DeepSeek Harness is MIT-licensed. Read the loop.

DeepSeek dropped Harness v0.1 into developer preview on August 13 and open-sourced the whole thing under MIT. Not a model. The runtime that wraps a model and turns it into an agent — the loop, the tool calls, the sandbox, the session state. That's the layer most teams rent without noticing, and it's the layer where your actual business rules live.

What actually happened

Per DeepSeek's own docs, Harness is built on a meta-framework called Cordis, which handles plugin mounting, unmounting, and dependencies. The design principle is blunt: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI are all plugins you can "select, swap, or extend in configuration without changing the DeepSeek Harness source code."

It ships with four runtime modes — Standard, Code, Minimal, and Creator — and installs with npx @deepseek-ai/dsh web. The code is at github.com/deepseek-ai/deepseek-harness under MIT, © 2026 DeepSeek. DeepSeek is explicit that this is a preview: core plugins and APIs will keep changing, and the repo warns of compatibility-breaking changes ahead.

Why the agent harness matters for your business

Here's what the harness actually decides for you. When does the agent stop? Which tools can it reach? What happens on a failed call — retry, escalate, or halt? Who approves a write? What lands in the audit log? None of that is the model. All of it is the harness, and all of it is your operating policy encoded in someone else's code.

Rent that layer from whoever sells you tokens and you've handed over your controls. The vendor changes the default loop, adds an auto-approve mode, or reworks tool scoping, and your governance changes in a release note you didn't read. We've watched exactly that play out — approval gates flipping to defaults is not a hypothetical.

MIT is the useful part here. You can read the seams: how Cordis separates the loop from the tools, how sandboxes get mounted, where session state is persisted. Even if you never ship a line of DeepSeek's code, that architecture is a free reference for building a harness you own — one where the model is a plugin and swapping providers doesn't touch your policy.

The obvious caveat: do not put a v0.1 developer preview in front of customers. DeepSeek says the APIs will break. Treat this as a teardown, not a dependency. Copy the shape, keep the boring parts of your stack boring, and put your approval gates somewhere a vendor upgrade can't move them.

Sources: DeepSeek Harness docs, deepseek-ai/deepseek-harness on GitHub.