# What Is DeepSeek Harness? The Open-Source Coding Agent Explained > Source: > Published: 2026-08-20 00:00:00+00:00 # What Is DeepSeek Harness? The Open-Source Coding Agent Explained DeepSeek Harness is an open-source, plugin-based coding agent that hit 165,000 GitHub stars in a week. Here's how it works. ## What is DeepSeek Harness? DeepSeek Harness is an open-source coding agent framework built entirely out of plugins. Instead of shipping a fixed set of features like Claude Code or Codex, DeepSeek built the harness itself, including its interface, tool calls, and agent loop, as composable plugins that developers can toggle, replace, or write from scratch. It gained roughly 165,000 GitHub stars within a week of release, signaling strong developer interest in a more open, customizable alternative to locked-down coding agents. ## TL;DR **DeepSeek Harness is built entirely from plugins**, meaning the chat interface, tool calls, context injection, and even the underlying agent loop are all swappable components rather than fixed code.- The project reached **165,000 GitHub stars in about a week**, an unusually fast adoption curve for a developer tool, though the creator notes it’s still rough around the edges. - Unlike Claude Code or Codex, which lock you into specific models, DeepSeek Harness lets you **plug in any provider**, including DeepSeek’s own API, OpenRouter, or other LLMs. - A built-in **creator mode** walks you through building new plugins conversationally, so you don’t need to hand-write extensions to add functionality. - The **trajectory view** gives full auditability into every action the agent takes, showing which plugin triggered which tool call or context injection, a level of transparency not found in comparable tools. - The harness supports **sub-agent delegation**, letting you call Claude Code or Codex from inside a DeepSeek Harness session for tasks better suited to those models. - It’s philosophically closest to **Pi**, an earlier self-extensible coding agent, but takes the plugin-based, self-modifying approach further by making the harness’s own internals editable. ## How does DeepSeek Harness work? The core idea is that nothing in the harness is hardcoded. Every discrete behavior, whether it’s how the agent injects context, how it calls a tool, how it asks a clarifying question, or even how a button renders in the UI, is implemented as an individual plugin. These plugins are composable units that stitch together to form the full agent system. Practically, this means you can go into a settings panel, see a list of installed plugins, and turn any of them off or swap them out. There’s also a plugin marketplace where third-party developers publish extensions, alongside whatever you build yourself. Want vision support for a model that can’t process images natively? There’s a plugin (called ModLens in the demonstrated setup) that acts as a bridge, routing image inputs to a model that can handle them and returning a text description. Want to pull GitHub star counts inside a conversation? That’s a plugin too, and one that can be built from scratch in a single guided session. Getting started involves installing from NPM or running from source, then connecting an LLM. The default model is DeepSeek V3.1, configured via a DeepSeek or OpenRouter API key, but because the harness isn’t tied to one provider, any model can be wired in through the models tab. ## What is creator mode? Creator mode is a built-in workflow for generating new plugins conversationally instead of writing them by hand. You describe what you want the harness to be able to do, the system asks clarifying questions, builds the plugin, and installs it automatically so it shows up alongside the harness’s default components. This is the mechanism that makes the harness “self-extensible”: the agent can reason about its own plugin architecture well enough to extend itself. A demonstrated example involved asking creator mode to build a plugin that fetches GitHub star counts for any repository. The result was a working plugin installed and ready to use in the same session, with no separate build step outside the conversation. ## What is the trajectory view and why does it matter? The trajectory view is an audit trail for every action the agent takes during a session. For each step, whether it’s a context injection, a tool call, or a clarifying question, you can click into it and see exactly which plugin produced it. If the agent injects a system prompt, the trajectory view shows it came from a specific system prompt plugin. If it delegates work to a sub-agent, the trajectory view shows that delegation as a distinct traceable event. This resembles what AI observability platforms like LangFuse or LangSmith provide for LLM applications generally, but built directly into the coding agent’s own interface. It gives a level of granularity into an agent’s inner workings that isn’t available in Claude Code, Codex, or even Pi, which is otherwise the closest comparable tool. ## How does sub-agent delegation work? ### Everyone else built a construction worker. We built the contractor. One file at a time. UI, API, database, deploy. DeepSeek Harness can call other coding agents, including Claude Code and Codex, as sub-agents from within a session. Rather than replacing those tools, the harness treats them as callable resources: you tell the agent you want to delegate a task to Claude Code or Codex, it makes a tool call to hand off the work, reasons over what comes back, and continues the conversation. This delegation is itself implemented as a plugin (a “subagent Claude Code” plugin, a “subagent Codex” plugin), which fits the overall pattern: any capability, including the ability to reach into another vendor’s coding agent, is just another composable piece of the system. It lets you use DeepSeek Harness as an orchestration layer while still tapping into other models’ strengths for specific tasks. ## Is DeepSeek Harness worth using instead of Claude Code or Codex? The honest answer, based on the current state of the project, is: not as a wholesale replacement yet, but worth understanding closely. DeepSeek Harness is described as rough around the edges, and it demands real time investment to configure plugins, tune the agent loop, and build out the extensions you need. Out-of-the-box tools like Claude Code and Codex remain more polished for people who just want something that works immediately without assembly. But the trade-off is control. With Claude Code or Codex, there’s no way to fix the specific behaviors that annoy you, like an overly verbose model, glitchy sub-agent handling, or an agent loop that wanders off track. You can nudge those tools with rules and prompts, but you can’t actually rebuild the parts that don’t work. DeepSeek Harness allows exactly that: you can modify, disable, or replace any plugin, including ones that govern the harness’s own internal loop, and adapt it to a specific codebase or workflow over time. The framing worth taking seriously is that this kind of architecture, not necessarily this specific project, looks like a plausible direction for coding agents generally. As agents get more capable of extending and modifying software autonomously, it follows that they’ll increasingly be used to extend and modify their own harnesses too. DeepSeek’s plugin system, and its creator mode in particular, is a concrete example of what a self-extensible agent looks like when the self-extension is built in as a first-class feature rather than bolted on. ## Frequently Asked Questions ### What makes DeepSeek Harness different from Claude Code or Codex? Claude Code and Codex are closed systems with fixed models and no access to their internal agent logic. DeepSeek Harness exposes everything, from the UI to the agent loop, as editable, toggleable plugins, and it isn’t locked to any single model provider. ### Do I need to know how to code to build a plugin for DeepSeek Harness? Not necessarily. Creator mode lets you describe the functionality you want in plain language, and it asks clarifying questions, builds the plugin, and installs it automatically within the same session. ### Can DeepSeek Harness use models other than DeepSeek’s own? Yes. While DeepSeek V3.1 is the default model, connected through a DeepSeek or OpenRouter API key, the harness supports any provider you configure in the models tab. ### Can DeepSeek Harness work alongside Claude Code or Codex rather than replacing them? ## Remy doesn't write the code. It manages the agents who do. Remy runs the project. The specialists do the work. You work with the PM, not the implementers. Yes. It supports sub-agent delegation, so you can call Claude Code or Codex from within a DeepSeek Harness session for specific tasks, then continue working in the harness once results come back. ### Is DeepSeek Harness ready to replace my current coding agent? It depends on your tolerance for rough edges. The project is functional and rapidly gaining attention, but it requires more setup and tuning than a polished out-of-the-box tool. It’s better understood right now as an early look at where customizable, self-extensible coding agents are heading.