DeepSeek just dropped an MIT-licensed agent framework that powers their own benchmark numbers. It’s free, runs locally, and you can try it in two minutes with zero setup. Here’s what you need to know.
If you’ve been watching the AI coding space, you know the pattern. Claude Code costs $20/month. Codex costs $20/month. They’re powerful, but you’re renting someone else’s infrastructure. DeepSeek Harness changes that equation.
What Is DeepSeek Harness? (Plain English)
The one-sentence summary
DeepSeek Harness is a complete AI agent system — web interface, command line, Python SDK — that you download, own, and run on your own machine. Every part is swappable. The model, the tools, the memory, the UI. All plugins.
Why “everything is a plugin” matters to you
DeepSeek’s exact words: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all swappable from configuration.
Translation: Don’t like the default model? Swap in GPT-4o, Claude, or a local Llama. Want a different file editor? Replace the plugin. Need custom orchestration logic? The agent loop itself is a plugin.
This isn’t marketing fluff. It’s built on Cordis, a battle-tested plugin kernel. DeepSeek assembled a complete agent product on top, then made every layer replaceable. You start from a working agent and swap parts. Most frameworks make you build from scratch.
The 5-Minute Test Drive (No Install Required)
You don’t need Git. You don’t need to build anything. You need Node 22+ and one command:
npx @deepseek-ai/dsh web
That’s it. Opens at http://127.0.0.1:3080. You get a full web UI — chat interface, file tree, session log, plugin inspector. Click around. Run a task. Watch the agent think.
If you want to modify the code later, clone the repo and run pnpm install && pnpm run build && pnpm dsh web. The npx route is for kicking tires. The clone route is for building on top.
Pro tip: The default branch is master, not main. Raw GitHub links against /main/ will 404.
Four Ways to Use It (Pick Your Mode)
Harness ships four presets. They’re not marketing tiers — they’re literally different folders on disk with different tool sets.
| Mode | Tools | Best For |
|---|---|---|
| Standard | Full suite: file edit, shell, search, skills, planning, goals, subagents, workflows | Daily driver coding agent |
| Code | Standard capabilities via Code Mode SDK; model writes TypeScript programs instead of discrete tool calls | Developers who want programmatic control |
| Minimal | Two tools only: persistent bash + str_replace_editor | Benchmarking, evaluation, understanding core behavior |
| Creator | Standard + runtime inspection, in-memory plugin experiments, preset authoring | Building custom modes, plugin development |
The Minimal mode secret
Here’s why Minimal mode matters: it’s the exact harness DeepSeek uses for its own published benchmark numbers. V4-Pro, V4-Flash — the code-agent scores in those model cards come from this two-tool configuration.
The repo’s BENCHMARK.md documents the eval path. If you evaluate models, you can run the same harness DeepSeek uses. That’s unprecedented — vendors usually keep their eval infrastructure private.
How It Compares to Claude Code & Codex
VentureBeat called it “an open-source, model-agnostic alternative to the agent infrastructure underlying Claude Code and Codex.” That’s the right framing. It’s not a full product replacement yet.
What’s better
| Harness | Claude Code / Codex |
|---|---|
| MIT license — you own it | Proprietary, rented |
| Any model via adapter | Locked to vendor models |
| Every layer inspectable | Black box |
| Free forever | $20/month |
| Local-first, no cloud required | Cloud-dependent |
What’s missing (honest gaps)
- No hosted background-agent service (you run it)
- No GitHub-native PR workflow (yet)
- No IDE extensions, no mobile, no Slack integration
- Version is 0.1.0-rc.5 — release candidate, zero tags, breaking changes promised
- Plugin interfaces will shift
The README says it in caps: “THERE WILL BE COMPATIBILITY-BREAKING CHANGES.” This is developer preview. Treat integrations as provisional.
Should You Use It Now? (Decision Guide)
Try it now if:
- You evaluate models — Minimal mode is the vendor’s own benchmark harness. Run it. Read it. Calibrate your own evals against it.
- You tinker with frameworks — The plugin kernel, append-only session log, and Creator mode are genuinely interesting architecture. Build experiments, not dependencies.
- You want leverage — Even if you never run Harness in production, an open, model-agnostic harness existing changes your negotiating position with closed vendors. That has value.
Wait if:
- You run agents in production — 0.1.0-rc.5, zero releases, zero tags, no hosted service. Watch for a first tagged release and interface stabilization.
- You need Claude Code/Codex parity today — Harness doesn’t match surface breadth or hosted workflows. The reason to care is strategic, not tactical.
The Bottom Line
DeepSeek published the gym they grade their own models in. That’s the story.
Strip the launch noise and three facts remain: real MIT-licensed harness, every layer a plugin, Minimal mode powers their benchmark numbers. The version label is softer than the coverage implies (rc.5, not v0.1). The install recipe circulating on social media misses a build step. The plugin noun list is DeepSeek’s, not the paraphrase.
But the hit rate is better than most launch-day hype. And the corrections are the part you only get from reading primary sources.
What happens next is the test: will independent teams rerun DeepSeek’s methodology against open weights? Will a dsh-plugin ecosystem form? An open harness makes both possible.
Possibility is where the vendor’s job ends and the community’s begins.
Your move: Run npx @deepseek-ai/dsh web. Spend 10 minutes in the UI. Decide if the architecture clicks for you. That’s the only evaluation that matters.
