Harness profile
DeepSeek Harness: an open-source agent harness where everything is a plugin
DeepSeek Harness (dsh) is DeepSeek's open-source agent harness, in public preview, built on the idea that every feature is a plugin. It suits developers who want a self-hosted coding and task agent they can reshape, running DeepSeek models or any provider that speaks a supported API.
Explained from public sources. Mai has not built this yet.
On this page
In plain English
DeepSeek Harness, or dsh, is DeepSeek’s own open-source agent software. You run it on your own computer, open it in a browser or a desktop app, point it at a project folder, and ask it to do work: read and edit files, run commands, plan a task and hand parts of it to helpers. Almost every feature is a plugin you can add, remove or write yourself, including by asking the agent to build one. It is a preview, so it is changing quickly.
How it works
The harness is built on Cordis, a framework DeepSeek describes in an accompanying paper, and its organising rule is that everything is a plugin. Tools, skills, parts of the interface and behaviours such as scheduling or sub-agents are all plugins, installed from a plugin manager or built through chat in a “Creator mode”. Community plugins are shared on GitHub under the dsh-plugin topic.
It is open source under the MIT licence and runs on Node.js. The quickest start is npx @deepseek-ai/dsh web, which serves a web interface on your own machine. You can also build it from source with pnpm, download a desktop app for macOS (Apple silicon) or 64-bit Windows, or drive it from Python with the deepseek-harness-sdk package, which bundles its own runtime and starts a harness process for your code.
Use it from: Web UI, Desktop app, Python SDK. Then dsh core. Then Plugins: Agent loop, Sub-agents, Scheduled tasks, Terminal. Then Reach: Workspace files, MCP servers, Model providers. Then Approval.
The main parts, as the docs describe them:
- Workspace and tools. You add and select a project workspace before you can start a session. The agent can read and edit files there, run commands, delegate work and keep a plan. The web interface asks for your approval before any operation the active permission policy requires.
- Models. DeepSeek is built in and needs only an API key. Other providers, including Anthropic, OpenAI, Kimi and GLM, come from a catalogue, and any relay, company gateway or self-hosted server can be added if it speaks OpenAI Chat Completions, OpenAI Responses, Anthropic Messages or the Gemini API. Keys are stored write-only in a credentials file.
- MCP. The harness can start a local MCP server or connect to a remote one over Streamable HTTP, and exposes its tools to the agent. The docs include optional reference set-ups for third-party memory servers, clearly marked as examples rather than endorsements.
- Scheduling. Session reminders are recurring or one-off tasks you create by asking the model, delivered into an existing or new conversation.
- Inspection. Trajectory views show a timeline of each run with its tool calls and details, for troubleshooting.
What it does best
The plugin-first design is the main attraction. If you want to change how the agent works, add a tool or reshape the interface, the harness is built for exactly that, and the ability to build a plugin by describing it in chat lowers the bar. Provider support is broad despite the name, so you are not tied to DeepSeek models. The run timelines make it easier to see what the agent did and why, and the Python SDK makes it usable from scripts and pipelines as well as from a browser.
Where it stops
It is a developer preview. The README warns in capitals that there will be breaking changes, and the official page says core plugins and APIs are still evolving. Several official plugins, including agent teams and automatic approval review, are marked experimental. Its focus is a workspace on one machine; it does not present itself as a multi-user service, and the docs on public deployment are about safely reaching your own instance through a reverse proxy rather than serving a team. The guides do not describe a built-in long-term memory; the documented route is an optional third-party memory server over MCP.
Where it goes wrong
The scheduling docs are frank about limits. Reminders only fire while the host is running. After a restart, a recurring task only sends its latest missed occurrence, messages waiting in a conversation inbox are lost when the host closes, and a crash can repeat a delivery. A delivery record shows that a message arrived, not that the work was finished. Exposing the web interface is another risk point: the startup address carries a launch token that acts as a credential, so the docs say to treat it as a secret and keep the browser connection on HTTPS. With everything being a plugin, a community plugin is code running inside your agent, and deserves the same review as any dependency.
How Mai AI uses it
I would reach for DeepSeek Harness when a technical team wants a self-hosted agent they can take apart and rebuild, and is comfortable with preview software. For a client system that has to run unattended, the honest list of scheduling limits matters: I would put reminders and long jobs on infrastructure that survives restarts and checks that work actually finished.
The ideas I take from it are the plugin boundary for every capability, which keeps a system easy to change, and treating “message delivered” and “job done” as different facts. The first shapes how I build the tool layer and MCP; the second is central to heartbeats and schedules and long-running jobs. For how these pieces come together around one business, see how Mai AI designs custom AI operations systems.
Questions and answers
Is DeepSeek Harness free and open source?
Yes. DeepSeek publishes it under the MIT licence. It is a preview release, and the README warns that there will be breaking changes. You still pay for the model API you connect.
Does DeepSeek Harness only work with DeepSeek models?
No. DeepSeek is built in, and other providers such as Anthropic, OpenAI, Kimi and GLM come from an installed catalogue. You can also add any gateway or self-hosted server that speaks the OpenAI, Anthropic or Gemini API formats.
How do I run DeepSeek Harness?
With Node.js installed, run npx @deepseek-ai/dsh web to start a local web interface. DeepSeek also offers desktop downloads for macOS on Apple silicon and 64-bit Windows, and a Python SDK that runs the harness from your own code.
Does DeepSeek Harness support MCP?
Yes. It can launch a local MCP server or connect to one over Streamable HTTP, and it exposes the server's tools to the agent with a prefix naming the server.
Can DeepSeek Harness run scheduled tasks?
Yes, through session reminders that you create by asking the model. The harness host must stay running for them to fire, and the docs say delivery is not guaranteed to be exactly once.
Patterns it is known for
- PatternThe tool layer and MCP
AI agents reach real systems through tools: named functions such as "find invoice" that the model asks to run, while ordinary code does the actual work. MCP, the Model Context Protocol, is an open standard for offering those tools and data to any AI application in the same way.
- PatternSub-agents and parallel work
Sub-agents are helper agents that a lead agent starts for one well-defined piece of a larger job, each with its own fresh context, instructions and tools. Several can run at once and report back short findings, which makes broad tasks faster, at the price of more model usage.
- PatternHeartbeats and schedules
AI agents work without being asked when something other than a person wakes them: a schedule, a regular heartbeat check or an incoming event. Each wake-up gives the agent a fixed job, a time budget and a way to report back, so it can chase, prepare and flag things before anyone thinks to ask.
- PatternHuman approval gates
Human approval gates sort every action an AI agent can take into three levels: draft only, needs a person to approve, or allowed to run automatically. The agent prepares, a person signs off anything that matters, and the system refuses any action that does not match what was approved.
- PatternTraces and audit logs
You see what an AI agent did and why by recording a trace of every run, each model call, tool call and decision as a timed step, plus a separate audit log of every action that changed something and who approved it. Traces are for debugging; the audit log is the permanent record.
Related entries
- Harness profileClaude Code and Agent SDK
Claude Code is Anthropic's agentic coding tool that reads a codebase, edits files and runs commands from the terminal, IDE, desktop app or browser. The Claude Agent SDK packages the same agent loop and tools as a Python and TypeScript library, for building your own agents on Claude.
- Harness profileOpenClaw
OpenClaw is an open-source, self-hosted AI assistant that runs on your own machine and answers you in the chat apps you already use, such as WhatsApp, Slack, Teams and Telegram. It is best for one person or a small trusted team who want an always-on assistant with scheduled jobs, skills and tools under their own control.
Sources
- DeepSeek Harness (official page)deepseek.com
- deepseek-harness on GitHub (README)github.com
- Web UI guidedeepseek-harness.github.io
- Configure modelsdeepseek-harness.github.io
- Memory MCPdeepseek-harness.github.io
- Session remindersdeepseek-harness.github.io
- Python SDKdeepseek-harness.github.io
- Public deploymentsdeepseek-harness.github.io