DeepSeek Harness (commonly referred to as dsh) is an open-source, local-first agent runtime and development environment created by DeepSeek AI. Designed to sit between a language model and a developer's workstation or repository, it provides the essential scaffolding—such as tools, sessions, sandboxing, and execution loops—required to run complex AI coding agents.
Unlike standalone language models or fixed, opinionated coding assistants, DeepSeek Harness is built on a modular plugin philosophy where "Everything is a Plugin".
Core Architecture and Design Philosophy
At the foundation of DeepSeek Harness is Cordis, a framework utilizing a programming paradigm for spatial and temporal composability. This architecture allows developers to treat nearly every component of the agent runtime as a swappable, extensible plugin.
Plugin Ecosystem: Model adapters, individual tools, skill sets, session stores, sandboxes, storage backends, scheduling systems, agent loops, and even user interfaces are structured as plugins.
Spatial Composability: Plugins declare their service dependencies dynamically rather than requiring rigid, manual boot sequences. They automatically activate when their required services are present and deactivate if those services disappear.
Temporal Composability: Cordis tracks runtime registrations—such as event listeners, prompt sections, and tool schemas—ensuring that removing a plugin cleanly unwinds its effects without leaving orphaned listeners or broken states.
Traceable Executions: Every operation performed by the model, including system prompts, reasoning traces, tool execution requests, sub-agent schedules, and context injections, is recorded in an append-only session log. Developers can inspect these records using the built-in Trajectory view to search, resume, fork, or replay execution streams.
Products, Services, and Interfaces
DeepSeek Harness delivers a local-first development experience complete with a web interface, programmatic interfaces, and multiple operational modes.
1. The Local Web Interface and API
DeepSeek Harness provides a browser-based user interface backed by a local HTTP API running on 127.0.0.1:3080. Developers can launch the system locally via Node.js using npx @deepseek-ai/dsh web or by cloning the repository from source, granting a live dashboard view of token consumption, execution times, cache hit rates, and active turns.
2. Multiple Runtime Modes
The harness ships with distinct operating profiles tailored to different development and benchmarking workloads:
* Standard Mode: A fully featured coding agent setup equipped with file editing, shell access, file and web search, skill libraries, planning frameworks, goal tracking, and workflow sub-agents.
* Code Mode: Extends Standard mode capabilities by exposing tools through a specialized Code Mode SDK, permitting the language model to orchestrate multi-step operations efficiently inside a single TypeScript program.
* Minimal Mode: A lightweight, two-tool coding agent configuration optimized for speed and strict benchmarking, retaining only a persistent bash shell and a file editor.
* Creator Mode: Designed specifically for agent developers and prompt engineers, this mode provides runtime inspection tools, in-memory plugin experiments, and preset-authoring guidance to help build custom agent configurations.
3. Data Processing and Local-First Security
DeepSeek Harness is engineered to run locally, processing user inputs, model outputs, session context, attachments, file paths, and execution logs directly on the host device without uploading data to external servers by default. Because coding agents operate with the ambient authority of a developer's shell—potentially accessing sensitive repositories, internal services, and credentials—the framework includes OS sandboxing layers designed to isolate execution environments.
Please note: This information is generated by AI. For legal use, verify with company sources.
