DeepSeek Harness

Name: DeepSeek Harness
Author: DeepSeek

AI Agents

DeepSeek

FreeOpen Source

About

Open-source agent harness by DeepSeek built on the Cordis plugin system. Everything is a plugin — models, tools, skills, sessions, sandboxes, storage, scheduling, and UI are all composable and replaceable

Visit

Share this tool

Twitter LinkedIn Facebook

Our Verdict

Game-Changing

Redefining agent architecture with a truly modular plugin system

DeepSeek Harness brings a paradigm shift to agent development by building everything on Cordis's plugin architecture. Every capability — from the model backend to the UI — is a swappable plugin, making it the most modular agent platform available. While still in developer preview with a nascent plugin ecosystem, its architectural elegance and backing by DeepSeek position it as a potential game-changer in the agent development space. The trajectory view, multiple agent modes, and session replay capabilities already provide a solid foundation for both development and production use.

Best for

•Developers building custom AI agents
•Researchers exploring agent architectures
•Teams wanting full control over agent stack

Consider alternatives if

•AutoGPT (mature agent framework with broader ecosystem)
•Dify (visual agent builder with more GUI tools)

Try DeepSeek Harness

Supported Platforms

Web AppWindowsmacOSLinux

Available platforms include Web App, Windows, macOS, and Linux.

Key Features

Cordis plugin architecture — everything is a plugin

Multiple agent modes: Standard, PTC (Code), Minimal, Creative

Trajectory view for full session replay and debugging

Plugin ecosystem with model, tool, skill, sandbox, storage, scheduling, and UI plugins

Web UI served at localhost:3080 by default

Session logging with append-only event stream for recovery, forking, and replay

Open source under MIT license

Pricing

free

Free and open source (MIT)

paid

N/A — fully open source

freemium

N/A

Use Cases

Building and deploying custom AI agents with a modular plugin architecture

Developing and testing new agent capabilities as Cordis plugins

Running coding agents in multiple modes (Standard, PTC, Minimal, Creative)

Debugging agent behavior through Trajectory session replay

Creating custom agent presets and configurations

Pros

Truly modular architecture — every capability is a plugin, enabling unlimited customization

Built on Cordis, a well-designed plugin system with academic foundations

Multiple agent modes cater to different use cases from full-featured to minimal

Trajectory view provides excellent debugging and observability

MIT licensed — fully open source with no restrictions

Backed by DeepSeek, a leading AI research company

Cons

Developer preview — rapid iteration with breaking changes expected

Very new project (August 2026) — ecosystem and documentation are still maturing

Requires Node.js runtime — not a standalone binary

Web UI is local-only by default, no built-in cloud deployment

Plugin ecosystem is nascent — limited third-party plugins available

Latest Update

2026-08-13 (Developer Preview)

Get Started with DeepSeek Harness in 5 Steps

Install and start the Web UI

Install Node.js, then run `npx @deepseek-ai/dsh web` to start the Web UI, served at http://127.0.0.1:3080 by default. The dsh process uses its invoking directory as the default filesystem location. To run from source, clone the repository, run pnpm install, pnpm run build, and pnpm dsh web. ![](images/tutorials/dsh-web-ui-welcome.png)

Start Web UI

npx @deepseek-ai/dsh web

Run from source

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

Configure a model

Open Settings → Models, enter a DeepSeek API key, and save it. The model route becomes usable immediately without restarting the server. To add other providers, choose Add provider and select one from the installed catalog (Anthropic, OpenAI, etc.), or choose Add a custom provider for a company gateway or self-hosted OpenAI-compatible endpoint — supply a lowercase Provider ID, base URL, API protocol, credential, and at least one model. You can also fetch available models from the endpoint before saving. ![](images/tutorials/dsh-settings-models.png)

Custom provider YAML (vision model)

llm-pi-ai:
  providers:
    my-gateway:
      apiKeyEnv: GATEWAY_API_KEY
      api: openai-completions
      baseURL: https://gateway.example/v1
      models:
        - id: vision-preview
          input: [text, image]

Choose a workspace

Click Choose workspace, add the project directory where you started dsh, and select it. A fresh Web UI has no selected workspace until you add one — the session composer remains unavailable until a workspace is selected.

Note

A fresh Web UI requires selecting a workspace before the session composer is available.

Run a task in a session

Start a session and send a prompt, for example: "Summarize this repository and identify its main packages." The agent can read and edit workspace files, run commands, delegate work, and maintain a plan. The Web UI asks before operations that require approval under the active permission policy.

Example prompt

Summarize this repository and identify its main packages.

Go further: CLI modes and the Python SDK

Beyond the Web UI, dsh offers CLI entry modes — run a single headless session with `dsh --profile headless "job"` for a one-shot persisted session. For programmatic use, install the Python SDK (`python -m pip install deepseek-harness-sdk`, Python 3.10+, Linux/macOS), set DEEPSEEK_API_KEY, and drive the harness from your own program.

Headless CLI

dsh --profile headless "Inspect the repository and fix the failing tests."

Install Python SDK

python -m venv .venv && . .venv/bin/activate && python -m pip install deepseek-harness-sdk && export DEEPSEEK_API_KEY=sk-...

SDK usage

from deepseek_harness import DeepSeekHarness

with DeepSeekHarness(
    provider="deepseek-official",
    model="deepseek-v4-flash",
    max_tokens=49152,
    cwd="/path/to/workspace",
    session_root="/path/to/sessions",
    cordis="examples/jsonrpc-agent/minimal.cordis.yml",
) as harness:
    result = harness.run(
        "Inspect the repository and fix the failing tests.",
        session_id="example-001",
    )

print(result.final_response)