Preface¶

When working on agent tasks in DeepSeek Harness (DSH), a common gap is that DeepSeek can reason around text, code, and APIs, but image pixels don’t automatically become model inputs. If developers want to put screenshot understanding, Browser operations, and MCP tool calls into the same task, while ensuring every step’s evidence is traceable, they need an auditable runtime. DeepSeekEyes, introduced below, is a DSH plugin designed for this scenario.

What It Is¶

DeepSeekEyes is maintained by dttxorg and licensed under MIT. It is an auditable vision and cross-platform Computer Use runtime for DeepSeek Harness. In simple terms, it connects image evidence, Browser automation, native Windows/macOS desktop control, and the MCP application layer into DSH tasks while preserving source evidence.

Core Features¶

Auditable Vision Evidence¶

After pasting an image in the Harness, the configured multimodal model reads the raw pixels, and DeepSeek receives the validated evidence and responds. User images are not scaled, converted, or recompressed; each re-read references the original content-addressed attachment.

In DSH rc.8+, if the upstream model explicitly declares image input, DeepSeekEyes directly uses the raw ImageBlock and records a native-bypass turn.

Cross-platform Computer Use¶

DeepSeekEyes provides Browser automation and native Windows/macOS desktop control. Browser automation can open pages, scroll, click, and verify results.

MCP Application Layer¶

Once the MCP server is enabled and allowed tools are selected, DeepSeek invokes applications, and DeepSeekEyes constrains, hashes, and audits the results. By default, optional automation and MCP are turned off.

Health Checks and Failover¶

DeepSeekEyes provides health check and failover capabilities for vision routing management.

Strict Evidence Validation¶

Evidence fields are strictly validated using JSON Schema/Ajv.

Token Usage Visibility¶

The plugin provides token usage visibility.

Installation and Enablement¶

First, confirm the Node.js version meets the requirements:

Node.js >= 22.19

Run the installation command in a macOS/Linux shell or Windows PowerShell:

npx -y @dttxorg/deepseekeyes@latest install

If an upgrade or diagnostics are needed, run respectively:

npx -y @dttxorg/deepseekeyes@latest upgrade
npx -y @dttxorg/deepseekeyes@latest doctor

If the current DSH profile is not web, append to the command:

--profile NAME

After installation or upgrade, restart dsh web once:

dsh web

Then complete configuration in the Harness under Settings → Models.

Typical Usage¶

Image Understanding¶

Paste an image in the session, the configured multimodal model reads the raw pixels, and DeepSeek receives the validated evidence and responds within the current task.

Browser Computer Use¶

Have DeepSeek perform page operations, such as opening pages, scrolling, clicking, and verifying results.

MCP Application Calls¶

Enable the MCP server, select allowed tools, DeepSeek invokes applications, and DeepSeekEyes constrains, hashes, and audits the returned results.

Recommended Use Cases and Notes¶

It is suitable for developers who need to put image evidence, Browser operations, and MCP tool calls into the same task within DSH. If you only need text reasoning, you may not necessarily need to enable its automation capabilities.

Note that by default, optional automation and MCP are turned off. The plugin runs with the current dsh process permissions, so you should review the source code and the MIT license before installation.

Resources¶

Community directory page:

https://www.skillhub.cn/plugins/dttxorg/deepseekeyes

GitHub repository:

https://github.com/dttxorg/deepseekeyes