BEGINNER · 15 min 01
What Is DeepSeek Harness?
From models to Agents to the Harness, build a concept map you won't confuse.
STEP 1 Read →
DeepSeek Harness field guide · verified 2026-08-15
From first launch to a plugin runtime.
A learning path in reading order: run dsh, understand the Agent Loop, find trustworthy Plugins, then extend it yourself.
$ npx @deepseek-ai/dsh web - 16
- Core tutorials
- 1805
- Plugins tracked
- 20
- Curated highlights
- 15
- Plugin categories
What is DeepSeek Harness
Not a new model — a runtime that lets models work
DeepSeek Harness (the dsh command) is an open-source Agent Harness from DeepSeek AI: it organizes workspaces, tools, permissions, sessions, model adapters and interfaces so a model can complete multi-step tasks. It is currently in developer preview.
Everything is a Plugin
Model adapters, tools, the Session Log and the Web UI are all pluggable. You extend behavior by composing plugins, not by patching an irreplaceable core.
Read →A complete Agent runtime
Workspaces, tools, permissions, sessions and the Agent Loop work out of the box, so the model can read real projects and finish multi-step tasks.
Read →Permissions stay transparent
Every tool call has a permission boundary and an event log. Plugin entries show source status, plus permission and security notes where we have independently reviewed them.
Read →Quick start
First real task in ten minutes
Get one real response first, then open the runtime layer by layer. Every step lists the expected result and the failure branches.
1
Prepare the runtime
Install Node.js 22 or later, and have a DeepSeek API Key or a compatible endpoint ready.
node --version # v22+ 2
One command to boot
No global install — start the Web UI directly and finish the model setup wizard.
npx @deepseek-ai/dsh web 3
Finish a first task
Open the Web UI in your browser, create a workspace, and let the Agent complete a first multi-step task.
http://127.0.0.1:3080 The 101 path
Not a doc maze — a route
16 core tutorials in reading order, each tagged with level, time, verification date and primary sources.
BEGINNER · 20 min 02
Launch DeepSeek Harness in 10 Minutes
Start the Web UI with the official npm command, check the port, and prepare your first session.
STEP 2 Read →
BEGINNER · 22 min 03
Configuring Models and Compatible Endpoints
Understand the boundaries of Providers, API Keys, and custom OpenAI-compatible Endpoints.
STEP 3 Read →
BEGINNER · 25 min 04
Workspaces, Sessions, and Permission Approvals
Understand what the Agent can see, which operations require approval, and how to set safe trial boundaries.
STEP 4 Read →
INTERMEDIATE · 24 min 05
How to Tell Plugin, Tool, Skill, and MCP Apart
The four most easily confused extension concepts, explained at once by responsibility and runtime boundary.
STEP 5 Read →
INTERMEDIATE · 28 min 06
Profiles and Bundles
Understand how dsh composes the plugin tree layer by layer at startup, and the difference between web and headless.
STEP 6 Read →
BEGINNER · 28 min 07
Troubleshooting Installation and Startup
Locate problems along the shortest path, from Node, ports, and networking to model configuration and workspace permissions.
STEP 7 Read →
INTERMEDIATE · 30 min 08
Installing Third-Party Plugins Safely
Four checks — source, permissions, versioning, and rollback — to reduce the risk of adding third-party code to the Agent Runtime.
STEP 8 Read →
INTERMEDIATE · 30 min 09
Use Plan Mode to Build an Executable Task Contract
Agree on goals, boundaries, acceptance evidence, and rollback before the Agent edits files.
STEP 9 Read →
INTERMEDIATE · 32 min 10
Build Feedback Loops and Independent Review
Use tests, logs, static checks, and a separate reviewer as external evidence.
STEP 10 Read →
ADVANCED · 35 min 11
Manage Context with Compaction, Sessions, and Forks
Preserve decisive facts, remove noise, and isolate alternate work paths.
STEP 11 Read →
INTERMEDIATE · 35 min 12
Build Project Memory with Skills
Load reusable workflows on demand instead of placing an entire handbook in every prompt.
STEP 12 Read →
ADVANCED · 38 min 13
Permissions, Sandbox, and Runtime Invariants
Separate soft instructions, human approval, filesystem enforcement, and tool policy.
STEP 13 Read →
ADVANCED · 42 min 14
Design Tools the Model Can Use Correctly
Build reliable names, schemas, canonical outputs, errors, and idempotent effects.
STEP 14 Read →
ADVANCED · 40 min 15
Subagent Orchestration: Split, Isolate, and Synthesize
Parallelize only independent work, with explicit ownership and acceptance evidence.
STEP 15 Read →
ADVANCED · 55 min 16
Build Your First Cordis Tool Plugin
Walk from apply(ctx) through configuration, tool registration, result observation, and unload verification.
STEP 16 Read →
Plugin radar
Before you grab a capability, see what permissions it asks for
Directory status and our editorial status are shown separately, with permission and security notes on reviewed highlights. A green label is not a permanent safety certificate — the original repository is always the final check.
PLG-001 DIRECTORY VERIFIED
dsh-vision-toolkit
Anionex
Adds image Q&A, long-screenshot OCR, UI reconstruction and visual diffing to a text-only model.
Skills & Agents TypeScript ★ 283
PLG-002 DIRECTORY VERIFIED
dsh_workflow
icetomoyo
Turns multi-agent orchestration into workflows that can be saved, governed, observed and resumed.
Skills & Agents TypeScript ★ 52
PLG-003 DIRECTORY VERIFIED
dsh-visualize
Nagi-ovo
Lets the model generate interactive HTML cards and visualizations inside the conversation.
Skills & Agents TypeScript ★ 67
PLG-004 DIRECTORY VERIFIED
DSH-better-sidebar
omdsh-dev
Brings files, terminal, Git and sub-agent entry points into the Web UI sidebar.
Other Plugins TypeScript ★ 590
PLG-005 DIRECTORY VERIFIED
dsh-security-audit
omdsh-dev
Checks the local dsh configuration, plugins and sensitive-file exposure risks.
Security & Policy TypeScript ★ 9
PLG-006 DIRECTORY VERIFIED
plugin-template
omdsh-dev
A template repository for getting started with DeepSeek Harness Plugin development.
Developer Tools JavaScript ★ 5
Skills & Agents 395 UI & Clients 188 Productivity 55 Developer Tools 75 MCP & Tools 51 Vision & Media 80 Git & DevOps 48 Web & Search 29 Memory & Context 54 Models & Providers 48 Sandbox & Runtime 20 Security & Policy 22 Data & Storage 8 Lists & Resources 30 Other Plugins 738
⚠️ Before installing a third-party Plugin, read the safe install guide: confirm the package name matches the repository, review the permission declarations, and prefer source-verified projects.
Core concepts
Align the concepts first
Six keywords that show up across the whole site. Come back here whenever a tutorial loses you.
Plugin
An extension unit that mounts interfaces, tools or adapters onto the runtime.
Cordis
The underlying plugin framework that drives DeepSeek Harness.
Agent Loop
The cycle of model generation → tool call → observation → next step.
Session
The conversational context of one task: messages, events and state records.
Tool & permissions
The capabilities a model may call, with boundaries — every call is auditable.
Profile / Bundle
A pre-composed set of plugins and configuration that switches runtime shape in one move.
Field notes
Writing the changes down
Architecture observations, Plugin methodology and release readings — kept out of the core tutorial path.
Architecture 01
Everything is a Plugin Is More Than a Slogan
Understand DeepSeek Harness's Plugin-first design through replaceable capabilities, event extension points, and configuration layers.
8/14/2026 Read →
Site Notes 02
How We Organize Plugins Without Creating Yet Another Noisy Directory
A complete index, editorial picks, and hands-on testing are three different promises, and they must be displayed separately.
8/14/2026 Read →
Observations 03
How to Read DeepSeek Harness During the Developer Preview
Keep release facts, official architecture, community hands-on reports, and personal inference separate, and you won't get lost amid rapid change.
8/14/2026 Read →
DEVELOPER PREVIEW
Commands expire. Verification records should not.
Every core tutorial carries its last verification date, applicable version and primary sources. When something feels stale, you know where to check.