AI Tools

What is DeepSeek Harness and is it free?

Published 2026-08-16 · Researched and price-checked on official sources

DeepSeek Harness is DeepSeek's free, MIT-licensed agent framework in which every capability is a plugin, launched in developer preview on August 13, 2026. It hit 115,907 GitHub stars within three days and competes with paid tools like Claude Code and OpenAI Codex.

Key facts

What matters

DeepSeek Harness is free and MIT-licensed; install it with npx @deepseek-ai/dsh web and the Web UI opens at http://127.0.0.1:3080.
It is a developer preview: the official README warns that compatibility-breaking changes are coming.
Every capability is a plugin (models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI), built on the Cordis framework.
It reached 115,907 GitHub stars and 11,319 forks within three days of the August 13, 2026 launch.
You bring your own model key: DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, or any OpenAI-compatible endpoint.

What is DeepSeek Harness?

DeepSeek Harness (CLI name dsh) is an open-source agent harness developed by DeepSeek AI. Every capability is a plugin: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI can all be swapped or recomposed. It is built on Cordis, a plugin framework whose design is described in the paper "A Programming Paradigm for Spatiotemporal Composability."

The project launched in developer preview on August 13, 2026, and the official README states that "there will be compatibility-breaking changes" as it iterates. The source is MIT-licensed TypeScript, and the repository passed 115,907 GitHub stars and 11,319 forks within three days. The Hacker News launch thread drew 730 points and 306 comments, making this the most-engaged AI tool story of the week.

Who is DeepSeek Harness for?

Developers and teams building AI agents who want full source-level control of the harness (MIT license).
Teams currently paying for Claude Code or OpenAI Codex that want to move model spend onto their own API keys.
Anyone who needs multiple model providers in one harness: DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, or any OpenAI-compatible endpoint.
Plugin authors: the project asks developers to tag plugin repositories with the dsh-plugin topic for discoverability.

Key features verified in the docs

Web UI: npx @deepseek-ai/dsh web starts a browser interface at http://127.0.0.1:3080 by default (Node.js required).
Python SDK: pip install deepseek-harness-sdk runs the agent from Python 3.10+ on Linux x64/arm64 or macOS 14+ on arm64.
Headless mode: a one-shot runner with no server, for CI or scripted jobs.
Credential safety: API keys are write-only and stored in $DSH_HOME/.credentials.yaml.
Approval policy: the Web UI asks before operations that require approval under the active permission policy, backed by a sandbox layer.
Session logs: sessions write JSONL logs containing model requests and tool calls.

DeepSeek Harness vs Claude Code vs OpenAI Codex

Claude Code is Anthropic's AI coding assistant that edits files, runs commands, and integrates with development tools, and most surfaces require a Claude subscription or an Anthropic Console account. OpenAI Codex is OpenAI's coding agent, accessed through ChatGPT and the API. Both are closed-source, released products.

DeepSeek Harness differs in two ways: it is MIT-licensed, so you can fork and modify the harness itself, and it is provider-neutral by design. The trade-off is maturity: Claude Code and Codex ship stable releases, while Harness is a developer preview with breaking changes ahead.

Getting started with DeepSeek Harness

Two install paths are documented. The quick path: install Node.js and run npx @deepseek-ai/dsh web, which starts the Web UI at http://127.0.0.1:3080. The dsh process uses its invoking directory by default, so you then choose a workspace and add a model API key under Settings, Models. The agent can read and edit workspace files, run commands, delegate work, and maintain a plan.

The programmatic path: clone the repository, create a Python virtual environment, pip install deepseek-harness-sdk, set DEEPSEEK_API_KEY (plus DEEPSEEK_BASE_URL if you use an OpenAI-compatible proxy), and run the bundled example against an isolated workspace and session directory.

Limitations and the honest catch

Developer preview: the README warns of compatibility-breaking changes; treat it as a moving target, not production infrastructure.
Windows support gap: the Python SDK docs list Linux x64, Linux arm64, and macOS 14+ on arm64 only.
Free software is not free to run: the Web UI needs a model API key, and DeepSeek API usage is billed per token under its peak and off-peak rates (see our DeepSeek API pricing coverage).
No issue tracker: the repo has GitHub issues disabled; feedback and bug reports go through GitHub Discussions.

How much does DeepSeek Harness cost?

The harness itself is free: the code is MIT-licensed, and the official page lists no subscription, cloud tier, or paid plan as of August 16, 2026. The cost is the model access you connect. The default provider is DeepSeek, and the Python SDK example defaults to the deepseek-v4-flash model, which DeepSeek bills per token: $0.66 per 1M output tokens off-peak and $1.32 at peak under the new peak/off-peak card that takes effect August 16, 2026.

Because you can configure Anthropic, OpenAI, Bedrock, Vertex, or Azure providers instead, your bill depends entirely on which model you point the harness at. Teams that want a flat subscription instead of usage-based billing should weigh that before switching.

Should you use DeepSeek Harness instead of Claude Code or Codex?

Choose it if you want source-level control of the harness, a provider-neutral setup, or a bill that tracks actual model usage rather than a flat subscription. The trade-off is maturity: Claude Code and Codex are released products with large ecosystems, while Harness is a developer preview that explicitly warns of breaking changes.

For a team evaluating agent infrastructure right now, the rational play is to prototype on the free MIT-licensed harness while keeping a released tool for anything production-critical. Revisit once DeepSeek Harness exits preview.

At a glance

	DeepSeek Harness	Claude Code	OpenAI Codex
License	MIT open source	Proprietary	Proprietary
Release stage	Developer preview (Aug 13, 2026)	Released	Released
Cost	Free software; you bring the model API key	Claude subscription or Console account	ChatGPT or API access
Model providers	DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, custom OpenAI-compatible	Claude models; third-party providers in Terminal CLI and VS Code	OpenAI models
Interface	Web UI, Python SDK, headless CLI	Terminal CLI, IDE, desktop, web	CLI, IDE, ChatGPT

DeepSeek Harness vs Claude Code vs OpenAI Codex (verified August 16, 2026)

FAQ

Is DeepSeek Harness free?

Yes. The harness is MIT-licensed open source with no paid plan listed on the official page as of August 16, 2026. You pay only for the model API you connect, such as DeepSeek API tokens.

How do I install DeepSeek Harness?

Install Node.js and run npx @deepseek-ai/dsh web to start the Web UI at http://127.0.0.1:3080. A Python SDK is also published as deepseek-harness-sdk for Python 3.10+ on Linux and macOS arm64.

What models work with DeepSeek Harness?

The default is DeepSeek, with built-in catalog providers for Anthropic and OpenAI. Bedrock, Vertex, Azure, and Codex use their native credentials, and any OpenAI-compatible endpoint can be added as a custom provider.