DeepSeek Harness 2.0 Desktop App: A Hands-On Guide for Coding Agents
How to install and use DeepSeek Harness 2.0's desktop app for coding agents, covering setup, plugins, creator mode, and scheduled automation.
What is DeepSeek Harness, and why does it need a desktop app?
DeepSeek Harness is the application layer that sits around a DeepSeek model and turns its text output into actual work on your machine. The model decides what to do, but the harness provides the tools to read files, edit code, run commands, and check results. Version 2.0 adds an official desktop app for macOS and Windows, plus a web interface for Linux, so you no longer need to run everything through a terminal or browser tab. The desktop build ships with its own runtime, meaning you don’t need to install Node.js separately just to use it.
TL;DR
- The desktop app (Mac version 0.2.0 release candidate 2 at time of testing) is a developer preview, so expect the interface to shift between releases.
- Linux users get a web-based setup instead, run through
npx @deepseek-ai/DSH webafter installing a supported Node version (22.19+ in the 22 series, or Node 24+). - The harness supports both API key billing and DeepSeek account sign-in, which draw from separate funding pools, so it’s worth checking which one is active before running tasks.
- In a real test, the agent fixed a broken revenue calculator by correctly excluding canceled orders and applying actual refunded units, moving the total from an incorrect $977 to a verified $662 net revenue.
- The app can generate self-contained HTML reports with charts and tables, previewable inside the app, that reuse the same corrected calculation logic as the underlying script.
- Plugins and creator mode let you extend the interface itself, including building small custom panels that install without restarting the app.
- An automation tasks bundle supports one-time and recurring scheduled jobs, though the app must keep running in the background for schedules to fire.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How do you install and set up Harness 2.0?
Installation starts at the official DeepSeek Harness download page, where Apple Silicon Mac and 64-bit Windows builds are available. Run the installer for your platform. Because the desktop app bundles its own runtime, there’s no separate Node.js install required.
Linux doesn’t get a native desktop build. Instead, the documented path is the web interface: install a supported Node version, then run npx @deepseek-ai/DSH web. This starts a local server, typically on port 3080, and opens it in your browser. Adding --no-open (referred to in the walkthrough as “the dash-no option”) stops it from automatically launching a new tab.
Once the app is open, go to Settings, either through the “More” menu or under Models, to configure the DeepSeek provider. You can enter an API key directly, which bills through standard API usage, or sign in with a DeepSeek account, which draws from a separate, account-funded quota. These are distinct billing paths, so confirm which one is selected before running anything that could rack up cost. As with any API key, keep it out of screenshots and project files you plan to share.
One setting worth flipping on early: under General, enable “Show coding view.” This exposes the coding trajectory, file diffs, and agent presets, which are central to actually understanding what the agent is doing rather than just reading a final chat response.
How do you give Harness a coding task?
Add a project folder using the plus button in the sidebar, then separately select that folder as the active workspace above the composer. These are two different actions. It’s easy to send a prompt against the wrong workspace if you skip checking the name first.
Model and mode selection happens per task. In one test run, the setup used DeepSeek V4.1 Flash with “max” thinking, “workspace write” access, and “standard” mode. Max thinking is a reasoning setting that affects response time and usage, so it’s a deliberate choice if you’re comparing performance against someone else’s run. “Workspace write” allows editing files and running project checks, “read only” is for inspection-only tasks, and “full access” extends permissions further. Whenever the app requests approval for an action, it’s worth actually reading what operation it wants to perform before approving it.
A useful way to evaluate a coding agent is to give it a task with a verifiable right answer. In one demonstration, a small fictional e-commerce project (a CSV of 16 orders, a Python reporting script, and a test suite) contained two deliberate bugs: it counted canceled orders as revenue, and it failed to subtract returned units correctly. The business rules were explicit: canceled orders contribute nothing, refunded orders still count but their units need to be subtracted using the actual refunded quantity, and all money math should stay in integer cents for precision.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Given a prompt to inspect the project, explain the bugs, fix the script, run the existing tests, and print corrected totals (while explicitly preserving the CSV and test files), the agent identified both issues: it excluded canceled rows and pulled real refunded-unit values instead of treating refunds as zero. All four original tests passed afterward. The corrected output showed 14 orders, $83 gross, $141 in refunds, $662 net revenue, and 41 net units, versus an original, incorrect $977 figure. Independently rerunning the tests (python3 -m unittest -v) and checking the CSV and test files confirmed nothing else had changed.
Can it generate reports and interactive dashboards?
Yes. In the same conversation, you can ask Harness to build a responsive HTML report from the same data, reusing the already-corrected calculation logic. A reasonable prompt specifies the exact fields needed (date range, order count, gross revenue, refunds, net revenue, net units, a product revenue chart, a product table) and constraints like “self-contained, no external libraries or downloaded assets,” which keeps the output easy to open directly without depending on an external service.
In testing, this produced a report.html file, a build_report.py generator script, and a calculations.md explanation file. The report rendered inside the desktop app’s right sidebar under “workspace files,” with a full-screen option for easier review. The number in the report ($662 net revenue, 41 net units) matched the independently verified calculation, confirming the report wasn’t just cosmetically correct but numerically consistent with the underlying fix.
One rough edge appeared here: the agent corrected a syntax error while building the generator, then spent extended time on a slow browser screenshot step and repeated checks. Once the deliverable was confirmed working, explicitly telling the agent to stop that job and finish up was necessary to move forward efficiently.
A follow-up request added an interactive product filter with a reset control and chart highlighting. This is where precision in the deliverable mattered: when filtering to a single product, the interface needed to clearly distinguish the “whole shop” totals from the filtered figures, since a working filter that doesn’t relabel totals can still mislead a reader. Testing individual product filters produced consistent numbers (for example, one product showed $200 net revenue and eight units, another showed $180 and three units), and resetting the filter correctly returned the view to the full 14-order, $662 shop total.
What do plugins and creator mode add?
Harness ships with an official plugins tab containing optional bundles that need to be explicitly enabled even though they come pre-included. An “installed” tab tracks anything added manually, via package name, repository, archive, or local path. After installing a plugin, you still need to check that it’s enabled separately, since installation and activation are different steps. Because plugins execute code with your permissions, checking the source before installing anything unfamiliar is a real security consideration, not just a formality.
Creator mode goes a step further: you describe a capability, where it should appear in the interface, and how to verify it works, and the agent attempts to build or configure the plugin itself. This is explicitly experimental, and a new plugin may require an app restart to appear. Keeping the first version of any custom plugin small enough to actually inspect makes it realistic to review before trusting it.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
In one test, a request for a small “report checklist” panel (three checkable items, a completed counter, and a reset button) led the agent to first spend time exploring the runtime environment before producing a working plugin. After reviewing the generated source, installing it via the local plugin folder path and clicking “enable now” made the panel appear in the sidebar without restarting the app. The checklist correctly tracked state when switching between panels, and reset cleared it back to zero. Because the state lives in memory rather than being persisted, it doesn’t survive a full app reload, a reasonable tradeoff for a lightweight demo plugin.
How does scheduled automation work?
A separate “automation tasks” bundle, also enabled from the plugins list, lets you request one-time or recurring scheduled jobs directly in conversation. A practical example would be regenerating a weekly report automatically whenever a source CSV updates. Specifying the input, desired output, schedule, and time zone explicitly (for example, “America/New York” rather than assuming local time) avoids ambiguity.
In a short test, a one-time task scheduled 90 seconds out instructed the agent to read the sales data, rerun the summary, and write a reminder file with verified figures. The saved task showed the correct one-time schedule and local time, and after it fired, a “delivery records” tab logged that the instruction had reached the conversation. The resulting file contained the correct net revenue and unit figures. Importantly, a delivery record only confirms the instruction reached the conversation inbox, not that the agent actually completed the work, so checking the real output afterward still matters.
Practical caveats: scheduled tasks return to their original conversation thread, and the app needs to keep running for schedules to execute. Closing the desktop window may leave the backend running, but fully quitting the app stops scheduling entirely.
What’s the difference between Harness modes?
Harness offers several task presets, chosen when starting a new conversation:
- Standard is the default for coding, file work, and research, and supports sub-agents for bounded pieces of work.
- PTC (programmatic tool calling) lets the agent write TypeScript that coordinates and combines multiple tool calls, which can help with batched operations but isn’t automatically faster or cheaper.
- Minimal is a lighter baseline with a persistent shell and fewer standard features.
- Creator adds the tools and guidance needed for building plugins, as covered above.
Switching the default preset doesn’t retroactively change an already-running session. There’s also an experimental “agent teams” bundle for multi-agent collaboration with shared tasks and messaging, though teammates share the same working directory, so overlapping files need clearly divided responsibilities.
Frequently Asked Questions
Is DeepSeek Harness 2.0 free to use?
Harness itself is the application layer; actual usage cost depends on which model access path you connect, either API key billing or a signed-in DeepSeek account with its own quota. Check which one is active in settings before running tasks, since they draw from separate funding sources.
Does the desktop app require installing Node.js separately?
No. The packaged Mac and Windows desktop builds include their own runtime. Node is only required for the Linux web setup, which needs Node 22.19+ (within the 22 series) or Node 24 and above.
What’s the difference between “workspace write” and “full access” modes?
Workspace write allows the agent to edit files and run commands within the selected project folder, suitable for most coding tasks. Full access extends permissions more broadly, while read-only mode restricts the agent to inspection without making changes.
Can Harness plugins run arbitrary code?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Yes, which is why the install dialogue explicitly warns that plugins run with your permissions. Checking the source of any plugin before installing it, especially ones from creator mode or third-party sources, is a real precaution rather than an optional step.
Will a scheduled automation task keep running if I close the app?
Closing the desktop window may leave the backend active, but fully quitting the application stops scheduled tasks from firing. For recurring jobs you actually rely on, the app needs to stay running continuously.
