DeepSeek Harness (dsh): What It Is, How to Install It, Plugins, and Cost

DeepSeek Harness (dsh) is DeepSeek's open-source, MIT-licensed agent harness, in developer preview. Install it with npx @deepseek-ai/dsh web, add a DeepSeek API key, and pick a workspace. Everything in it is a plugin, including subagents and agent teams.

November 19, 2026 · 1 min read
License
MIT
License
Developer preview status
Preview
Developer preview status
Default local Web UI port
:3080
Default local Web UI port
deepseek-flash per 1M tokens, off-peak
$0.15 / $0.60
deepseek-flash per 1M tokens, off-peak

What Is DeepSeek Harness?

Last updated November 19, 2026.

DeepSeek Harness in one paragraph

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI, MIT-licensed and in developer preview. Start it with npx @deepseek-ai/dsh web, enter a DeepSeek API key under Settings, Models, and choose a workspace. The agent reads and edits files, runs commands, delegates work, and keeps a plan, asking before any operation the permission policy marks for approval.

An agent harness is the code around a model that turns it into an agent. It owns the tool loop, the file and shell access, the permission checks, and the context that goes back to the model each turn. DeepSeek built dsh on its Cordis framework, described in the paper A Programming Paradigm for Spatiotemporal Composability, and made every part of it a plugin. See harness engineering for how harness design moves agent scores.

What ships

A local Web UI, a desktop app (DSH Desktop), a Python SDK, other CLI modes, and automation for GitHub review sessions and session reminders.

What to expect

Developer preview. The README states there will be compatibility-breaking changes and asks you to read the safety notice before running it.

How to Install DeepSeek Harness

The npm route needs only Node.js. The command starts the Web UI at http://127.0.0.1:3080 and opens your default browser. Over SSH it prints the host URL instead, since your SSH client owns the forwarded address.

# Run from npm
npx @deepseek-ai/dsh web

# Run without opening a browser
npx @deepseek-ai/dsh web --no-open

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

A fresh Web UI has no workspace selected. Click Choose workspace, add the project directory, and select it; the session composer stays disabled until you do. The docs also cover publishing the Web UI behind a reverse proxy and uninstalling DSH Desktop.

Configure Models in DeepSeek Harness

Open Settings, Models, paste a DeepSeek API key, and save. The model route works immediately, with no server restart. The same screen takes other providers and any custom OpenAI-compatible endpoint, so dsh is not tied to DeepSeek's API.

DeepSeek API models dsh can call (per 1M tokens, off-peak / peak, api-docs.deepseek.com, November 19, 2026)
ModelInput (cache miss)Input (cache hit)OutputContext
deepseek-flash (DeepSeek-V4.1-Flash-1017)$0.15 / $0.30$0.003 / $0.006$0.60 / $1.201M
deepseek-v4-pro (DeepSeek-V4-Pro-0813)$0.66 / $1.32$0.022 / $0.044$1.98 / $3.961M

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Every other hour bills at half. Full rate detail, caching behavior, and API key setup are on DeepSeek API pricing.

DeepSeek Harness Plugins and Creator Mode

Tools, skills, and parts of the interface all load as plugins. The terminal, the agent loop, and subagents are plugins too, so you can replace any of them. DeepSeek lists four official plugins as experimental.

Plugins listed on deepseek.com/en/harness (November 19, 2026)
PluginStatus
Agent teamsOfficial, experimental
Auto approval reviewOfficial, experimental
Scheduled tasksOfficial, experimental
Voice inputOfficial, experimental
Terminal, agent loop, subagentsInstalled by default

Creator mode writes plugins from a chat request. DeepSeek's demo asks for a floating Pomodoro timer; the agent loads its plugin-development skill, writes the package, installs the bundle, and verifies it in 5 minutes 24 seconds. Memory MCP and plugin indexes are documented integrations, and community plugins are found through the dsh-plugin GitHub topic.

DeepSeek Harness Cost

The harness costs nothing. The bill is model tokens. An agent session re-sends its system prompt and tool definitions every turn, so most input lands as cache hits at $0.003 per million on deepseek-flash off-peak. Output dominates the cost: 2M output tokens in a day of agent work is $1.20 off-peak or $2.40 at peak on deepseek-flash.

DeepSeek Harness vs Claude Code on DeepSeek

DeepSeek documents a second route to the same models. Claude Code can call DeepSeek through its Anthropic-format endpoint by setting a few environment variables.

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash[1m]
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-flash
DeepSeek Harness vs Claude Code
DeepSeek HarnessClaude Code
LicenseMIT, open sourceProprietary
InterfaceLocal Web UI, desktop app, Python SDKTerminal, IDE, desktop
Extension modelEverything is a plugin; Creator modeSkills, hooks, MCP servers, subagents
ModelsDeepSeek API, other providers, any OpenAI-compatible endpointAnthropic models; DeepSeek via ANTHROPIC_BASE_URL
StatusDeveloper previewGenerally available

Pick dsh to change how the agent works, since every part of it is a plugin you can swap. Pick Claude Code on DeepSeek for a stable, documented agent loop at DeepSeek token prices. The broader field is ranked on best AI coding agents.

DeepSeek Harness: Pros and Cons

Strengths
  • MIT-licensed and open source
  • Every component is a plugin, including the agent loop and subagents
  • Creator mode builds plugins from a chat request
  • Works with any OpenAI-compatible endpoint, not only DeepSeek
  • deepseek-flash input on a cache hit costs $0.003 per million tokens off-peak
Limitations
  • Developer preview with announced compatibility-breaking changes
  • Agent teams, scheduled tasks, and auto approval review are experimental
  • No public agent benchmark entry for dsh on Terminal-Bench 4.1 as of November 19, 2026
  • Peak-hour DeepSeek rates are double the off-peak rates

Run DeepSeek Harness on Other Open-Weight Models

Because dsh accepts custom OpenAI-compatible endpoints, the model behind it is a setting. Morph serves open-weight coding models on custom kernels at https://api.morphllm.com/v1. GLM-5.3 (morph-glm53-744b) costs $1.19 per million input tokens and $6 per million output with a 1M context.

Base URL: https://api.morphllm.com/v1
API key:  YOUR_MORPH_API_KEY
Model:    morph-glm53-744b

See Morph models for the full lineup and DeepSeek V4 for the model family behind the DeepSeek API.

Frequently Asked Questions

What is DeepSeek Harness?

DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI and released under the MIT license. It is in developer preview, and the README warns of compatibility-breaking changes. The agent works inside a workspace you choose, reading and editing files, running commands, delegating to subagents, and keeping a plan.

How do I install DeepSeek Harness?

Install Node.js and run npx @deepseek-ai/dsh web. That starts the Web UI at http://127.0.0.1:3080 and opens it in your browser. Pass --no-open to skip the browser. To run from source, clone github.com/deepseek-ai/deepseek-harness, then run pnpm install, pnpm run build, and pnpm dsh web.

Is DeepSeek Harness free?

The harness is free and MIT-licensed. You pay for the model it calls. On the DeepSeek API, deepseek-flash costs $0.15 per million input tokens on a cache miss, $0.003 on a cache hit, and $0.60 per million output tokens off-peak, with peak hours billed at double.

Is DeepSeek Harness on GitHub?

Yes, at github.com/deepseek-ai/deepseek-harness. Documentation lives at deepseek-harness.github.io/deepseek-harness. Plugin authors are asked to tag their repositories with the dsh-plugin topic.

How is DeepSeek Harness different from Claude Code?

DeepSeek Harness is open source and plugin-first, with a local Web UI and a desktop app. Claude Code is Anthropic's proprietary terminal agent. You can also run Claude Code on DeepSeek models by pointing ANTHROPIC_BASE_URL at https://api.deepseek.com/anthropic, which DeepSeek documents in its Agent Integrations guide.

Can DeepSeek Harness use models other than DeepSeek?

Yes. The model configuration guide covers other providers and custom OpenAI-compatible endpoints, so any API that speaks the OpenAI format can back it, including Morph's open-weight models at https://api.morphllm.com/v1.

Private deployments

The fastest endpoints are private deployments

Morph's top speeds come from dedicated deployments, not shared public endpoints: speculators trained on your traffic, caching tuned to your workload, and volume discounts over public per-token rates. Over 1 trillion tokens per day run this way.

Talk to us about a private deployment

Give any harness faster code search

WarpGrep is an agentic code search tool that drops into any agent harness. It finds the right files so the model reads less. $0.80 per 100K tokens.

Sources