DeepSeek Harness: Why AI Agents Need a Workspace

Open a chat window and ask for a bug fix. The model may return a promising patch, but the real work still sits around it: finding the right repository, opening files, running tests, checking the diff, and remembering what happened five minutes ago. An AI agent can do those things only when something gives the model tools, a workspace, and a controlled way to use them.

That surrounding software is an agent harness. DeepSeek Harness, usually shortened to dsh, is DeepSeek AI's open-source answer: a local workspace for everyday files, coding, research, background jobs, and plugins. What is DeepSeek Harness, and why use it instead of a normal chat page? The useful distinction is that a chat page returns an answer; a harness coordinates a sequence of actions around that answer. The official site describes a public preview built on Cordis's plugin architecture, while the repository still calls the project a developer preview and warns that compatibility-breaking changes are possible. (deepseek.com)

The missing piece between a model and a useful agent

A language model predicts text from the context it receives. An agent is a model placed inside a loop: it can receive a task, choose a tool, inspect the result, and decide what to do next. A tool call is one of those deliberate actions, such as reading a file, running a shell command, or searching a source.

An agent harness manages that loop. It keeps session history, connects the agent to a workspace directory, routes requests to a model provider, handles permissions, and records the trail of actions. Think of the model as a skilled mechanic and the harness as the workshop: benches, tools, parts bins, and a clipboard showing what was done. DeepSeek Harness's preview is built around that workshop idea, with examples spanning documents, spreadsheets, code, research, scheduled tasks, and execution traces.

This matters for beginners because the interface can stay conversational while the work becomes concrete. A request such as find the empty labels in this project, fix them, and run the tests can become a visible sequence of file reads, edits, commands, and results instead of one large block of generated text.

Why everything is a plugin

A plugin is a small add-on that gives an application a new capability. DeepSeek Harness calls its design composable, meaning you can combine those add-ons without rebuilding the whole application. The underlying Cordis framework treats services, tools, interface pieces, and automation as parts that can be loaded, connected, and unloaded.

The developer documentation starts with a TypeScript module. TypeScript is JavaScript with a type system that helps catch certain mistakes before the code runs. A minimal plugin looks like this:

import type { Context } from '@deepseek-ai/cordis'

export const name = 'heartbeat-plugin'

export function apply(ctx: Context) {
 ctx.effect( => {
 const timer = setInterval( => {
 console.log('[heartbeat] still loaded')
 }, 5_000)

 return => clearInterval(timer)
 })
}

The apply function runs when the plugin loads, and ctx is the context object through which the plugin registers behavior. The returned cleanup function stops the timer when the plugin unloads. That detail is more important than it first appears: a plugin should not leave listeners, timers, or connections behind after it disappears. Plugins can also declare dependencies, so the framework waits until services such as tools or language-model access are ready. (deepseek-harness.github.io)

The preview page shows the same idea at a larger scale. A scheduler, terminal, subagent system, voice input, or custom Pomodoro timer can live as a separate capability. Creator mode then becomes a practical entry point: describe the feature, let the agent draft the files, inspect the result, and install it into the running workspace.

Get a local session running

The shortest route uses npx, the Node.js package runner, so there is no global command to maintain:

npx @deepseek-ai/dsh web

The command starts the local Web UI, a browser-based interface, on port 3080 by default. After it opens, configure a DeepSeek API key, which is the secret credential authorizing requests to the model service, and choose a workspace directory. The workspace is the project folder the agent is allowed to inspect and modify. (github.com)

For source work, check out the repository, install its dependencies, build the project, and start the same interface:

pnpm install
pnpm run build
pnpm dsh web

The current development guide lists Node.js 22.19 or newer in the 22 line, or Node.js 24, plus Corepack-enabled pnpm; the repository pins pnpm 11.7.0. That matters when an installation fails before the interesting part begins: a mismatched runtime can look like an application bug. (github.com)

When a browser is too much

DeepSeek Harness also has a headless mode. Headless means the task runs without a graphical interface or a long-lived browser server:

dsh --profile headless 'run the tests'

This form prints the final answer and exits, which fits one-off maintenance tasks and continuous integration (CI), the automated system that builds and tests code after changes. With --json, the runner can emit newline-delimited events for another script to consume, and a session identifier can resume a conversation later. The same project can therefore serve a person at a desk and a small automation job in a terminal. (github.com)

The model connection is not locked to one narrow screen, either. Current settings support DeepSeek, selected third-party providers, and custom model APIs for relays, company gateways, or self-hosted services. That makes the harness less like a single chatbot and more like a host application whose model can be swapped.

The preview label is part of the story

As of October 2, 2026, the newest visible GitHub release is v0.2.0-rc.2, a pre-release published on September 29, 2026. The fast release cadence is exciting, but it also tells you how to approach the software: expect APIs and plugins to move. The public website says public preview; the repository says developer preview. Those descriptions point in the same direction—usable software that is still being shaped.

There is a more important caution. The project’s safety notice says it has not undergone a security audit and can run model-generated commands, load third-party plugins, reach the network, and access files or credentials made available to it. Start with a disposable project folder, use the least access the task needs, keep backups, and review commands and plugins before approving them. A local agent that can edit your files is powerful precisely because it is connected to real tools.

DeepSeek Harness is best understood as a bet on the layer around the model. The model supplies reasoning and language; the harness supplies memory, tools, workspaces, inspection, and a way to package new abilities. That makes dsh interesting not because it replaces the chat window, but because it gives the chat window somewhere useful to go.