$ buildwith info deepseek-harness

DeepSeek Harness

#5 in Agent Ops · Free / OSS

sources differ by section

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI, released under the MIT license. It is built on Cordis with an everything-is-a-plugin design: tools, skills, and even UI panels are plugins you can install or have the agent write for you. It runs as a local web UI (npx @deepseek-ai/dsh web) or as a desktop app, and DeepSeek calls it a developer preview with compatibility-breaking changes still to come.

for Builders who want a hackable, self-hosted agent workspace for code, documents and background jobs, and who are comfortable running developer-preview software in a sandboxed environment.

use this if you want an open-source, plugin-first agent harness you can run as a local web UI or desktop app and point at DeepSeek, Claude, OpenAI, Kimi or your own gateway.

open deepseek.com ↗

Features

$ features --list

  • ›Everything-is-a-plugin architecture on Cordis, with a plugin market and a dsh-plugin GitHub topic for discovery
  • ›A Create mode where the agent writes, installs and checks a new plugin from a chat request
  • ›Workspace agent that reads and edits files, runs commands, delegates to subagents and keeps a plan, asking before operations that need approval
  • ›Scheduled tasks (for example a weekly report every Friday) and background jobs with progress you can check later
  • ›Built-in DeepSeek models plus third-party providers such as Anthropic, OpenAI, Kimi and GLM, or any custom OpenAI Chat Completions, OpenAI Responses or Anthropic Messages endpoint
  • ›A developer trace view of every turn, tool call, argument, result and timing, plus a Python SDK, MCP memory and GitHub review guides

Pricing

$ pricing --tiers

Open source

Free

MIT license. Install with npx @deepseek-ai/dsh web or build from source. Model usage is billed separately by whichever provider's API key you add, such as the DeepSeek API.

Why this rank

$ diff --rank 5

Ranked #5 in agent ops by editorial hold (manualOverride), immediately after Hermes, Herdr, OpenClaw and Paperclip. DeepSeek Harness has huge early adoption (278,715 GitHub stars on 2026-11-11, a verified adoption score of 100), but it is an unaudited developer preview with breaking changes promised, and its other dimensions have not been measured yet. Placeholder scores are not shown on the site. With adoption measured, its weighted total (88) also puts it at #5, between Paperclip and OpenHuman. Pick it if you want DeepSeek's open, plugin-first harness and can run it in a sandbox.

Score breakdown

$ scores --rubric

Measured so far · how we rank

Adoption

20% weight

100verified

How widely builders already use it for this kind of work. — 283,433 GitHub stars on 2026-11-18. Agent Ops adoption scale: 20 × log10(stars), rounded, capped at 100 (1k stars = 60, 10k = 80, 100k+ = 100); here 20 × log10(283,433) = 109.0, capped at 100. Stars are a proxy for builder adoption, not a usage count. · verified 2026-11-18

Strengths

$ strengths

  • ›Backed by DeepSeek and already widely adopted: 278,715 GitHub stars and 34,273 forks on 2026-11-11, less than three months after the repo was created on 2026-08-13.
  • ›Not locked to DeepSeek models: third-party and custom API providers plug in from Settings → Models.
  • ›Plugins can extend tools, skills and the interface, so the harness can be reshaped rather than just configured.
  • ›Rapid release pace: tagged dsh-v0.2.x pre-releases continue, the latest dsh-v0.2.2-rc.2 on 2026-11-11.

Limitations

$ limits

  • ›Developer preview: the README warns there will be compatibility-breaking changes, and all recent GitHub releases are alpha or release-candidate tags.
  • ›SAFETY.md says it has not had a security audit and must not be treated as secure or production-ready; it recommends a disposable VM or container.
  • ›It can run model-generated commands and load third-party plugins, and sandboxing and approval prompts reduce risk but do not guarantee isolation.
  • ›The official product page at deepseek.com/harness is in Chinese; the English documentation lives in the GitHub repo.

Sources

$ sources --verified

$ buildwith alternatives deepseek-harness

  • #1 Hermes AgentFree / OSS

    use this if you want an always-on, model-agnostic agent that learns skills over time and lives in your terminal and chat apps.

  • #2 HerdrFree / OSS

    use this if you want your coding agents to keep running after you close the laptop and to reattach from any terminal.

  • #3 OpenClawFree / OSS

    use this if you want a hugely popular open-source personal agent that actually does things on any OS and platform.

  • #4 PaperclipFree / OSS

    use this if you want to run a team of agents (Claude, Codex, Cursor, Hermes and more) toward one goal, with roles, budgets and tickets.

  • #6 OpenHumanFree / OSS

    use this if you want a fast, lightweight open-source agent harness with built-in memory that can run many agents on a small server.

  • #7 BuzzFree / OSS

    use this if your team wants humans and agents working in the same rooms with every message, workflow step and approval in one auditable log.

  • #8 ComposioFreemium

    use this if your agent needs authenticated access to real apps (Gmail, Slack, GitHub and many more) without you building OAuth and integrations yourself.

  • #9 OpenMausBotFree / OSS

    use this if you want a chat-app roster of bots running on the Claude, Codex or Grok CLIs and subscriptions you already have, each with its own computer.

  • #10 RakazoFree / OSS

    use this if you want persistent, self-hosted AI teammates with their own browser and shell, on your own model keys and machines.

  • #11 OpenBotFreemium

    use this if you want a self-hostable workspace where AI coworkers get their own computers, under company sign-in, tool permissions, approvals and audit.

  • #12 HarnessFreemium

    use this if you run Claude Code, Codex, Cursor and others across several machines and want one place to watch and steer them.

sponsor this spot →