The fastest way to try DeepSeek Harness is to run the official npm package, open the local Web UI, configure a model, and select a workspace.
Path A: run the npm package
npx @deepseek-ai/dsh web
Open this address in your browser:
http://127.0.0.1:3080
If the port is already in use:
dsh web --port 3081
Path B: run from source
Source development requires Node.js 22.19+ or 24+ and Git 2.26+. The official repository pins [email protected].
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web
The launcher does not determine whether existing build artifacts are stale. If the UI still looks old after a source change, run pnpm run build again.
Complete first-time setup
- Open Settings → Models and save an API key.
- Click Select workspace, add a project directory, and select it.
- Create a session and send a read-only task:
Summarize this repository and list its main directories. Do not modify any files.
The input remains disabled until a workspace is selected. That is an application state, not a frozen page.
DSH101 independent run record
On 2026-08-14, DSH101 completed a local acceptance run on macOS with Node.js 22.22.2, pnpm 11.7.0, and DSH 0.1.0-rc.5. The page at http://127.0.0.1:3080 returned HTTP 200.
This result proves only that the recorded machine and version worked at that time. It does not guarantee the same outcome on another system, network, model, or later release.
Permission boundary
New sessions use workspace-write by default. It restricts file writes, but does not mean reads, networking, and process visibility are completely isolated. Confirm the workspace scope and approval settings before real work.
The current CLI rejects dsh web --host 0.0.0.0. Do not bypass that restriction to expose an unauthenticated Web UI to the public internet.
Stop the service
Press Ctrl+C in the launch terminal. The CLI performs cleanup first; the final status code for SIGINT is 130.
Evidence and revision
Primary sources
This guide is intentionally concise; use the official source or documentation below as the authority for commands, behavior, and risk boundaries.