DeepSeek HarnessOpen-source agent

DeepSeek Harness Locally, let AI edit code, run commands, and compose plugins

Free & open source
Edit files & shell
Everything is a plugin

DeepSeek Harness is more than chat: the model reasons; Harness connects files, terminal, and tools to real work. One command starts the local UI — install, modes, and guides begin here.

Install DeepSeek Harness Read the tutorials

Open source MIT
One-command start
Runs locally

One command to start the local Web UI — copy and run in your terminal

npx @deepseek-ai/dsh web

Plugins

All capabilities

Model, tools, session, sandbox, UI…

Cordis

Kernel

Load, unload, wire dependencies only

4 modes

Presets

Standard · PTC · Minimal · Creative

npx

Try in minutes

@deepseek-ai/dsh web on :3080

Agent = Model + Harness

Harness keeps agents working in real environments

The model reasons

Model adapters handle inference — the harness does not replace which LLM you plug in.

Harness connects the environment

Files, terminal, tools, sessions, sandboxes, and the local Web UI so work continues beyond a chat transcript.

Architecture tutorial →

What is DeepSeek Harness?

DeepSeek Harness (DSH) is the execution layer around a model. Instead of a hard-coded agent loop, Cordis loads plugins for models, tools, skills, sessions, sandboxes, storage, scheduling, and even the UI. Developers compose behavior in configuration, inspect runs in Trajectory views, and extend the stack with community dsh-plugin packages. This site explains install paths, safe setup, and architecture in plain language — always verify commands against the official GitHub repository.

Install guide →Local Web UI →Run modes →Community plugins →

Three layers you compose

Cordis kernel

Loads, unloads, and wires plugin dependencies only — no hard-coded agent loop in the core.

Run modes →

Capabilities as plugins

Models, tools, skills, sessions, sandboxes, storage, scheduling, and UI collaborate through Cordis.

Plugin hub →

Compose in configuration

Swap a provider or tool without forking source — pick plugins in config and reload.

Install guide →

Why teams evaluate DSH

Composable by design

Replace a model provider or tool without rewriting the harness core — configuration selects plugins.

Observable runs

Append-only session logs capture prompts, tool calls, and injections so you can replay and fork trajectories.

Multiple run modes

Full coding agent, code-mode PTC, minimal dual-tool bench, or creative preset authoring.

Open MIT preview

Study the source, contribute plugins, and keep ownership of how agents touch your repos.

Core capabilities

Everything is a plugin

Capabilities collaborate through Cordis services and events instead of a monolithic agent binary.

From official design notes — rewritten here

Trajectory observability

Inspect what the model saw: system prompts, tool results, sub-agent scheduling, context injections.

Same event stream for resume & replay

Four practical modes

Pick Standard for daily coding, PTC for multi-step code orchestration, Minimal for evals, Creative for presets.

See /modes

Web UI in one command

After Node.js is ready, npx boots a local UI — add an API key, pick a workspace, send a task.

http://127.0.0.1:3080

Where DSH fits

Coding agents

Repo edits & shell
Skills and workflows
Sub-agents for long tasks

Platform builders

Custom presets
In-memory plugin experiments
Config-level composition

Eval & research

Minimal tool surface
Reproducible trajectories
Harness vs model ablation

Four modes at a glance

Need detail? Open the modes page for Standard, PTC, Minimal, and Creative — then jump back to install when you are ready to run.

Four modes at a glance →

Standard

Full coding-agent surface

File edits, shell, search, skills, planning, goals, sub-agents, and workflows. This is the default after you trust the workspace permissions and you have watched Trajectory on a disposable folder. It is the profile that most “just code this repo” prompts expect. Pair it with the workspace warning on the [install](/en/install/) page: your entire home directory is the wrong first target.

Use when: Day-to-day patches, tests, and refactors in a repository whose root you chose on purpose.

PTC

Code-mode multi-step tools

Keeps Standard-class capabilities but presents tools through a Code Mode SDK so the model can author a TypeScript program that composes several operations in one shot. Reach for it when the pain is orchestration — many tool hops that should be one program — not a single chatty call. If preview docs for Code Mode moved, believe GitHub over a blog date on this site.

Use when: Pipelines that should be a short TypeScript script rather than twenty chat turns.

Minimal

Two-tool evaluation harness

Public notes describe persistent bash plus str_replace_editor only. That tiny surface is how you stop lying to yourself during model evals. If web-search, marketplace UIs, and appearance packs stay loaded, you are measuring the harness cocktail. Keep secrets out of the eval folder anyway; bash is still bash.

Use when: Benchmarks, regression of a pinned model, or teaching the difference between model skill and plugin luck.

Creative

Preset authoring lab

Standard powers plus runtime inspection and in-memory Cordis experiments so you can design and document new agent presets without pretending the session is production. Plugin authors and skin authors should live here until unload is boring. See [contribute](/en/contribute/) if you intend to publish a dsh-plugin rather than screenshot a theme.

Use when: Building presets, inspecting Cordis, trying UI skins, writing plugins.

Quick try or install from source

Official developer-preview paths — if flags change, trust the GitHub README over a blog date on this site.

Quick try

With Node.js on PATH, one npx command starts the local Web UI (commonly :3080).

npx @deepseek-ai/dsh web

From source

Clone upstream and follow pnpm build steps — typical for plugin authors and contributors.

git clone https://github.com/deepseek-ai/deepseek-harness

Full install checklist

Get productive in three steps

1

Install & workspace

Node, npx, API key, and a disposable folder — not your entire home directory.
→
2

Pick a run mode

Standard for daily coding; PTC, Minimal, or Creative when the job differs.
→
3

Tutorials & plugins

Walk the tutorial track, then browse curated dsh-plugin packages safely.
→

FAQ

FAQ →

Is DeepSeek Harness the same as DSH or DS Harness?

Yes — DSH and DS Harness are common short names for DeepSeek Harness. On this site we use DeepSeek Harness (dsh) in titles and introduce aliases in the first paragraphs for search clarity.

Is DeepSeek Harness a new DeepSeek model?

No. Agent = Model + Harness: the model reasons; the harness connects files, the terminal, tools, and the UI. This site is about that open-source execution layer.

How do I install DeepSeek Harness quickly?

Install Node.js, then run npx @deepseek-ai/dsh web, open http://127.0.0.1:3080, add a DeepSeek API key, choose a workspace, and send a task. Full steps live on the install page.

What does “everything is a plugin” mean?

Model adapters, tools, sessions, sandboxes, storage, loops, scheduling, and UI are Cordis plugins. You compose them in config instead of forking a monolith.

What is PTC mode in DeepSeek Harness?

PTC is programmatic tool calling: the model writes a short program that composes several tool calls instead of chatting through each hop. Standard is the daily coding surface. Details live on the modes page.

Is there a hosted DeepSeek Harness online?

No cloud Web UI. Start it locally with npx @deepseek-ai/dsh web and open http://127.0.0.1:3080. Project files and API keys stay on your machine. Operator checklist: the Web UI guide on this site.

Where is the official DeepSeek Harness product page?

Product copy: https://www.deepseek.com/harness/. Source, flags, and LICENSE: github.com/deepseek-ai/deepseek-harness. harness-ds.com is a community portal with install, PTC, plugins, and troubleshooting — not a vendor homepage.

Is DeepSeek Harness made by ByteDance?

No. DeepSeek Harness (dsh) is published by DeepSeek AI under MIT. It is unrelated to ByteDance agent products.

How is DeepSeek Harness different from DeepSeek Chat?

DeepSeek Chat is the consumer chat app. DeepSeek Harness is a local developer framework: Cordis plugins, files, shell, and a Web UI on :3080 that you run yourself.

Is MIT license OK for commercial use?

The upstream deepseek-ai/deepseek-harness repository is published under MIT — confirm LICENSE on the tag you pin. Community plugins have their own licenses. This portal does not grant a license.

DeepSeek Harness vs Claude Code, Codex, or OpenCode?

DSH is an MIT plugin-first preview harness. Claude Code and Codex are commercial coding agents; OpenCode is another open coding agent. See the Claude Code comparison page for the workflow matrix we actually maintain.

DeepSeek Harness vs Claude Code — what changes?

DSH centers on npm plugins via Cordis; Claude Code emphasizes MCP-centric tooling. See our comparison page for workflow fit, not hype.

Latest guides

2026-10-09

DeepSeek Harness on macOS: Homebrew, PATH, Full Disk Access

Same npx command as every other OS — on Mac the surprises are which node is on PATH, TCC prompts, and not granting Full Disk Access on day one.

2026-09-28

What's New in DeepSeek Harness 0.1.7 RC

v0.1.7-rc.2 is Pre-release. Scheduled tasks stay off, session logs move to V4, and PTC names become ptc-runtime.