Ecosystem
Cordis, the composability paper, the dsh-plugin topic, benchmark interpretation and a neutral comparison with other agent harnesses
DeepSeek Harness is part of a larger ecosystem. This page lists the projects most directly related to it, and places them alongside other agent harnesses to help you evaluate and choose.
Cordis
Cordis is the plugin framework underneath DeepSeek Harness, maintained by the cordiverse organization. It supplies the runtime model for composing plugins, services, typed events and lifecycle (fibers); "everything is a plugin" is its core idea. The harness vendors the Cordis source.
Cordis has three core primitives: plugins (the smallest unit implementing a service), services (capabilities mounted on ctx and looked up by key), and events (communication between plugins). Together they underpin the harness's composability. If you want to write plugins, start with the Cordis primer.
The composability paper
A Programming Paradigm for Spatiotemporal Composability describes the design behind Cordis. "Spatiotemporal composability" means composing capabilities across space — how many plugins coexist and cooperate — and across time — how lifecycles start and unwind. Read it to understand why the harness takes a plugin-based architecture.
Preprint status
The paper is currently marked by its authors as a preprint under active revision and has not been peer-reviewed. Treat it as design reference, not as validated conclusions.
Interpreting benchmark numbers
Before reading any score attributed to DeepSeek Harness, get the framing right:
A system score, not a harness score
DeepSeek's official announcements measure public Code Agent benchmarks on the whole system of "model + DeepSeek Harness minimal mode + parameter configuration + tool environment". The official wording: "For Code Agent tasks in public benchmark suites, DeepSeek-V4-Pro-0813 is evaluated with the DeepSeek Harness minimal mode as the framework (max effort, top_p=0.95, temperature=1.0); results may differ slightly under other frameworks." (see the DeepSeek API announcement)
So the correct citation is "a system score of V4 Pro + Harness minimal mode", not "a DeepSeek Harness score". As of this writing, there is no independent controlled study comparing DeepSeek Harness against other harnesses under the same model, tasks and budget; the harness's own contribution must be measured with your own A/B tests instead of being inferred from official combined scores.
Plugin discovery: the dsh-plugin topic
Community plugins are discovered through the GitHub dsh-plugin topic. Add the topic to your plugin repository so others can find it on the topic page, and browse that page for existing plugins. Installation is covered in the FAQ.
The upstream repository
deepseek-ai/deepseek-harness is the upstream repository for this project, containing all source, documentation and issue tracking. Beyond the core framework, the repository also holds:
- Python SDK and TypeScript SDK: for programmatic use.
- Example bundles and runnable examples: under
examples/. - CLI: the
dshcommand that boots profiles.
This help center is adapted and localized from that repository; when content drifts from the latest version, the upstream takes precedence.
Beware of same-name projects
The name "deepseek-harness" was used by third-party projects before the official release, creating a risk of confusion and supply-chain impersonation. When installing or forking, verify the deepseek-ai organization and the @deepseek-ai npm scope. See security & compliance.
Comparison with other agent harnesses
An agent harness is the runtime frame that hosts agents: it wires model adapters, tools, filesystem, command execution, sandbox and approvals together. Several comparable tools exist, and the table below is factual and makes no quality judgment:
| Project | Description | License | Deployment | Position |
|---|---|---|---|---|
| DeepSeek Harness | An everything-is-a-plugin agent harness powered by Cordis | MIT | Local web / headless | An open, composable agent runtime |
| Codex CLI | OpenAI's open-source coding agent, running in the terminal | Apache-2.0 | Local CLI + cloud model | A mature terminal coding agent |
| Claude Code | Anthropic's coding tool, running in the terminal | Not released under an open-source license | Local CLI + cloud model | A mature terminal coding agent |
| Gemini CLI | Google's open-source CLI with MCP and extensions | Apache-2.0 | Local CLI + Google cloud | A terminal agent for the Google ecosystem |
| OpenCode | A model-neutral open-source coding agent | MIT | Local CLI / desktop + your provider | A model-neutral open runtime |
Codex CLI and Claude Code are terminal tools oriented at code repositories; DeepSeek Harness likewise lets an agent read, write and run commands in a local workspace. Its design difference is the composition model: every capability is a plugin that can be composed and swapped through Cordis, which emphasizes customizability and extensibility. Which one to pick depends on whether you prioritize ecosystem, model binding or composition freedom.
License verification
The licenses above reflect each repository at the time of writing (Apache-2.0 / MIT are verifiable on the corresponding GitHub repositories). A license only covers the software code itself — not model weights, third-party plugins or API service terms. Enterprises should produce their own SBOM and dependency license inventory before adoption, see security & compliance.
Keep exploring
- For the core concepts, see the glossary.
- For data paths, deployment risk tiers and security baselines, see security & compliance.
- To join discussions or report issues, see community & support.