DeepSeek Harness

license

MIT(OSI)

source

public

core features withheld

no — the tree at 687f282 carries one root MIT LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path, and every package manifest that declares a license declares MIT or BSD-3-Clause. DeepSeek calls the harness open source and sells model access, which it reaches through an API key like any other provider

DeepSeek Harness is MIT licensed, and its public repository holds the whole harness, Web and desktop apps included, so it builds and runs from published source. What DeepSeek sells is access to its models, which the harness reaches with an API key in the same way it reaches other providers.

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/index.md recorded 2026-10-19

  The Web UI guide's first step is to open Settings, Models and enter a DeepSeek API key, linked to platform.deepseek.com, and it points to the model configuration guide for other providers.

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/LICENSE recorded 2026-10-19

  The root LICENSE is the MIT License text, "Copyright (c) 2026 DeepSeek".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/README.md recorded 2026-10-19

  The README says "DeepSeek Harness (`dsh`) is an open-source agent harness developed by" DeepSeek AI, documents running it from a clone with pnpm install and build, and closes "## License [MIT](LICENSE)"; no paid tier or license key is mentioned.

• https://registry.npmjs.org/@deepseek-ai/dsh/latest recorded 2026-10-19

  The npm record for @deepseek-ai/dsh 0.2.0-rc.2 declares "license": "MIT", a repository of deepseek-ai/deepseek-harness under apps/cli, and the dsh command.

Adoption is the trailing month of npm downloads of @deepseek-ai/dsh, the package the README runs with npx. Installs of the Python SDK, which carries its own copy of the dsh command, and of the desktop app are not counted.

• https://api.npmjs.org/downloads/point/last-month/@deepseek-ai/dsh recorded 2026-10-19

  1,723,718 downloads of @deepseek-ai/dsh between 2026-09-17 and 2026-10-16

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/README.md recorded 2026-10-19

  The README's install instruction is "npx @deepseek-ai/dsh web", which starts the Web UI.

DeepSeek Harness covers what a coding agent is expected to do, with sub-agents, plan mode, MCP, many model providers and an SDK, and adds per-platform process sandboxing and approval prompts that limit what the agent can touch. It sits a step below OpenHands, which publishes benchmark results, because it is still a developer preview with no published benchmark result and no security audit.

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/apps/cli/README.md recorded 2026-10-19

  The CLI's entry modes include "dsh web", a headless mode that runs one session and prints the answer, an SDK server over JSON-RPC stdio, and "Serve automation clients over ACP stdio".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/BENCHMARK.md recorded 2026-10-19

  The benchmark page only explains how to run benchmark tasks through the Python SDK; it reports no result.

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/index.md recorded 2026-10-19

  The Web UI guide says "The agent can read and edit workspace files, run commands, delegate work, and maintain a plan" and that the Web UI "asks before operations that require approval under the active permission policy".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/providers.md recorded 2026-10-19

  The model guide offers built-in providers such as "`anthropic`, `openai`, `moonshotai` for Kimi, or `zai` for GLM" and custom endpoints speaking "OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/packages/README.md recorded 2026-10-19

  The package map lists groups for subagents, plan collaboration, skills, "External Model Context Protocol servers exposed as native tools", an LSP tool, "POSIX remote connection", "Host-owned scheduled follow-ups", an "Out-of-process SDK" and an "Automation-only Agent Client Protocol server".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/packages/sandbox/README.md recorded 2026-10-19

  The sandbox group runs commands "read-only", under "workspace-write" or "danger-full-access", with backends "Linux bwrap then Landlock, macOS Seatbelt, Windows restricted token".

• https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/SAFETY.md recorded 2026-10-19

  The safety notice says the harness "is experimental developer-preview software" that "has not undergone a security audit".