DeepSeek Harness: A Hands-On Guide to the Everything-is-a-Plugin Agent Framework
Quick verdict
DeepSeek Harness (dsh) v0.1 is DeepSeek's open-source agent framework released on August 13, 2026. Every component is a plugin, you can start it with `npx @deepseek-ai/dsh web`, and the Web UI serves at http://127.0.0.1:3080.
DeepSeek Harness — codenamed dsh — is DeepSeek's open-source agent framework, released as a developer preview (v0.1) on August 13, 2026 under the MIT license. Its core idea is simple and radical: everything is a plugin.
It is also the harness DeepSeek used for the V4 Pro 0813 agent benchmarks, which means the evaluation setup behind scores like Terminal Bench 2.1 and DeepSWE is now open for anyone to inspect and reuse.
What Is DeepSeek Harness?
dsh is an agent harness built by DeepSeek AI. It provides the runtime, tools, model adapter, session log, and agent loop for building autonomous agents — and every one of those pieces is a replaceable plugin.
The framework is powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability. There is no privileged core to patch: you extend dsh by mounting a plugin next to the others, and registrations unwind cleanly when a plugin unloads.
Why It Matters
- Open evaluation: DeepSeek used Harness (minimal mode) for the V4 Pro 0813 agent benchmarks. With the public release, developers can reproduce those numbers and run comparisons under the same conditions.
- Everything is a plugin: the model adapter, tool registry, session log, and even the agent loop itself are replaceable from configuration.
- Rapid ecosystem: the community has already tagged hundreds of repositories with the
dsh-plugintopic, adding things like long-term memory and sandboxed execution.
Quick Start
Prerequisites
- Node.js (official docs recommend recent versions, e.g. 22.19+ or 24+)
- A DeepSeek API key (or another OpenAI-compatible endpoint) to run real agents
Install and run from npm
The fastest way to start is via npx:
npx @deepseek-ai/dsh web
The command boots the Web UI, served at http://127.0.0.1:3080 by default.
⚠️ DeepSeek Harness is a developer preview. The project is iterating rapidly and will introduce breaking changes.
Run from source
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web
Web UI Walkthrough
Once the server is running:
-
Open the Web UI at
http://127.0.0.1:3080. -
Configure a model: go to Settings → Models, enter your DeepSeek API key, and save. The model route becomes usable immediately — no restart needed.
-
Choose a workspace: click Choose workspace, add the project directory where you started
dsh, and select it. The session composer stays disabled until a workspace is selected. -
Run a task: start a session and send something like:
Summarize this repository and identify its main packages.
The agent can read and edit workspace files, run commands, delegate work, and maintain a plan. For operations that need approval, the Web UI asks first based on the active permission policy.
The dsh process uses its invoking directory as the default filesystem location, but a fresh Web UI has no workspace until you add one.
Key Concepts
Everything is a plugin
Model adapters, tool registry, session logging, system-prompt assembly, telemetry, and the agent loop are all contributed as plugins. You extend dsh by mounting a plugin beside the others — there is no core fork needed.
Profiles and bundles
A profile is a named composition of bundles. The web and headless profiles ship as templates:
web— the browser applicationheadless— a one-shot runner with no server
dsh-base provides the first layer of every profile (model adapters, tools, persistence, sandbox and approval policy, settings, credentials, telemetry).
To inspect the tree your machine actually boots:
dsh --profile web --dump-config
Any row it prints can be replaced with your own patch.
The plugin ecosystem
Community plugins are discoverable via the dsh-plugin topic on GitHub. Early community additions include long-term memory and sandboxed/remote execution plugins. This is the fastest-growing part of the project right now.
Headless / CLI Mode
Beyond the Web UI, dsh ships a headless mode — a one-shot runner with no server. This is useful for scripting agent runs, CI evaluation, and reproducing DeepSeek's benchmark setup in batch.
Developer Notes
- Language / runtime: TypeScript monorepo; requires recent Node.js and pnpm.
- Breaking changes expected: do not pin a critical production workflow to
v0.1without a strategy for updates. - Community: feedback and bug reports go through GitHub Discussions; there is also an official Discord community.
Should You Use It?
Yes, if you want to build or evaluate DeepSeek-powered agents, reproduce DeepSeek's own benchmarks, or want a fully plugin-driven agent stack.
Wait, if you need a stable, production-hardened agent framework today — this is a fast-moving developer preview.
For a ready-to-use API experience meanwhile, DeepSeek V4 Flash remains the most proven budget option, and DeepSeek V4 Pro is the stronger agentic model.
Related Articles
Related Articles
Qwen 3.8-Max: Alibaba's 2.4 Trillion Parameter Open-Source Flagship (2026)
Qwen 3.8-Max is here — 2.4T parameters, 1M context, open weights coming next week. Full review covering autonomous coding, agent benchmarks, pricing, and how it compares to GPT-5.6 and Fable 5.
DeepSeek V4 Flash: Official Release With Major Agent Performance Boost (2026)
DeepSeek V4 Flash official release review. 284B MoE model, 1M context, 384K output, and dramatically improved agent benchmarks. Full review with pricing, benchmarks, and use cases.
DeepSeek: Complete Guide to the R1 & V4 AI Models (2026)
An in-depth review of DeepSeek—covering R1 reasoning, V4 Flash and V4 Pro (official release 0813), pricing, and real-world performance. Is DeepSeek still the best value AI in 2026?
Qwen3.8-27B: The 27B Open-Source Model That Packs Agentic Coding Into Your GPU
Qwen3.8-27B is a ~27B dense, Apache-2.0 multimodal model with strong agentic coding (DeepSWE 42.2, Terminal-Bench 73.0) and a ~17GB GGUF, made for consumer GPUs.