What Is DeepSeek Harness? The Plug-In Coding Agent Explained
DeepSeek Harness is a plug-in based agentic coding framework that can even call Claude Code or Codex as sub-agents. Here's how it works.
What is DeepSeek Harness?
DeepSeek Harness is DeepSeek’s own agentic coding system, released alongside the V4 Pro model as a developer preview. Instead of shipping a single fixed workflow like most coding agents, it treats every part of the system, tools, skills, sessions, even other coding agents, as a plug-in you can swap in or out. That includes the ability to call Claude Code or Codex as sub-agents inside your own DeepSeek-orchestrated workflow, which makes it less a competitor to those tools and more a framework that can sit above them.
TL;DR
- DeepSeek Harness ships as a developer preview alongside the V4 Pro model, and it comes with a local web app interface rather than just a CLI.
- Its core idea is plug-in composability: tools, skills, sessions, and even entire agent harnesses like Claude Code or Codex can be mixed, matched, and swapped.
- The framework is built on a stated philosophy from DeepSeek’s own paper on “programming paradigm for spatial temporal composability,” which frames agent behavior as something you assemble rather than something baked in.
- In a real-world test building a real-time International Space Station tracker, the harness burned through roughly 20 million tokens across two turns and about 35 minutes, with a 100% cache hit rate by the end of the run.
- The interface exposes live stats most harnesses hide: tokens per second, cache hit rate, turn count, and running time, giving builders more visibility into cost and performance while a task runs.
- Early hands-on impressions favor the Flash model over Pro for coding tasks, citing better cost-to-output value even though Pro is a substantial capability jump.
- The project is MIT licensed, and its config lives in a YAML file, meaning you can add other model providers or point it at a locally hosted model instead of DeepSeek’s hosted API.
How is DeepSeek Harness different from Claude Code or Codex?
Claude Code and Codex are both opinionated, largely closed coding agents: you pick the model, you pick the workflow, and the harness around it is fixed by the vendor. DeepSeek Harness inverts that. Its selling point is that the harness itself is decomposable. Tools, skills, and even the underlying agent loop are plug-ins, and the project explicitly supports calling Claude Code or Codex as sub-agents from within a DeepSeek-orchestrated session. That means you could, in theory, build a multi-agent pipeline where DeepSeek’s harness manages orchestration while delegating specific subtasks to whichever agent (its own model, Codex, or Claude Code) is best suited for that step.
This composability is described in DeepSeek’s own paper on what it calls a programming paradigm for spatial temporal composability, essentially a framework for treating agent capabilities as interchangeable parts rather than a monolithic pipeline. Practically, this shows up in the settings UI, where plug-ins can be enabled or disabled, models can be swapped per task, and permissions can be scoped from read-only to full dangerous-skip-permissions access, similar to modes long available in Claude Code.
One interesting side note from early testers: the interface and interaction style of DeepSeek Harness reportedly feels similar to Codex, which may not be a coincidence. There have been claims that a meaningful share of commits and pull requests in DeepSeek’s coding tooling came out of a Codex work tree, which would explain some of the shared UI vibes.
What does the plug-in architecture actually let you do?
In practice, the plug-in system covers several categories:
- Tools: things like shell access, file editing, and web search are individually toggleable plug-ins rather than fixed capabilities.
- Modes: the harness ships with several operating modes, including a full coding agent mode with file editing, shell, and web search; a leaner code mode; a minimal mode; and a creator mode where you define your own custom agent profiles.
- Models: with a DeepSeek API key, both Flash and Pro versions of V4 are selectable inside a session, and the reasoning effort for each can be tuned.
- Providers: because the project is MIT licensed, users can add other model providers beyond DeepSeek’s own API, including pointing the harness at a self-hosted or local model.
- Sub-agents: entire external harnesses (Claude Code, Codex) can be wired in as plug-ins, letting you build a small multi-agent system using DeepSeek’s orchestration layer as the glue.
All of this configuration lives in a YAML file, so changes to providers, models, or reasoning effort are editable directly rather than buried in a settings menu you can’t inspect.
How do you set up DeepSeek Harness?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
The most reliable path, based on hands-on testing, is cloning the repository directly and installing from source rather than relying on a packaged release, since the tool is still a developer preview with rough edges. Once installed, it ships with a local web app you start with a simple command, which spins up a browser-based interface on a local port (by default, something like port 3018, though this is configurable).
On first login, the app asks for a DeepSeek API key, which you generate from a DeepSeek account’s API keys page. After that, setup involves:
- Selecting a workspace folder where the project’s files will live.
- Choosing an operating mode (full coding agent, code mode, minimal, or creator mode).
- Selecting a model (Flash or Pro, assuming a DeepSeek key is provided).
- Optionally configuring web search, which only activates if a DeepSeek key is present.
- Setting permission levels, ranging from read-only to full access.
Because it’s a local web app, there’s also the option to host it somewhere accessible over the internet, effectively giving you a personal, remotely reachable coding agent instance.
Is DeepSeek Harness worth using right now?
For builders comfortable with developer previews, yes, with caveats. The system delivered genuinely impressive output in testing: a real-time International Space Station tracker built against a live-updating API, complete with an accurately rendered Earth shader and correct sun positioning relative to real-world time. The agent worked through the task with a visible to-do list, debugged issues with Playwright-based verification, and ultimately produced a result described as one of the better outputs seen for that kind of prompt.
The tradeoff is token efficiency. The ISS tracker task burned through around 20 million tokens over two turns and roughly 35 minutes of running time, with output tokens around 240,000 at about 130 tokens per second. That’s a heavy token bill for what is, on paper, a fairly simple app. Cache hit rates were consistently strong, ranging from 95 to 100 percent across runs, which helps offset raw cost, but the overall token hunger is notable and consistent with a broader pattern in open models generally trailing closed frontier models on efficiency.
Between the two available models, Flash appears to offer better return on investment for coding tasks than Pro, despite Pro representing a bigger capability jump over its predecessor. Given DeepSeek’s already competitive pricing on both models, Flash is the more practical starting point for most coding workflows, with Pro reserved for harder problems where the extra reasoning effort pays off.
Frequently Asked Questions
What is DeepSeek Harness built on?
It’s DeepSeek’s own agentic coding framework, released as a developer preview alongside the V4 Pro model, and built around a plug-in architecture where tools, skills, sessions, and even sub-agents are interchangeable components rather than fixed parts of the system.
Can DeepSeek Harness use Claude Code or Codex?
Yes. One of its core design goals is letting other coding agents, including Claude Code and Codex, be called as plug-in sub-agents within a DeepSeek-orchestrated multi-agent workflow.
Is DeepSeek Harness free or open source?
The project is MIT licensed, which means the code is open and modifiable, including adding new model providers or pointing it at locally hosted models. Usage costs depend on which model API you connect, such as DeepSeek’s hosted Flash or Pro models.
How do I install DeepSeek Harness?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The recommended approach is cloning the repository and installing directly from source, since it’s an early developer preview. After installation, a command starts a local web app interface where you configure your API key, workspace, mode, and model.
Should I use Flash or Pro with DeepSeek Harness?
Early testing suggests Flash offers better value for typical coding tasks, being faster and cheaper while still producing strong results. Pro is a meaningful upgrade in raw capability but comes at a higher cost, making it better suited to harder problems rather than everyday coding work.
