One-line positioning
DeepSeek broke Agent capability into plugins — Harness is an open-source "everything-is-a-plugin" agent runtime you can recombine at every layer
DeepSeek Harness is a developer preview (v0.1) released by DeepSeek on August 13, 2026, open-sourced the same day under the MIT license. It is not a chat product but a software framework that turns AI models into autonomous agents. The core principle is "everything is a plugin": models, tools, skills, sessions, sandboxes, storage, scheduling, and UI are all composed from plugins, freely swappable and recombinable. Underneath sits the Cordis plugin system (powering the Koishi chatbot ecosystem for four years) — the framework itself has no privileged core.
Getting Started (based on official docs and OSS community tests, not first-person)
Note: compiled from the official launch post, GitHub repo, and third-party technical reviews, rewritten and condensed by editors — not a claim that we tested each point ourselves.
- Run it: with Node.js installed, one command `npx @deepseek-ai/dsh web` launches the Web UI locally at http://127.0.0.1:3080, no source clone needed.
- Pick a mode: Standard (full tool set), PTC (model-generated code orchestrating multi-round tool calls), Minimal (only Shell + file edit, for model benchmarks), Creative (experiment with Cordis plugins in memory, compose new modes).
- Swap models: the inference layer is a plugin, so the same Harness can drive DeepSeek's own models or point at Anthropic, OpenAI, Bedrock, Vertex, Azure, or any OpenAI-compatible endpoint — switching provider is a config change, not a rewrite.
Real Pain (OSS community and review consensus)
Note: recurring points from official warnings and early reviews, condensed by editors, not individually verified by us.
- Early preview: v0.1 is explicitly flagged "may have breaking changes"; core plugins and base interfaces will iterate fast, not production-ready.
- No GA date or pricing: still a developer preview, no general-availability date, no commercial pricing (MIT OSS itself is free).
- Developer-facing: needs Node.js toolchain and some engineering skill; not out-of-box for non-technical users — a different category from desktop Agent products.
- No independent benchmarks: official self-reported Terminal-Bench 2.1 at 87.9 etc., but no third-party (e.g. Artificial Analysis) reproduction yet; treat magnitude as company-reported.
Hard Comparison (one-line verdicts)
- Claude Code: closed commercial Agent tool, works out of the box; Harness is an open framework, more flexible but self-assembled — "scaffold" not "finished product".
- OpenCode: more polished but less flexible than Harness; reviews call Harness more radical on "composability".
- Pi: deliberately minimal, non-modular core; Harness goes opposite, even the DI layer is swappable.
Who It's For / Not For
- For: developer teams wanting to control Agent behavior and deployment, engineering orgs needing one multi-model base, companies willing to fork and repackage the eight layers commercially.
- Not for: casual users wanting a chat app, those refusing CLI and config, or enterprise buyers needing vendor SLA and stable GA.
Editor's Take (our site's view)
Note: our judgment from a design / front-end / tech-selection angle — the angle that sets us apart from spec-repost sites.
DeepSeek's move is interesting: it isn't just selling models, it open-sourced the "connective layer between model and software" too. For small teams this means forking and replacing any layer (e.g. point the filesystem at a remote sandbox, Bash/PTY/LSP migrate together, zero changes) — a composability closed Agent tools can't match. The MIT license matters: commercial use, modification, redistribution allowed, not a gesture. But a cold splash: v0.1 is early, the breaking-changes warning isn't courtesy, real production waits for GA and third-party benchmarks. It is not the same thing as "out-of-box desktop Agents" like WorkBuddy or QwenWork — Harness is the foundation to build houses on, not a furnished house. Our advice: technical teams can prototype now and validate whether "pluginized Agent" is worth betting on; enterprise buyers shouldn't treat it as a finished Agent to purchase — it is a candidate base for your future Agent platform, not today's ops object.
Ratings
Architecture flexibility: ★★★★★
Open-source ecosystem (MIT/commercial): ★★★★★
Ease of use (onboarding barrier): ★★☆☆☆
Docs and maturity: ★★★☆☆