Model
Interprets the task, reasons about the next step, and selects an available tool.
Independent English overview
Understand and run DeepSeek Harness
A practical map of the open-source agent harness: what it does, which runtime surface to choose, how to start safely, and which claims we have actually verified.
$ npx @deepseek-ai/dsh web
Open http://127.0.0.1:3080
Browser UI · local workspace · persistent sessions
Tracked release0.1.0-rc.5
Source revisionmaster@47f943859b
Last verified2026-08-14
MaturityDeveloper Preview
The mental model
The model reasons. The harness makes work possible.
DeepSeek Harness—whose command is dsh—is an open-source agent harness from DeepSeek AI. It is not a model. It connects a model to tools, permissions, session state, persistence, and user interfaces so an agent can carry a task through multiple steps.
Agent loop
Executes tool calls, records their results, and continues until the task ends or needs intervention.
Tools and policy
Define what the agent can inspect or change and which operations require approval.
Session
Preserves events and state so work can be resumed, replayed, or forked.
Choose a surface
Four ways into the same runtime
Start with the surface that matches your job, then follow the English deep dive for commands, constraints, and evidence.
Web UI
A local browser interface for workspaces, sessions, settings, and approval prompts.
npx @deepseek-ai/dsh webRead the guide Headless runner
Run a one-off task, persist its session, and print the final assistant response to the terminal.
dsh --profile headless "<task>"Read the guide Python SDK
Start and drive the Harness runtime from Python over JSON-RPC.
pip install deepseek-harness-sdkRead the guide Cordis plugins
Extend tools, model providers, services, and runtime behavior through the Cordis plugin model.
export function apply(ctx) { /* ... */ }Read the guide First local run
Launch the Web UI in a few minutes
The shortest path is the published npm package. Keep the UI on localhost, configure a model, select a small workspace, and begin with a read-only task.
npx @deepseek-ai/dsh webhttp://127.0.0.1:3080- 01Configure a model
Open Settings → Models and save the provider credentials.
- 02Select a workspace
Choose the smallest project directory that contains the files you need.
- 03Start with a bounded task
Ask it to summarize the repository without modifying files.
Before real work
Treat permissions and plugins as security decisions
Developer Preview means commands and configuration may change. More importantly, an agent harness can execute shell commands and modify files, so its operating boundary matters.
Minimize the workspace
workspace-write limits writes, but it is not a complete isolation boundary for reads, networking, or processes.
Keep credentials out of Git
Use the model settings or local credential files. Never commit literal API keys or paste them into issue logs.
Keep the Web UI local
Do not work around the localhost restriction to expose an unauthenticated agent interface to the public internet.
Review third-party installs
A plugin install script may run outside the agent session sandbox. Review it, pin a revision, and test it in a disposable environment.
Evidence, not mirroring
Every claim has a scope
DSH101 separates facts found in official documentation or source code from behavior reproduced locally. Version and verification dates stay visible because a Developer Preview can change quickly.
Official factDocumentation or source reviewedLocal testReproduced in a recorded environment
English coverage
The complete public knowledge base is available in English under /en/. Each English page links to its Chinese counterpart and preserves the same evidence and version scope.