DeepSeek Harness: The Open-Source Coding Agent Built Like Minecraft

Three separate DeepSeek stories landed within hours of each other, and together they tell one story: DeepSeek is making a very public run at the agentic-coding crown that Anthropic's Claude Code currently wears.

What happened

Bloomberg reported that DeepSeek is openly positioning its coding-agent stack against Claude Code. The substance behind that report is DeepSeek Harness (DSH) v0.1, released as a developer preview and fully open-source under the MIT license. A hands-on all-night test by 36Kr's Silicon Star Pro came away impressed — their verdict was that DSH "outperforms Claude Code in a Minecraft-like way," which sounds odd until you see what they mean.

DSH is a web-based coding agent that runs on your own machine: session list on the left, conversation in the middle, every tool call visible on a timeline. It ships with four modes — a full-featured standard mode, a "PTC" mode where the model writes a single TypeScript program instead of ten back-and-forth tool calls, a stripped-down minimal mode for benchmarking, and a creation mode for building new modes. There's also a feature called Trajectory: an append-only log of everything the model saw, thought, and did, so you can replay and audit any run like a flight recorder.

The genuinely unusual part is the architecture. Everything is a plugin — and I mean everything, including the model itself, the tools, the sandbox, and even the default UI. The kernel is Cordis, a microkernel from the Koishi chatbot-framework community, whose signature feature is reversible side effects: plugins can be installed and uninstalled while the system runs, no restart, no leftover state. The 36Kr team installed and removed more than 20 plugins in a session without a single restart.

And the community went feral. In under 24 hours, 288 plugin repositories appeared under the dsh-plugin topic — dual-model routing to cut API bills, vision bridges that pair DeepSeek with any OpenAI-compatible vision model, Telegram and WeChat bot channels, and even plugins that import your Claude Code session history and continue it inside DSH. The Minecraft comparison is apt: Mojang never sold a pre-packaged game so much as a moddable foundation, and the community built the rest. DeepSeek seems to be betting the same play works for coding agents.

The third piece is pricing. Alongside the V4 lineup, DeepSeek introduced peak/off-peak API rates — off-peak is 50% cheaper — effective August 16, 16:00 UTC.

What it means for how you work

If your team runs coding agents or any agentic workload on a schedule, that off-peak pricing is not a footnote. Batch jobs, overnight refactors, report generation — shifting them to off-peak hours cuts the model bill in half before you've optimized anything else. And because DSH is open, local, and plugin-based, you're not locked into one vendor's harness: swap the model, swap the sandbox, swap the frontend, keep your workflows.

The audit angle is underrated too. Trajectory's full event log is the kind of thing compliance teams ask about and most agent products can't provide.

My take

My opinion on DeepSeek's coding harness is simple: let's wait and see how it does. Early hands-on enthusiasm is not production evidence, and a plugin ecosystem that's one day old is enthusiasm, not maturity.

But the part I'm confident about is the economics. Cache hit and token efficiency are the most important parts of a harness — and Claude Code, while being pretty feature-full, can be very inefficient much of the time. Z.ai's ZCode already does very well with cache hits, which shows how much room there is to win on efficiency rather than raw model quality. DeepSeek has highly talented engineers, so I expect their harness to be very good and to be the cheapest option for running DeepSeek V4 models on the DeepSeek API. If you're already planning to build on V4-Pro or V4-Flash, DSH is worth a serious look — just don't rip out anything that works on the strength of one all-night review.