI, and browser interfaces. The project is licensed under Apache-2.0 and supports major providers including Anthropic, OpenAI, Google Gemini, and xAI.

Benchmark Performance

The project reports results from Terminal-Bench 2.1 using the official Harbor verifier, comparing Mixdog against native CLI tools for specific models. The tests used identical tasks and models, changing only the harness.

• GPT-5.6 Sol xhigh (vs. Codex CLI):

  • Cost: Mixdog averaged $0.476/trial compared to Codex CLI's $0.782/trial (39% lower).
  • Context: Median final context was 18.5k tokens for Mixdog vs. 34.3k tokens for Codex CLI (46% smaller).
  • Accuracy: Mixdog achieved 86.5% (385/445) vs. Codex CLI's 86.1% (383/445).
  • Speed: Comparable performance, with Mixdog at 415s/trial vs. Codex CLI at 437s/trial.

• Claude Opus 5 (vs. Claude Code):

  • Cost: Mixdog averaged $104.29/run compared to Claude Code's $129.21/run (19% lower).
  • Context: Median final context was 27.6k tokens for Mixdog vs. 38.2k tokens for Claude Code (28% smaller).
  • Accuracy: Mixdog scored 79/89 vs. Claude Code's 77/89.
  • Speed: Mixdog was 1.16x faster (610s/trial vs. 708s/trial).

Key Efficiency Mechanisms

Mixdog reduces overhead through several technical strategies:

• Context Management: Uses lightweight system instructions, structured compaction, and idle-time context reduction to minimize token usage.
• Code Intelligence: Integrates tree-sitter and ast-grep for native code parsing, supporting 31 languages for parsing and 24 languages for symbol/import extraction. This allows structural queries without dumping entire files.
• Caching: Implements provider-aware caching to reuse processed context, though cached tokens still count toward context limits.
• Tooling: Features purpose-built tools with bounded results and on-demand prompt loading to avoid unnecessary data transfer.

Features and Integration

• Workspace: Desktop app includes Monaco editor, Git integration, terminal, and file explorer. Supports multi-device session continuity between Desktop, TUI, and paired browsers.
• Beyond Code: Supports Browser Use (Chromium), Computer Use (Windows), and document/image/video generation.
• Local Execution: Includes a Mixdog Local Provider for running models on Windows x64 with NVIDIA GPUs, eliminating the need for separate server setups.
• Extensibility: Supports MCP servers, skills, hooks, and plugins. Workflows and agents are defined via Markdown files (WORKFLOW.md, AGENT.md).