DeepSeek Open-Sources Harness Agent Framework Under MIT License, Ships V4-Pro With Native Responses API Support

Overview

DeepSeek has released DeepSeek Harness, an open-source agent runtime framework, as a developer preview under the MIT license, according to the project’s GitHub repository. The release came alongside an API update to DeepSeek-V4-Pro that adds native support for the OpenAI Responses API format, as reported by TechNode.

What We Know

DeepSeek Harness — described in its own repository as “DeepSeek Harness: Everything is a Plugin” — is built on a design philosophy the project’s README describes as “an architecture where everything is a plugin,” powered by what the project calls the Cordis framework. The command-line tool is known as dsh, and developers can launch it with npx @deepseek-ai/dsh web, which starts a local web interface, according to the README.

The repository’s package manifest lists the current version as 0.1.0-rc.5, and the project is licensed under MIT, according to GitHub. The repository quickly drew more than 85,000 stars and over 7,500 forks on GitHub in the day following its creation.

TechNode reported that DeepSeek is “positioning it as an open-source alternative to Anthropic’s Claude Cowork,” describing the tool as offering “a web interface for running AI agents” with a plugin-based system for expanding functionality. The outlet noted that the project “operates on DeepSeek’s Cordis architecture.”

The project remains explicitly unstable. Its README states that “DeepSeek Harness is currently in developer preview and is iterating rapidly,” followed by a bolded warning reading, in full capitals, “THERE WILL BE COMPATIBILITY-BREAKING CHANGES.” TechNode separately reported that DeepSeek’s preview notice cautioned the software “is still changing rapidly and may introduce compatibility-breaking changes.”

Separately, DeepSeek updated its V4-Pro model’s API on August 13. According to TechNode, the API documentation now identifies deepseek-v4-pro as an available model, with the current version designated “DeepSeek-V4-Pro-0813,” and the model supports the Responses API format as well as tool calls. TechNode reported the model’s context length at 1 million tokens with a maximum output of 384,000 tokens, and listed API pricing at $0.003625 per million input tokens on cache hits, $0.435 per million input tokens on cache misses, and $0.87 per million output tokens.

What We Don’t Know

No formal GitHub release tag or git tag has been published for DeepSeek Harness as of this writing — the 0.1.0-rc.5 designation exists only in the project’s package manifest, not as a tagged release. It is also not yet clear how DeepSeek Harness’s plugin architecture will evolve once it exits developer preview, or how directly it will end up competing with established coding-agent products given the software’s early, breaking-change-prone state.

Analysis

The pairing of an open-source, MIT-licensed agent runtime with a simultaneous model API update signals DeepSeek is pushing further into the layer that determines how models call tools, manage sessions, and execute multi-step agent workflows — territory increasingly contested among AI labs shipping their own coding-agent tooling. By open-sourcing the harness itself rather than keeping it proprietary, DeepSeek is inviting third-party developers to build plugins and extensions on top of its own models, a strategy that mirrors the plugin ecosystems competitors have cultivated around their own agent products.