DeepSeek Harness

What it is

DeepSeek Harness, or dsh, is DeepSeek's own coding agent. A "harness" is the program wrapped around an AI model that lets it read files, run commands and edit code. Claude Code is Anthropic's harness for Claude. This is DeepSeek's, and it's free and MIT-licensed.

Two things set it apart. First, it runs in a web page or a desktop app, not a terminal. Second, almost everything in it is a plugin, built on a framework called Cordis that can load and unload plugins while the app keeps running. You can add features, or have the agent write its own plugins, without restarting.

It works best with DeepSeek's own models, but you can also connect Anthropic, OpenAI, Kimi, GLM, any OpenAI-compatible service, or a model running on your own machine. It came out on August 13, 2026 as a "developer preview", and the README warns in capital letters that there will be compatibility-breaking changes. Feedback goes to GitHub Discussions, which has passed 10,500 threads, because GitHub issues are switched off.

What actually happens when you use it

1. Start it. Run one command and it opens in your browser, or install the desktop app.
2. Add a model. Paste a DeepSeek API key in Settings, or add another provider.
3. Pick a workspace. Choose the project folder the agent may work in.
4. Ask for something. The agent reads and edits files, runs commands, keeps a plan, and can hand parts of the job to helper agents. It asks before risky steps, depending on the permission mode you choose.
5. Look back at everything. Every prompt, tool call, result and helper agent is recorded in a session log. A Trajectory view lets you inspect it, and you can resume, fork or replay a session.

Shell commands can run in a sandbox that limits what they can write: read-only, project folder only, or full access. It uses the operating system's own tools on macOS, Linux and Windows.

What people like

• You can see exactly what it did. One Hacker News user called the full, replayable record of everything the model saw a killer feature.
• It's very cheap to run. One Reddit user built a simple single-page app for about six cents, with almost all of the input served from cache at a lower price.
• It works with local models. A user running a small Qwen model on their own machine said it works great for small Python projects and felt fast. Another runs it against a home AI server.
• Helper agents can talk back. One user likes that the main agent and its helpers can message each other mid-task, so you can redirect a helper without starting over.
• A friendly app, not a terminal. A Reddit user described it as a mix of the Codex or Claude app and DeepSeek's chat site. Terminal fans wish it had a proper command-line version.

What people run into

• Simple jobs can take ages. In one popular help thread, a quick code question took over 30 minutes and 40 steps, where Claude took a minute. Most of the time went on searching across several repos. Pointing it at the right folder fixed it for the poster, and others suggest turning the reasoning level down.
• It isn't built for beginners yet. One newcomer's first hour went badly. Even some who defended it in the replies agreed it isn't really aimed at beginners yet.
• Platform gaps. There's no Linux desktop app, though the web version works there. On Windows, one developer found the sandbox permanently rewrote their project folder's permissions, and another user hit the same thing.
• Plugins carry risk. A community store already lists thousands of plugins, and they run real code on your computer. One commenter worries plugins might be malware or get abandoned.
• DeepSeek accounts get suspended. Several users report their DeepSeek accounts, including API access, being suspended with no reason given and no working appeal. That isn't the harness's fault, but it stops you working.

Before you install

• Compared with Claude Code: cost. The software is free and DeepSeek's API is far cheaper per token than Claude's. DeepSeek's prices are halved outside peak hours, and its web search tool is free through the harness until November 1, 2026. A local model costs nothing per token at all.
• Compared with Claude Code: privacy. The code is open, so you can read what it does. But DeepSeek's privacy policy says it stores data in China and may use your inputs to improve its models, with an opt-out by email. The US just banned DeepSeek from federal civilian agencies, so check your clients' rules.
• It uploads session logs by default. When you use DeepSeek's official API, it sends a copy of each session's log along with your requests. Turn it off under Settings → General → Upload Session Log when using the official model API. Sending feedback or rating a message in the app also uploads that session, unless you set DSH_TELEMETRY_DISABLED=1 before starting it.
• Compared with Claude Code: maturity. The safety notice says it hasn't had a security audit and isn't production-ready, and recommends running it in a throwaway virtual machine or container. New versions come out every few days and are all marked pre-release.
• Beware lookalikes. A third-party site has been repackaging it and ranking above DeepSeek in search. Use deepseek.com, npm or GitHub.

The short version

DeepSeek Harness is the most interesting new open-source agent of the year: a browser-based app with full replay of every step, a plugin system you can reshape, and running costs that make Claude look expensive. It's also a ten-week-old preview with slow runs on simple jobs, no security audit, and a default that sends your session logs to DeepSeek. If you're a vibe coder who relies on the agent getting it right, stick with Claude Code for real work and give this a few weeks. If you like to tinker, try it on a side project in a throwaway environment, with the log upload switched off.