Coding agent (CLI) · DeepSeek (GitHub org deepseek-ai)
DeepSeek Harness
DeepSeek's own MIT-licensed agent harness: per its site, a coding agent with file edits, shell, planning and subagents, run through a local Web UI, a desktop app or a headless CLI. README calls it a developer preview.
DeepSeek Harness review
DeepSeek Harness, or dsh, is DeepSeek's own open-source agent harness, published on GitHub as deepseek-ai/deepseek-harness under the MIT licence. It is not DeepSeek-Reasonix, a separate project under an individual's GitHub account. We read the official site, repository docs and GitHub and npm records on 12 November 2026; we did not install or run it.
A web and desktop agent first
The documented main surface is a browser, not a terminal screen. `npx @deepseek-ai/dsh web` starts a local Web UI, by default at 127.0.0.1:3080, and the desktop README calls the desktop app an Electron shell around it. The CLI offers one-shot `headless` runs with optional JSON output, plus ACP and SDK profiles. A terminal UI appears in the docs only as an example profile.
The site describes Standard mode as a full coding agent with file editing, shell, search, planning and subagents, and Minimal mode as two tools, a persistent bash and an editor. The docs describe MCP client support and a sandbox with read-only, workspace-write and full-access modes. The pages we read show no software price. The quickstart asks for a DeepSeek API key and the providers guide lists other providers, so model usage would be billed by whichever provider you configure; we did not read API pricing.
What to weigh
- The README calls it a developer preview and warns of compatibility-breaking changes. Every GitHub release so far is a prerelease; the newest, 0.2.2-rc.2, is from 11 November.
- SAFETY.md says it has had no security audit, must not be treated as production-ready, and that sandboxing does not guarantee isolation.
- Desktop downloads cover Apple silicon Macs (macOS 13 or later) and 64-bit Windows 10 or later. A site string says Intel chips are currently not supported; we saw no Linux build.
- CONTRIBUTING.md says external pull requests are not accepted at the moment.
The repository showed about 279,000 stars on 12 November. Stars measure attention, not quality.
Facts: platforms, price and licence
- platforms
- macOSWindowsOfficial desktop downloads listed on deepseek.com: macOS (Apple silicon, macOS 13 or later; a site string says "Intel chips current not supported") and Windows (64-bit, Windows 10 or later). The same code also runs from the npm package @deepseek-ai/dsh (Node.js >=24.9.0 in package.json) as a CLI launcher: dsh web serves a local Web UI (default 127.0.0.1:3080), dsh headless runs one task and exits, plus ACP and SDK profiles. No Linux desktop build was seen on the site; the README and docs read do not state which operating systems the npm launcher supports, so linux is left out. The Web UI is served from the user's own machine, not a hosted service, so web is left out. Developer preview.
- price
- FreeSoftware is open source under MIT. The Harness page, README and user guide read show no price for the app; the quickstart says to save a DeepSeek API key from platform.deepseek.com under Settings > Models, and other providers or OpenAI-compatible endpoints can be added with their own keys, so model usage is billed by whichever provider is configured. API prices on DeepSeek's separate pricing page were not read (seen 2026-11-12).
- source
- Open sourceLicence: MIT License (LICENSE file: Copyright (c) 2026 DeepSeek)
- github
- deepseek-ai/deepseek-harness 279.7k stars, seen 2026-11-13
DeepSeek Harness features
- ✓ yes MCP
- ✓ yes Multi-agent
- ✓ yes Headless / CI
- ✓ yes Sandbox
- ? not verified Local models
- ✓ yes Your own API key
- ? not verified Git worktrees
- ? not verified SSH
✓ documented · – absent or not in its feature list · ? not verified
DeepSeek Harness strengths
- MIT licence, published by DeepSeek itself; source, docs and prerelease tags are public on GitHub and npm
- Several ways to run it: desktop app for macOS and Windows, npx Web UI, one-shot dsh headless with --json output, plus ACP and SDK (including a Python SDK guide)
- Docs describe a file-effect sandbox with read-only, workspace-write and full-access modes, approval prompts, and MCP client support
- Model-agnostic setup: DeepSeek API key first, but other providers and OpenAI-compatible endpoints can be added in Settings
DeepSeek Harness weak points
- No terminal UI found: the CLI docs list web, headless, acp, sdk and sdk-minimal profiles and mention a tui profile only as an example you would install; the docs describe the Web UI and desktop app as the interactive surfaces
- README says it is a developer preview with compatibility-breaking changes expected; SAFETY.md says it has not had a security audit, must not be treated as secure or production-ready, and sandboxing does not guarantee isolation
- Desktop downloads on the site are macOS Apple silicon and Windows x64 only; a site string says Intel chips are currently not supported and no Linux build was seen
- CONTRIBUTING.md says external pull requests are not accepted at the moment; agent teams, auto approval review, scheduled tasks and voice input are labelled Experimental in the site's plugin preview
Who DeepSeek Harness is for
- best for
- Developers who want to try a vendor-made, MIT-licensed agent harness in a local browser or desktop window, or script it headlessly, and accept developer-preview stability.
- look elsewhere if
- Anyone who needs a terminal UI today (none found), an Intel Mac or Linux desktop app, or software the vendor documents as production-ready.
Lists that include DeepSeek Harness
Sources and what we could not verify
Checked 2026-10-21. Prices and features change; the vendor's page is the authority.
- www.deepseek.com/en/harness — official page: preview status, desktop downloads (macOS Apple silicon macOS 13+, Windows 64-bit Windows 10+), Intel not supported string, Standard/Minimal/Code/Creator modes, experimental plugin labels, MIT footer, npx command, links to GitHub and docs
- www.deepseek.com/en/download — download page lists macOS arm64 dmg (Apple silicon, macOS 13 or later) and Windows x64 exe (Windows 10 or later); its strings include 'Intel chips current not supported'; no Linux mention
- github.com/deepseek-ai/deepseek-harness — official repository, MIT, developer preview, run instructions
- api.github.com/repos/deepseek-ai/deepseek-harness — 278908 stars (authenticated gh api, 2026-11-12; an earlier read gave 278905), 34308 forks, MIT, created 2026-08-13, pushed 2026-11-11, not archived, homepage deepseek.com/harness, default branch master, owner deepseek-ai (Organization)
- api.github.com/repos/deepseek-ai/deepseek-harness/releases/latest — returned 404 Not Found
- api.github.com/repos/deepseek-ai/deepseek-harness/releases?per_page=8 — dsh-v0.2.2-rc.2 published 2026-11-11; all 38 releases (paginated) flagged prerelease, no release assets
- registry.npmjs.org/@deepseek-ai/dsh — npm dist-tags latest 0.2.2-rc.1 (published 2026-11-04), no rc.2 version listed, license MIT
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/README.md — agent harness by DeepSeek AI, developer preview, compatibility-breaking changes, npx @deepseek-ai/dsh web on 127.0.0.1:3080, MIT
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/apps/cli/package.json — npm package @deepseek-ai/dsh 0.2.2-rc.2, engines node >=24.9.0, bin dsh
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/LICENSE — MIT License, Copyright (c) 2026 DeepSeek
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/SAFETY.md — experimental developer-preview; no security audit; must not be treated as secure or production-ready; sandboxing, approval prompts and permission controls do not guarantee isolation
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/CONTRIBUTING.md — external pull requests not accepted at the moment; community plugins via dsh-plugin topic
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/apps/cli/README.md — entry modes: web, headless, acp, sdk, sdk-minimal; tui only as an example profile (assuming it is installed); Node launcher
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/cli-help.md — headless options (--json, --session-id, --permission-mode, --approval-policy, --add-dir), list of shipped profiles
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/index.md — Web UI quickstart, DeepSeek API key under Settings > Models (platform.deepseek.com), agent reads/edits files, runs commands, delegates, plans, asks for approval; links a Python SDK guide (python-sdk.md exists, not read in full)
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/providers.md — third-party providers and custom OpenAI/Anthropic/Gemini-protocol endpoints
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/apps/desktop/README.md — desktop is an Electron shell around the complete dsh Web application, bound to 127.0.0.1; Manage dsh Command installs dsh on macOS and Windows; known limitations; product analytics policy for Desktop
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/packages/sandbox/README.md — read-only, workspace-write, danger-full-access modes; Linux, macOS and Windows backends; confinement shares host kernel and filesystem
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/packages/mcp/README.md — MCP client and resources
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/packages/team/README.md — agent team shared task board for subagent trees
- raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/tool-catalog.md — subagent, agent-team, experimental worktree, workflow, schedule tools
- github.com/esengine/DeepSeek-Reasonix — different project: separate repository under the individual GitHub user account esengine (type User, not an organization); not mentioned in the deepseek-harness README or guides read; relationship to DeepSeek not checked
- github.com/orgs/deepseek-ai — owner org deepseek-ai: display name DeepSeek, blog www.deepseek.com, 46 public repos; GitHub 'verified' badge false; ownership tied to the vendor by the official site linking to this repo and the repo homepage pointing to deepseek.com/harness
Not verified
- Which operating systems the npm launcher officially supports (README and docs read do not say); linux therefore not listed even though the sandbox package has a Linux backend and CI runs Linux lanes
- ssh: packages/ssh is described as a remote-execution provider for headless or custom profiles, not an SSH client; left null
- git_worktrees: only an experimental create_worktree tool appears in the tool catalog; left null
- local_models: a custom endpoint can point at a self-hosted server per the providers guide, but local-model support is not stated; left null
- phone_remote and mobile_app: no mobile app or phone control found; left null, not false
- Desktop account sign-in: the desktop README's known-limitations list says sign-in is not connected (button disabled) while the same README describes a working sign-in flow; unresolved
- Desktop product analytics: the desktop README says desktop follows a product collection policy with a live app setting; the policy was not read
- DeepSeek API pricing: the pricing page was not read, so model cost is not stated
- Benchmark scores and the reported Hugging Face harness leaderboard: not on any official page read (BENCHMARK.md only explains how to run benchmarks); a web search surfaced Terminal-Bench figures on secondary sites only and found no leaderboard entry, so none are recorded
- Whether the project is hands-on usable as described: nothing was installed or run
- Star count 278,908 confirmed via authenticated GitHub API on 2026-11-12; star history and any earlier press figures were not verified
- The site's 'Intel chips current not supported' is a string in the page's locale bundle; whether it is shown to visitors on Intel Macs was not checked
- Site 'Terminal' plugin entry is described as 'Set command time and output limits', not a terminal UI; whether any terminal UI ships is not confirmed beyond the CLI docs
- Experimental plugin labels (agent teams, auto approval review, scheduled tasks, voice input) were read from a plugin preview on the site; the Desktop README also says desktop product analytics follow a collection policy that was not read
- Sandbox enforcement and its defaults were not tested; the docs describe modes only