DeepSeek Harness: what it is, how to install it, and how it compares

DeepSeek Harness at a glance

How to install and use DeepSeek Harness

1. Pick desktop or web

   Download the desktop app from deepseek.com/en/harness for macOS (Apple silicon) or Windows (64-bit); on macOS it can also install a dsh command. On any machine with Node.js, run the web UI with npx, or build from source.

   # Run the web UI with Node.js
   npx @deepseek-ai/dsh web
   # then open http://127.0.0.1:3080
   
   # Or build from source
   git clone https://github.com/deepseek-ai/deepseek-harness.git
   cd deepseek-harness && pnpm install && pnpm run build && pnpm dsh web

2. Sign in or add an API key

   On first launch, either sign in with your DeepSeek account in the browser or paste an API key from platform.deepseek.com under Settings → Models. Keys are stored in ~/.dsh/.credentials.yaml ($DSH_HOME).

3. Give it project instructions

   DeepSeek Harness reads AGENTS.md or CLAUDE.md in your project (plus .local.md variants), and a global ~/.dsh/AGENTS.md. Existing instructions written for Claude Code or Codex carry over; see the CLAUDE.md guide and Codex AGENTS.md guide.

4. Choose how much it may do

   Pick a sandbox mode (read-only, workspace-write or danger-full-access) and an approval policy (ask or never). DeepSeek's safety notes say the harness is experimental, hasn't had a security audit, and the sandbox doesn't guarantee isolation, so start read-only on code you care about.

5. Run tasks headless and add plugins

   Use dsh headless for one-off tasks in scripts, --resume or --session-id to continue a session, and dsh plugin to add plugins such as scheduled tasks, subagents or agent teams.

   # One task, no UI, with a permission mode
   dsh headless "Add tests for src/auth" --permission-mode workspace-write
   
   # Stream events as JSON, e.g. for scripts
   dsh headless "Summarize open TODOs" --json
   
   # Add a plugin to a profile
   dsh plugin --profile default add <package>

Source: DeepSeek Harness repository and docs

DeepSeek Harness pricing: what it costs to run

The harness is free and open source. Usage is billed by the DeepSeek API at these rates, in US dollars per million tokens (api-docs.deepseek.com, November 6, 2026). Off-peak rates are half of peak; peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. DeepSeek documents no free quota, so "using it for free" in practice means free software plus whatever API credit you have.

Main features

• Plugins for everything: tools, models, interfaces and behaviors are plugins; install them per profile or create one by asking in chat ("Creator mode"). Community plugins use the GitHub topic dsh-plugin.
• Official and experimental plugins listed by DeepSeek include agent teams, auto approval review, scheduled tasks, voice input, a terminal, an agent loop and subagents.
• Skills (SKILL.md folders), MCP servers, and hooks, including an experimental bridge that runs existing Claude Code hooks.
• Subagents in-process, or backed by ACP, Codex, Claude Code or another Harness.
• Sessions you can resume, compaction, headless mode with JSON events, and an SDK for building on top.
• The Super-agent: a separate MIT bundle, dsh-superagent, makes the harness act proactively on a GitHub repository; DeepSeek calls it a research prototype. See DeepSeek Super-Agent explained.
• The paper often linked with it, "A Programming Paradigm for Spatiotemporal Composability" (arXiv, August 2026), describes Cordis, the framework underneath, not the harness itself.

DeepSeek Harness vs Claude Code vs Codex vs OpenCode

DeepSeek doesn't publish a comparison; this is our reading of each tool's own docs on November 6, 2026. DeepSeek also documents how to use its models inside Claude Code, Codex and OpenCode, so you can mix and match.

Questions people search for

• Linux: there's no Linux desktop download yet; on Linux, run the web UI with npx @deepseek-ai/dsh web or build from source.
• VS Code: DeepSeek doesn't ship a VS Code extension; use the desktop or web UI alongside your editor, or use DeepSeek models inside an editor agent DeepSeek documents, such as Claude Code or Codex.
• Ollama and local models: not documented. The harness accepts a custom OpenAI- or Anthropic-compatible endpoint for a self-hosted server, which is the documented route for your own models.
• TUI: no terminal UI ships by default; a custom profile or the experimental terminal plugin are the documented options.
• Paper: see the Cordis paper above; DeepSeek asks to cite the harness itself as "DeepSeek Harness: Everything is a Plugin".

Give DeepSeek Harness a job, not just a task

Crewmark supports DeepSeek Harness alongside Claude Code, Codex and OpenCode. It turns a session into an AI employee with a role, a goal and working hours, and wakes it on a heartbeat to take the next issue from a shared board.

1. Pick DeepSeek Harness as the agent

   Run the setup command in your project; the employee runs in a DeepSeek Harness session on your machine, billed to your DeepSeek API account.

2. Give it a role and a goal

   A software engineer, an SEO specialist, a growth marketer: each with a number to move. Read what an AI employee is.

3. Let the heartbeat start the work

   During working hours it wakes every 15, 30 or 60 minutes, takes the next issue and messages you when it needs a decision. Compare approaches in what is a superagent.

DeepSeek Harness FAQ