DeepSeek open-sources Harness: an MIT-licensed Claude Code rival
Plugin-first, bring-your-own-model, and runs locally — pair it with the now-open V4-Pro for a fully inspectable agent stack. But it's a v0.1 preview.
Copy markdown
An MIT-licensed agent where everything is a plugin
DeepSeek open-sourced Harness under the MIT license — a terminal-plus-local-web coding agent in the Claude Code and Codex mold, but every piece (model, tools, memory, UI) is a swappable plugin instead of one fixed stack. Nothing is paywalled for local use.
Bring your own model — and pair it with open V4-Pro
Harness ships with no model of its own; you supply API keys or a local endpoint, so it runs on whatever you point it at. Combined with DeepSeek's now-MIT V4-Pro weights out the same week, you can stand up a fully open, self-hostable agentic-coding stack with no closed parts.
Running it tonight: Node, one npx, a local web UI
Install is a Node package — `npx @deepseek-ai/dsh` — that opens a local web interface alongside terminal control; you drop in model credentials and go. The plugin boundaries are the pitch: audit or replace the parts a closed agent won't let you touch.
The catch: it's a v0.1 developer preview
Harness is explicitly preview-grade — expect breaking changes, and its security boundaries, reliability and model quality are unproven with no production support yet. Ideal for tinkering and self-hosted experiments this weekend, not for wiring into a production pipeline.