DeepSeek Harness Opens Developer Preview: A Fully Plugin‑Based Open‑Source Agent Framework

DeepSeek Harness, now in a globally open developer preview under the MIT license, introduces a thin Cordis core and a fully plugin‑based architecture with over 130 interchangeable capabilities, four preset modes, exhaustive session tracing, and simple npx or source‑code startup, marking a shift from model‑only competition to agent‑infrastructure innovation.

Open Source Tech Hub
Open Source Tech Hub
Open Source Tech Hub
DeepSeek Harness Opens Developer Preview: A Fully Plugin‑Based Open‑Source Agent Framework

Agent Definition

Agent = Model + Harness . Model provides reasoning; Harness acts as the execution nervous system, enabling perception of the environment, tool scheduling, session management, and closing task loops, turning textual output into file edits, command execution, and API calls.

Core Kernel (Cordis)

The Cordis kernel implements only three responsibilities:

Plugin lifecycle management (load, start, stop, unload)

Dependency injection between plugins

Global event bus

It contains no model, tool, or UI logic, allowing independent replacement of any capability.

Plugin Architecture

More than 130 individual plugins cover model integration, sandboxed terminals, file editors, skill system, session storage, scheduling loops, and the Web UI. Examples:

Replace the DeepSeek model with another model plugin.

Swap the default editor for a custom one.

Disable unwanted plugins such as web search.

Add proprietary tool plugins for internal systems.

All modifications are performed via configuration files; no source code changes are required.

Preset Configurations

Four preset configurations are provided:

Standard – full toolset (file editing, shell, web search, skills, planning, sub‑agents, workflow); suited for daily development, project refactoring, general coding tasks.

PTC – includes all Standard capabilities plus TypeScript‑driven multi‑round tool orchestration; suited for complex multi‑step tasks and programmable workflows.

Minimal – retains only persistent Bash and string‑replace editor; suited for model benchmarking, lightweight debugging, simple tasks.

Creative – full capabilities with runtime checks, experimental plugins, and creation guidance; suited for plugin developers and custom agent presets.

Traceable Execution

Execution is recorded with an additive session‑log design. Logged items include:

System prompts injected each round

Complete chain‑of‑thought outputs

Parameters and results of every tool call

Sub‑agent scheduling information

All injected context

The Trajectory view allows filtering by source or round, session recovery, forking, and replay, enabling precise debugging.

Quick Start

Option 1: One‑click npx (recommended)

npx @deepseek-ai/dsh web

Starts a Web UI at http://127.0.0.1:3080; the first run prompts for a DeepSeek API key.

Option 2: Run from source (for extensibility)

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

The project uses a pnpm monorepo, TypeScript as the primary stack, and is released under the MIT license with unrestricted commercial use.

Developer Preview Status

The framework is in a developer‑preview stage; breaking changes are expected before a production‑grade release.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

TypeScriptAI agentsplugin architectureOpen SourceDeepSeekAgent Framework
Open Source Tech Hub
Written by

Open Source Tech Hub

Sharing cutting-edge internet technologies and practical AI resources.

0 followers
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.