DeepSeek Harness (dsh) is DeepSeek’s open-source, MIT-licensed coding-agent harness built on an everything-is-a-plugin kernel where the model, tools, UI, and even the agent loop are hot-swappable plugins.
Facts below verified as of 2026-09-13.
The bet is not another agent but a harness with nothing built in: if the plugin architecture holds, forking a harness to change it becomes obsolete. Four weeks in, the bet is unproven and the project says so itself.
What it is #
A TypeScript CLI and web UI (npx @deepseek-ai/dsh web) that runs coding agents against local workspaces with file editing, shell, search, skills, planning, subagents, and workflows.
Four runtime modes set the tool surface: Standard, Code (the model orchestrates tool calls through generated TypeScript), Minimal (bash plus editor only, aimed at fair benchmarking), and Creator (authoring new plugins at runtime).
Everything sits on Cordis, a plugin kernel with hot reload and reversible side effects, described in a companion paper from Peking University and DeepSeek-AI.
Surfaces beyond the web UI: a headless one-shot mode, a JSON-RPC SDK, an ACP adapter for editors, and a Python SDK wheel.
It is local-first and BYOK: API keys stay in a local credentials file, with providers including Anthropic, OpenAI, Bedrock, Vertex, Azure, and any OpenAI-compatible endpoint.
Status #
New and extremely loud: 221,878 stars and 26,319 forks as of 2026-09-13, just over four weeks after the repo was created on 2026-08-13. No stable release exists, only alpha and rc prereleases (latest: dsh-v0.1.5-rc.2 on 2026-09-10, still current as of 2026-09-13), and the README warns there will be compatibility-breaking changes. The launch thread drew 747 points and 301 comments on Hacker News, with the author answering questions directly. An ecosystem is already forming: a Tauri desktop port with 2,000 stars, a plugin directory site, and an MCP plugin catalog. I read the star count as attention, not adoption, and the safest status label is developer preview.
Strengths #
- Full traceability: an append-only session log records prompts, reasoning, tool calls, and subagent scheduling, with resume, fork, and replay over one event stream.
- Capability swaps happen through configuration, not forks, which is the cleanest test of the plugin thesis.
- Minimal mode exists specifically so model benchmarks run on identical harness scaffolding.
- DeepSeek shipped it MIT with multiple provider support, so it is usable without DeepSeek models.
Cautions #
- The project’s own SAFETY.md states it has not undergone a security audit, executes model-generated code, and that its sandboxing and approval prompts do not guarantee isolation.
- Alpha software with announced breaking changes; anything built on it will churn.
- Hacker News debate flagged plugin ecosystems as an attack surface and a long-term compatibility burden, and found the Cordis paper’s novelty overstated.
- Issues are disabled in favor of Discussions, which makes the bug backlog harder to read than peers.
Pricing #
Free and open source under MIT. No hosted tier and no pricing page; costs are whatever your model provider charges.
Compared to #
- Claude Code: the turnkey choice with subscription billing and polished defaults; choose dsh when you want open internals and full traces instead.
- OpenCode: the stable community-governed daily driver; choose dsh when you want to modify the harness itself rather than configure it.
- Codex: OpenAI’s CLI with real OS-level sandboxing; dsh has stronger observability but no sandbox guarantees today.
Bottom line #
Recommended for harness builders and benchmark runners who want a hackable, fully traced agent platform, with eyes open that it is an alpha. Not for anyone whose threat model includes untrusted repos on day one, or who wants a stable daily driver this quarter.
Changes #
- 2026-08-30 - Created in an owner-directed star sweep, with five references fetched.
- 2026-09-09 - Recorded the desktop-port repo’s move to the dsh-tauri-desk org and fixed the redirected link.
See also #
- Harness Feature Matrix - the capability comparison this note joins
- Agentic Coding Tools Landscape - the map of the harness layer
- OpenCode - the stable open-source alternative
- Model Selection for Coding Tasks - what the models you plug into dsh change
References #
https://github.com/deepseek-ai/deepseek-harness - repository, README, install, MIT license, 221,878 stars as of 2026-09-13
https://www.deepseek.com/harness/en/ - runtime modes and plugin architecture claims
https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/HEAD/SAFETY.md - the project’s own security warnings
https://news.ycombinator.com/item?id=49285244 - the 747-point launch thread, including critical takes on plugins and the Cordis paper
https://arxiv.org/abs/2608.25512 - the Cordis composability paper behind the architecture