How to Install and Set Up DeepSeek Harness Locally
A step-by-step guide to installing DeepSeek Harness from source, adding your API key, setting permissions, and configuring plugins.
What is DeepSeek Harness?
DeepSeek Harness is DeepSeek’s own agentic coding system, released alongside the newer DeepSeek V4 Pro model as a developer preview. Instead of being a single tool, it’s built around a plugin architecture where tools, skills, sessions, and even other coding agents like Codex or Claude Code can be plugged in, swapped out, or combined. It ships with a local web app interface, which is the main way most people will interact with it. To install it, you clone the repository, install it from source, and launch the bundled web server on your machine.
TL;DR
- DeepSeek Harness is an agentic coding framework from DeepSeek, currently in developer preview, designed around treating tools, models, and even other agents as interchangeable plugins.
- Installation is done by cloning the GitHub repo and installing from source, which reportedly works reliably compared to other install paths.
- Once installed, running a single command starts a local web app (by default on port 3018, though the port is configurable) that serves as the main interface.
- You need a DeepSeek API key, created from your DeepSeek account’s API keys section, before the harness will let you do anything.
- The harness offers multiple operating modes (full agent mode, code mode, minimal mode, and a creator mode for custom agent profiles) plus four permission tiers ranging from read-only to full unrestricted access.
- Plugins, including web search, model providers, and local model support, are configured through a YAML file, and the project is MIT licensed, so it’s straightforward to extend or swap in other models.
- Early hands-on testing suggests the Flash variant of DeepSeek’s models offers better cost-to-performance for coding tasks than the Pro variant, despite Pro being the flagship.
How do you install DeepSeek Harness?
The recommended installation method is cloning the source repository directly and building from there, rather than relying on a packaged installer. This is the approach that worked without issues during testing, and it’s the one recommended if you hit any snags with other installation routes.
The basic steps:
- Clone the DeepSeek Harness repository from GitHub.
- Install the project from source using its standard build process.
- Launch the web interface with the provided command (referred to as
ds webin the project), which starts a local server.
By default, the web app runs on localhost at port 3018, but you can specify a different port if that one is already in use or you prefer another setup. Because the whole thing runs as a local web app, there’s also the option of deploying it to a remote server, which would let you access your own DeepSeek Harness instance from anywhere rather than just your local machine.
How do you get and configure a DeepSeek API key?
The harness requires a DeepSeek API key before it will run any agentic tasks. To get one, log into your DeepSeek account, navigate to the API keys section, and generate a new key. The first time you open the harness’s web interface, it prompts you directly for this key. Without it, most of the interesting functionality, including web search, stays locked. Once you add a valid DeepSeek key, features like web search get enabled automatically as part of the plugin system.
Depending on which key and account tier you use, you may get access to both the Flash and Pro versions of the model family, letting you switch between them per task or per session.
What are the different modes and permission levels?
After entering your API key, the harness asks you to set up a workspace, essentially a folder on your machine where all files and session data for a given project get stored.
From there, you choose an operating mode:
- Full mode: the complete coding agent, with file editing, shell access, and web search (web search needs to be set up separately even in this mode).
- Code mode: a narrower focus on coding tasks.
- Minimal mode: a stripped-down configuration for lighter tasks.
- Creator mode: lets you build and save your own custom agent profiles.
Alongside modes, there are four permission levels that control how much the agent can do on your system:
- Read-only, which limits the agent to viewing files without changing anything.
- Workspace, which restricts write access to the designated workspace folder.
- Right (a middle tier with somewhat broader access).
- Full access, which removes safety restrictions entirely, comparable to the “dangerously skip permissions” mode found in other coding agents like Claude Code.
Choosing the right permission tier matters if you’re running the harness on a machine with sensitive files nearby, since full access effectively gives the agent free rein over your system within its process permissions.
How does the plugin and model configuration work?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Everything in DeepSeek Harness, from tools to skills to sessions, is designed as a plugin. In the settings panel, you’ll find a list of available plugins, some enabled by default and some not, which you can toggle on or off depending on what you need.
Model and plugin configuration lives in a YAML file. This is where you can see the currently selected provider (DeepSeek’s own API by default), the active model, and settings like reasoning effort for models that support flexible reasoning (a feature introduced with the Pro and Flash versions of V4). Because the project is MIT licensed, adding other model providers, including locally hosted models, is technically straightforward. Support for running local models through the harness was demonstrated on a multi-GPU DGX setup, showing the framework isn’t locked to DeepSeek’s hosted API.
The web search plugin is worth calling out specifically: it only activates once you’ve supplied a valid DeepSeek API key. Skip that step and web search stays unavailable, even if everything else is configured correctly.
Is DeepSeek Harness worth setting up right now?
For anyone already experimenting with agentic coding tools like Codex or Claude Code, DeepSeek Harness is worth trying, with the caveat that it’s a developer preview and will have rough edges. In hands-on testing, the interface felt fast and polished, with detailed logging of what the agent is doing at each step, live stats on tokens per second, cache hit rate, and turn count. Cache hit rates in particular stood out, with figures in the 95 to 100% range observed during a coding task, which is unusually high compared to other harnesses.
The tradeoff is token efficiency. A moderately complex task (building a real-time ISS tracker with a live API feed and a 3D globe visualization) burned through roughly 20 million tokens over two turns and about 35 minutes, with output alone hitting around 240,000 tokens. That’s a lot of token consumption for what is, on paper, a fairly contained project. Open models in general have a reputation for being less token-efficient than some closed alternatives, and this pattern held here too.
On model choice, the Flash variant came out ahead of Pro in practical coding tests, offering a better balance of speed and cost even though Pro is the more capable flagship model. If you’re optimizing for value rather than raw capability, Flash is the one to reach for first.
Frequently Asked Questions
Do I need a DeepSeek API key to use the harness?
Yes. The web interface prompts for it on first launch, and core features, including web search, won’t function without a valid key generated from your DeepSeek account.
What port does the DeepSeek Harness web app run on by default?
It runs on localhost port 3018 by default, but the port number can be changed when you launch the web server.
Can I use other AI models besides DeepSeek’s own inside the harness?
Yes. The harness is MIT licensed and built around a plugin architecture, so you can add other model providers, including locally hosted models, through the YAML configuration file.
What’s the difference between DeepSeek Flash and Pro for coding tasks?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Pro is the larger, more capable flagship model, but in hands-on coding tests, Flash delivered better cost-to-performance, making it the recommended default for most coding work.
Is DeepSeek Harness stable enough for daily use?
It’s currently a developer preview, so expect rough edges and possible instability. It’s functional and fast in testing, but treat it as an early-stage tool rather than a finished product.
