DeepSeek AI Releases Open-Source Agent Harness with Plugin-First Architecture

DeepSeek Harness Enters Developer Preview

DeepSeek AI has published DeepSeek Harness (dsh), an open-source agent harness built on an "everything-is-a-plugin" architecture. The project is powered by Cordis, a framework whose design is described in the paper *A Programming Paradigm for Spatiotemporal Composability*.

The repository is currently in developer preview and the maintainers warn that compatibility-breaking changes will occur as the project iterates rapidly.

Key Features and Design

• Plugin-centric model: Every capability — tools, memory, planning, UI — is implemented as a plugin.
• Cordis runtime: Provides spatiotemporal composability, allowing plugins to be composed across time and space dimensions.
• Local Web UI: The `dsh web` command starts a server at `http://127.0.0.1:3080` and opens it in the default browser. An `--no-open` flag runs the server without launching the browser.
• SSH-aware launch: When started over SSH, only the host URL is printed because the SSH client or editor owns the forwarded address.
• Development workflow: `pnpm run dev:web` builds, serves, and hot-reloads client bundles in a single terminal. `make help` lists matching Make targets for Web and Desktop builds.

Safety and Community

• A safety notice is included in the repository and should be reviewed before running the project.
• Third-party dependencies and their licenses are disclosed in `THIRD_PARTY_NOTICES.md`.
• Feedback and bug reports are accepted through GitHub Discussions.
• Plugin authors are encouraged to add the `dsh-plugin` topic to their repositories for discoverability.
• A Discord community is available for real-time discussion.

Practical Implications

Developers interested in building modular, composable AI agents can experiment with DeepSeek Harness today, but should treat it as unstable pre-release software. The plugin architecture may appeal to teams wanting fine-grained control over agent capabilities without vendor lock-in.

How to Verify

• Source repository: <https://github.com/deepseek-ai/deepseek-harness>
• Documentation: <https://deepseek-harness.github.io/deepseek-harness/>
• Cordis design paper: *A Programming Paradigm for Spatiotemporal Composability* (linked from the repo)
• All claims above are drawn from the public GitHub repository and its documentation. No private access or insider information was used.