DeepSeek Harness (DSH) — an open-source agent framework from DeepSeek, released in August 2026 — has received an official public registry of plugins and Agent profiles, dsh-index. The registry is hosted on xlings.org and includes 71 packages with SHA256 verification and CN mirrors. The ecosystem has grown to approximately 1,183 projects on GitHub, and the main repository has gathered 64,500 stars in a few weeks.
What happened
An official dsh-index registry appeared on xlings.org, containing 71 packages: 68 plugins, 2 ready-made Agent profiles, and 1 group package in 78 versions. 41 packages are mirror-published via the CN-mirror xlings for access from China. DSH installation is done with a single command npx @deepseek-ai/dsh web, after which plugins for TUI, desktop applications, vision analysis, and MCP integrations become available.
Context
DSH is built on Node.js 22.19+/24+ with the pnpm package manager and uses an "everything is a plugin" architecture: each agent component is an npm package installed via a standard package manager. The framework works on top of any LLM via API and does not contain its own model. The approach is similar to how the npm ecosystem accelerated the spread of Node.js — but applied to an agent framework. DeepWiki has prepared detailed documentation on the project's architecture.
Why this matters for the industry
DeepSeek officially establishes the npm plugin ecosystem as the standard extension for the agent framework — this is a platform shift, not just a feature. A centralized registry with SHA256 verification solves the trust problem for open-source agent plugins. Rapid growth — 1,183 projects under the dsh-plugin tag and 64,500 stars on GitHub in a few weeks — signals a shift of developers from proprietary agent platforms like Claude Code to open infrastructure where the model is a replaceable component. Without public benchmarks (SWE-bench, AgentBench), the exact performance of DSH has not yet been verified.
Why this matters for users
Developers can install DSH with a single command and get a working AI agent with the ability to choose from a catalog of verified plugins. The platform allows testing agent solutions with various models through a single interface without being tied to a specific AI company. For prototyping and internal tools, DSH is ready to use right now. For production deployment, it is recommended to wait for the appearance of latency benchmarks, data on the stability of plugins under load, and inference cost estimates.
What is still unknown / limitations
There are no public benchmarks for agent task performance. There is no data on the stability of the plugin system under production load and inference cost estimates. Comparing the growth rate of DSH with Llama 3.1 or langgraph is an incorrect analogy — Llama 3.1 is a new-generation model, and langgraph is a mature orchestration framework. The Node.js architecture limits performance for compute-intensive tasks.
Sources
- dsh-index — registry of plugins and Agent profiles
- DSH Base — framework documentation
- deepseek-ai/deepseek-harness — architecture overview on DeepWiki
Author
Look at AI, editorial team