DeepSeek Harness

• Running a local coding agent on a repository with DeepSeek models and reviewing its file changes as side-by-side diffs
• Pointing the harness at a company gateway or self-hosted model server through a custom OpenAI- or Anthropic-compatible endpoint
• Writing a plugin that adds a tool, model adapter, or sandbox without forking the harness
• Driving the agent from a Python program through the SDK, with a disposable workspace for each task
• Starting an automatic read-only review session when a GitHub pull request moves from draft to ready for review

DeepSeek Harness is most interesting as a runtime you can take apart. Because models, tools, sandboxes, and the UI are all plugins, and every context injection is logged, it suits developers who want to see and change how an agent works. Multi-provider support and a Python SDK widen its use beyond DeepSeek models. For day-to-day coding, its maturity is the constraint: only pre-release versions exist, breaking changes are expected, DeepSeek itself says it is not security audited, and the newest capabilities are experimental. Run it in an isolated environment and treat it as a preview.

About DeepSeek Harness

DeepSeek Harness is DeepSeek's own agent harness, built on the Cordis plugin framework with an everything-is-a-plugin design: models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling, and the UI are all plugins, composed through configuration rather than source changes. You start it with `npx @deepseek-ai/dsh web`, which serves a web UI on 127.0.0.1:3080. You then add a DeepSeek API key under Settings → Models, choose a workspace folder, and give the agent a task. It can read and edit files, run commands, delegate work, and keep a plan, and it asks before actions that the active permission policy marks for approval. DeepSeek is the default provider, but the built-in catalog also covers providers such as Anthropic, OpenAI, Moonshot AI (Kimi), and Z.ai (GLM), and you can add any gateway or self-hosted server that speaks OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages. Providers that sign in with OAuth, such as Codex, are not supported yet. DeepSeek describes four modes: Standard (the full coding agent), Code mode (called PTC in the Chinese docs), where the model combines multi-step tool calls in one TypeScript program, Minimal, with only persistent bash and a file editor, and Creator, for building custom agent presets. The same `dsh` command also runs one-shot headless jobs, serves editors over the Agent Client Protocol (ACP), and backs a Python SDK. Sessions are logged append-only, including system prompts, reasoning, tool calls, and every context injection, so you can trace why the agent did something. Pre-releases in September 2026 added experimental Browser Use and Computer Use, terminals in the web sidebar, an experimental Auto review mode, and scheduled reminders. The repository also contains an Electron desktop app, but DeepSeek's harness page and the GitHub releases do not offer an official download, so the npx web UI is the documented way to run it. The tradeoff is maturity. Every release so far is a 0.1.x alpha or release candidate, the README warns of compatibility-breaking changes, and DeepSeek's safety notice says the project has not been security audited and should run in a disposable VM or container. It fits developers who want to experiment with or extend an agent runtime more than teams looking for a stable daily coding tool.