DeepSeek opens its AI agent tool to everyone

Chinese AI company DeepSeek has released DeepSeek Harness, abbreviated dsh, as open source under the MIT license. An agent harness is the software that surrounds an AI model and lets it carry out multistep tasks, such as working with files or running commands.

The project is still in developer preview, though. The README warns that upcoming changes may break compatibility and urges users to read SAFETY.md before running it.

Everything is a plugin

The architecture follows the principle that everything is a plugin and is built on the Cordis framework. According to an independent review from Developers Digest, the model adapter, tools, storage, the agent loop itself and the web interface are all plugins. The same review says DeepSeek uses a forked version of Cordis with 18 local patches.

According to an introductory guide from CodePick, the harness is model-agnostic and supports providers such as DeepSeek, OpenAI and Anthropic. It can run file operations, shell commands, network calls and MCP servers. Plugin discovery is also available via the GitHub topic dsh-plugin.

Security and audit trails

According to Developers Digest, every request to the language model checks that the outgoing messages match an append-only session log byte for byte. This makes audit trails part of the run itself.

The sandbox, meaning the isolation layer that limits what the agent can touch, is "fail-closed," according to the review. It uses bubblewrap and Landlock on Linux, seatbelt on macOS and token restriction on Windows, and it refuses to run if the protection is unavailable.

Getting started

The harness can be launched locally as a web interface with the command npx @deepseek-ai/dsh web, which opens at http://127.0.0.1:3080 by default. There is also a desktop application, and it can be built from source with pnpm.

The adapter uses the models deepseek-v4-flash and deepseek-v4-pro by default, according to Developers Digest with a 1 million token context window and three reasoning levels: off, high and max.

Limitations

There are no published benchmarks, as the file BENCHMARK.md is only a three-line stub. All development arrived as a single PR, so the history is limited. According to Developers Digest, the codebase consists of roughly 453,000 lines of TypeScript spread across about 219 packages. The figures come from that one review and have not been confirmed by DeepSeek.

CodePick concludes that the tool is not suitable for production environments that require stability, and that the README does not mention any Python SDK.