DeepSeek Harness
Open SourceEverything is a plugin.
Published 27 September 2026 · Last updated 6 October 2026
Scores
Popularity4/5
Exceptional launch trajectory (95K+ stars in 48 hours, 130K+ within days) is a strong early signal, though it's still a very recently launched project without a sustained multi-month track record.
Learning Curve3/5
The web UI and CLI are approachable quickly, but understanding and customizing the plugin architecture (swapping loops, sandboxes, or the UI layer) takes real familiarity with the Cordis framework underneath.
Flexibility5/5
The plugin-everything design is the most architecturally open-ended approach in the subcategory. Model, tools, sandboxes, orchestration, and UI are all independently swappable rather than baked into a fixed loop.
Performance3/5
Too new for independent benchmarks at time of evaluation. Backed by a well-resourced model lab and DeepSeek's own V4-Pro model, but explicitly a developer preview with known compatibility instability ahead.
Portability5/5
MIT-licensed, model-agnostic across major providers, and built on a swappable plugin kernel with no hard vendor lock-in required to use the core harness.
About DeepSeek Harness
DeepSeek Harness (command dsh) is DeepSeek's open-source, MIT-licensed agent harness, launched in August 2026 as a developer preview and widely compared with Claude Code. Its organising idea is everything is a plugin: the model adapter, tool registry, session log, agent loop, sandboxes, filesystem access, orchestration, and the interface itself are swappable components composed on Cordis, an open-source plugin framework from the Cordiverse project, rather than a fixed agent loop.
The default surface is a local web UI started with npx @deepseek-ai/dsh web, and desktop apps for Apple Silicon Macs and 64-bit Windows wrap the same interface. The same dsh launcher also runs headless one-shot jobs, serves programmatic clients over the Agent Client Protocol, and backs a Python SDK installed from PyPI, so one configuration covers interactive work, scripts, and embedding. The agent reads and edits workspace files, runs commands, keeps a plan, delegates subtasks, and asks for approval under a configurable permission policy.
DeepSeek's API is the first-class model route, set up with a single API key. Third-party providers such as Anthropic, OpenAI, Kimi, and GLM come from a bundled provider catalog, and any relay, company gateway, or self-hosted server can be added as a custom endpoint speaking the OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages protocol. The harness itself is free; the cost is the inference billed by whichever provider is configured.
It is still a developer preview: the README warns of compatibility-breaking changes, and plugin APIs and profile configuration shift between releases. Community plugins are collected under the dsh-plugin GitHub topic.
Key Features
- Everything-is-a-plugin design: model adapter, tools, sessions, sandboxes, filesystem, orchestration, and UI are all swappable
- Profiles built as layered Cordis plugin bundles, with optional hot reload of configuration
- Local web UI, plus desktop apps for Apple Silicon Macs and 64-bit Windows
- Headless one-shot mode and an Agent Client Protocol server for programmatic clients
- Python SDK (deepseek-harness-sdk on PyPI) that drives the same profiles without a system Node.js
- DeepSeek API first-class, with Anthropic, OpenAI, Kimi, GLM, and more from the bundled provider catalog
- Custom endpoints over the OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages protocol
- MIT licensed, with community plugins discoverable through the dsh-plugin GitHub topic
Pros
- Deep customisation is first-class: swap the agent loop, the sandbox, or the UI without forking
- One launcher covers web, desktop, headless, ACP, and SDK use from the same configuration
- Not locked to DeepSeek: third-party providers and self-hosted endpoints work alongside it
- Fully open source (MIT) and maintained by a major model lab
- Large community on GitHub, with a plugin ecosystem already forming
Cons
- Developer preview, with compatibility-breaking changes announced in the README
- The plugin-everything design adds concepts (profiles, bundles, patch layers) a fixed-loop agent doesn't need
- No terminal UI ships by default; interactive work goes through the web UI or the desktop app
- Desktop builds cover only Apple Silicon Macs and 64-bit Windows, so Linux users run the web UI
DeepSeek Harness Pricing
Open Source- · Free, MIT-licensed harness — bring your own API key to DeepSeek or any supported provider
- · Usage-based inference cost depends on the provider and model chosen
Tech Stacks with DeepSeek Harness
DeepSeek Harness Dev Stack
DeveloperDeepSeek Harness as the coding agent, in its local web interface or desktop app, running DeepSeek or models such as Kimi, GLM, Claude, and GPT, with GitHub or GitLab for version control.
Development
LLM
Tools Related to DeepSeek Harness
Integrates with DeepSeek Harness(1)
DeepSeek Harness ships a direct DeepSeek adapter: one DeepSeek platform API key in its Models settings makes DeepSeek's models the default route, with reasoning controls and DeepSeek's API wire extensions supported.
Alternatives to DeepSeek Harness(11)
Both are hackable, open-source harnesses, and DeepSeek Harness routes third-party models through Pi's own pi-ai library. DeepSeek Harness makes every component a Cordis plugin behind a web UI; Pi is a minimal terminal agent with four core tools extended in TypeScript. The harness ships more, Pi stays smaller.
DeepSeek Harness is the open-source rival most often compared with Claude Code, built from swappable plugins and provider-agnostic; Claude Code is Anthropic's proprietary agent for the terminal and IDEs, tuned for Claude models. The harness is fully rebuildable, Claude Code more polished.
DeepSeek Harness is a plugin-built agent run from its own web UI or desktop app; Kilo Code spans VS Code, JetBrains, a CLI, and cloud agents with 500+ models at zero markup. The harness wins on rebuilding the agent itself, Kilo Code on editor reach.
Both are open-source agents from model labs: DeepSeek Harness is a web and desktop agent assembled from swappable Cordis plugins, Grok Build a terminal agent that fans work out to parallel subagents, optionally in separate git worktrees. Choose the harness for customisation, Grok Build for parallel terminal work.
DeepSeek Harness is an open-source, plugin-built agent with a local web UI and desktop app that runs DeepSeek or other providers; ZCode is Z.ai's open-source desktop app tuned for GLM, with Goal Mode for long-running tasks. Pick the harness to customise the agent itself.
Both are open-source agents from model labs: DeepSeek Harness is a plugin-built web and desktop agent with DeepSeek first-class, Qwen Code a terminal-first agent tuned for Qwen models that also runs in VS Code, Zed, and JetBrains. The harness goes deeper on plugins, Qwen Code reaches more editors.
Vendor
DeepSeek
Website →Tags
Details
- Maintained
- Yes
- Primary languages
- JavaScript, TypeScript, Python, Go, Rust
- Type
- AI IDE
- Open source
- Yes
- GitHub stars
- 234k
- AI features
- Chat, Agent, Inline edit
- Stars updated
- 2026-09-23