How to use ego (lite) with DeepSeek Harness
Set up DeepSeek Harness browser automation with ego (lite). Enable Full access in the Web UI or headless Terminal, load ego-browser, and complete a browser task.
DeepSeek Harness can control ego (lite) from its local Web UI or a one-shot headless Terminal command. In this guide, DeepSeek Harness opens Anthropic's X profile in an ego (lite) Agent Space, reads the first five posts, and returns the result.
The complete video above follows the Web UI workflow. DeepSeek Harness does not currently ship a separate native desktop app or a built-in TUI, so this guide uses the official Web UI and headless profile instead.
Before you start
Install ego (lite), finish onboarding, and make sure Node.js and a working model provider are available.
- Download the DMG and open it
- Follow the instructions and double-click the ego (lite) icon to install
- Open ego (lite) and finish onboarding
-
- Install Node.js
-
- Launch DeepSeek Harness and connect a model in Settings > Models
ego (lite) installs ego-browser under ~/.agents/skills/. DeepSeek Harness scans this shared user Skill directory, as well as .agents/skills/ inside the current project.
After installing ego (lite), fully stop and relaunch DeepSeek Harness before continuing. This lets the Harness reload the newly installed Skill and show the /ego-browser slash command in the Web UI. Starting only a new session may not reload the Skill.
Choose the Web UI or headless Terminal
To use the Web UI, open Terminal in your project directory and run:
npx @deepseek-ai/dsh web
DeepSeek Harness opens its local Web UI at http://127.0.0.1:3080. If dsh is already installed globally, you can use dsh web instead.
For Terminal, use the headless profile. It runs one task, prints the final answer, and exits. Here, “headless” describes the Harness interface; ego (lite) still opens a visible browser window while the task runs.
Step 1: Enable Full access
Web UI
Open the permission menu below the prompt box and select Full access. Read the warning, check the acknowledgement box, and select Enable Full access.
Full access maps to the danger-full-access preset: commands are not confined by the Harness sandbox, and approval prompts are disabled. Only enable it for a task and workspace you trust.
Headless Terminal
Set the permission mode for the same process that runs the headless task:
DSH_PERMISSION_MODE=danger-full-access dsh --profile headless \
"Use the ego-browser skill to open Anthropic's official X profile and collect the first five posts shown in the Posts feed."
The environment variable enables danger-full-access with the approval policy set to never for this run. The quoted text at the end is the task prompt, not another configuration option.
Step 2: Load the ego-browser Skill
Web UI
In the prompt box, type /ego-browser and select the matching Skill from the suggestion list. Continue writing or paste the browser task after the Skill name, then submit the message.
A direct /ego-browser token loads the Skill before the task begins. DeepSeek Harness can also select a model-invocable Skill from the task intent, but the explicit token makes the recorded workflow easier to verify.
Headless Terminal
The headless command in Step 1 already says to use the ego-browser Skill, so a separate slash command is not required. You can also include /ego-browser at the beginning of the quoted task if you want to invoke it explicitly.
Step 3: Let DeepSeek Harness complete the browser task
Web UI
DeepSeek Harness loads the Skill, calls the ego-browser command-line helper, and opens Anthropic's X profile in a separate ego (lite) Agent Space. Keep the Web UI open while the Agent reads the requested posts and prepares the response.
Headless Terminal
The same browser workflow runs from Terminal without a Harness conversation window. The Terminal remains the controller on one side while ego (lite) shows the live browser activity on the other.
Step 4: Complete the task
When the run finishes, the Web UI shows the five posts with their visible details, direct links, and a short summary of each post. In headless mode, DeepSeek Harness prints the final answer in Terminal and exits. The Anthropic profile remains open in ego (lite).
Troubleshooting
/ego-browser does not appear
Confirm that ~/.agents/skills/ego-browser/SKILL.md exists. Then start DeepSeek Harness from the intended project directory and create a new session. DeepSeek Harness also scans <project>/.agents/skills/ for project-specific Skills.
The task stops for permission approval
In the Web UI, confirm that the current session shows Full access before submitting the task. In Terminal, make sure DSH_PERMISSION_MODE=danger-full-access appears before dsh in the same command.
dsh is not found in Terminal
Use the package runner form instead:
DSH_PERMISSION_MODE=danger-full-access npx @deepseek-ai/dsh --profile headless \
"Use the ego-browser skill to open Anthropic's official X profile and collect the first five posts shown in the Posts feed."
Why is there no desktop app or TUI in this guide?
DeepSeek Harness currently provides an official local Web UI and a headless CLI profile. The Web UI runs locally on your computer but opens in a browser; it is not a native desktop app. A TUI requires a separate community plugin and is outside this guide.
X asks you to sign in or verify the account
Complete the login or verification in the ego (lite) Agent Space, then run the task again.
Go deeper
- Quick start: install ego (lite) and run your first browser task
- Claude Code browser automation: use ego-browser from Claude Code
- Codex browser automation: use ego-browser from Codex Desktop or the TUI
- OpenCode browser automation: use ego-browser from OpenCode Desktop or the TUI
- OpenClaw browser automation: use ego-browser from the OpenClaw macOS app or TUI
- ego-browser: learn the browser helper commands
- DeepSeek Harness official site: see the product overview and quick start
- DeepSeek Harness repository: read the source and current release notes
- DeepSeek Harness user guide: configure the Web UI and model providers
- DeepSeek Harness permission presets: understand Full access, sandbox mode, and approval policy