DeepSee: Vision and Smart Model Routing for DeepSeek Harness

Overview

DeepSee is a plugin designed to transform the DeepSeek Harness into a multimodal, multi-model coding workspace. Its core philosophy is "Gemini sees, DeepSeek codes." The tool enables users to paste screenshots into text-only DeepSeek sessions, route tasks based on cost and difficulty, and visually verify results before delivery. It integrates seamlessly with existing AI coding environments like Claude Code, Codex, OpenCode, and Pi.

Benefits

DeepSee offers several key advantages for developers and coding enthusiasts. First, it uses a dual-model orchestration system that pairs DeepSeek V4 Flash for quick tasks like documentation and small edits with DeepSeek V4 Pro for complex work like architecture design and security reviews. This smart routing helps save money by using the cheaper model for simple jobs and the powerful model only when needed. Second, the tool includes a visual verification loop that automatically checks UI-related code. It uses a vision engine to spot defects and ensure the code looks correct before delivering the final result. Third, DeepSee supports a wide range of vision sources, including free options like the built-in Gemini API and various local CLI tools. This flexibility ensures users can find a vision service that fits their needs and budget. Finally, the plugin is native and removable, meaning it can be installed easily and removed at any time without leaving behind any local proxy daemons.

Use Cases

DeepSee is ideal for developers who work with AI coding assistants and need to handle visual tasks. It is perfect for users who want to paste screenshots into their coding sessions to get help with UI issues. The tool is also useful for teams that need to balance cost and performance by automatically routing simple tasks to faster models and complex tasks to more powerful ones. Developers working with various AI coding environments like Claude Code or Codex can use DeepSee to enhance their workflow with visual verification. It is also a great choice for users who want to leverage existing logins from other tools like Grok or Pi without setting up new accounts. The ability to configure custom providers makes it suitable for advanced users who want to integrate specific vision models into their setup.

Pricing

DeepSee offers a free-first routing policy called DeepSee Auto. This means users can start with the free DeepSeek V4 Flash model for initial tasks and only pay for the more expensive DeepSeek V4 Pro model when complex tasks require it. The tool also supports free vision engines like the built-in Gemini API and the antigravity-cli, which requires no API key. For users who need more control, the DeepSee Customize option allows them to manually assign models to specific work categories. While specific pricing for the Pro model is not detailed in the provided information, the tool is designed to help users minimize costs by using the right model for the right job.

Vibes

DeepSee has received positive attention for its innovative approach to combining vision and coding. Users appreciate the seamless integration with existing AI coding environments and the ability to visually verify code results. The tool's flexibility in supporting multiple vision sources has been highlighted as a significant advantage, allowing users to choose the best option for their needs. The community has welcomed the open-source nature of the project, with contributions and feedback driving its development. The MIT license and the author's commitment to transparency have fostered a sense of trust among users. Overall, DeepSee is seen as a valuable addition to the AI coding ecosystem, offering a practical solution for handling visual tasks in a text-based environment.

Additional Information

DeepSee is designed, developed, and maintained by chang416. The project is open-source and welcomes contributions via focused pull requests. Before submitting code, contributors are encouraged to run linting, type checking, testing, and building commands to ensure quality. The tool is licensed under the MIT license, which allows for free use, modification, and distribution with no warranty. DeepSee integrates with various upstream engines like Antigravity CLI, Gemini, OpenAI, and Anthropic APIs, respecting their individual terms of service and quotas. The project acknowledges early exploration referenced MIT-licensed code and expresses gratitude to the community for their support.