DeepSeek Harness (dsh) is DeepSeek AI’s open-source agent harness — the local runtime that lets an AI model use your files, commands, and a browser UI — shipped as a developer preview on August 13, 2026 (product page). It earns 7.6/10: free, MIT-licensed, and packed with features, but a developer preview with expected breaking changes and no security audit. Best treated as a learning tool on a throwaway setup, not a daily driver.
What Is DeepSeek Harness (dsh)?
An agent harness is the software layer between an AI model and your computer: it runs the conversation loop, turns model replies into tool calls and file edits, and logs results. DeepSeek Harness is that layer as a local runtime, adding sandboxed commands (a sandbox confines what software may touch) and a browser UI (product page). DeepSeek AI released it open-source as a developer preview under MIT on August 13, 2026 — roughly 210K GitHub stars so far (GitHub repository). The slogan “Everything is a plugin” is literal: models, tools, skills, sessions, sandbox, and UI are swappable plugins — bolt-on modules you can add or replace — built on the open-source Cordis kernel. It is provider-agnostic: any DeepSeek or OpenAI-compatible API key — the pay-per-use credential your provider issues — works (product page).
The Scores: 7.6/10 for a Promising Dev Preview
The overall score is the mean of the sub-scores: (8.0 + 9.0 + 7.0 + 8.0 + 6.0) / 5 = 7.6. Ease of Use rewards the one-command install and browser UI; Features the Web UI, CLI, Python SDK, sandbox, MCP bridge, and ACP in one free package; Performance is docked for breaking changes and no security audit; Documentation is solid but fast-moving; Support has no paid options, GitHub issues disabled.
| Overall Score | 7.6/10 |
|---|---|
| Ease of Use | 8.0/10 |
| Features | 9.0/10 |
| Performance | 7.0/10 |
| Documentation | 8.0/10 |
| Support | 6.0/10 |
The Good: Free, Open Source, and Packed With Features
One command — npx @deepseek-ai/dsh web (Node.js required) — opens a Web UI at http://127.0.0.1:3080; pick a model, API key, and workspace, then chat, with approval prompts before risky actions (documentation). Every session logs a trajectory — a step-by-step record of what the model saw: system prompts, reasoning, tool calls, subagent runs — that you can resume, fork, search, or replay (documentation). Sessions run one of four presets (ready-made agent configurations): Standard (full toolset), PTC (multi-step operations in one TypeScript program), Minimal (persistent shell plus file editor), or Creator (author your own) (documentation). A headless CLI (command-line interface) runs dsh --profile headless "job" for one fresh session and a printed answer; a Python SDK (pip install deepseek-harness-sdk, Python 3.10+) exposes the same runtime to code (PyPI). Sandboxing fails closed with three modes — read-only, workspace-write, or danger-full-access — via Landlock on Linux 5.13+, Seatbelt on macOS, and restricted tokens on Windows (documentation). An opt-in MCP bridge — MCP (Model Context Protocol), the open standard for attaching external tools — connects third-party servers whose tools act like native ones (tools are the only MCP capability bridged); dsh --profile acp speaks the Agent Client Protocol, a standard for client apps driving agents, so automation runs without a human (documentation). Everything is free forever under MIT, with no paid tier or gated features (GitHub repository).
The Not-So-Good: It’s a Developer Preview
The project’s own warnings are blunt: documentation states “THERE WILL BE COMPATIBILITY-BREAKING CHANGES,” so presets, scripts, or workflows may break between releases (documentation). There is no security audit, and official SAFETY guidance says the tool is not production-ready: it can execute model-generated code and commands, load third-party plugins, and reach the network, credentials, and files. The sandbox is same-world confinement sharing your host kernel and filesystem — not a container or VM — so it reduces risk but doesn’t guarantee isolation (documentation). The MCP bridge covers tools only; MCP resources and prompts are not supported. GitHub issues are disabled, so feedback goes through Discussions (GitHub repository). Finally, long-term tutorials and community answers don’t exist yet — in our view, the gap beginners feel first.
DeepSeek Harness vs. the Alternatives
dsh is not the only free local agent runtime. Hermes Agent (Nous Research) is the closest free, open-source analog — a more established local agent framework with plugins and skills — and the safer pick for a general-purpose agent today (Beginner’s Guide to Hermes Agent). Claude Agent SDK / Claude Code (Anthropic) is the polished commercial option if you live in the Claude ecosystem and don’t mind proprietary software (Agent SDK docs). OpenClaw is a free, open-source, model-agnostic gateway for connecting agents to your messaging channels, but it lacks dsh’s built-in web UI (Beginner’s Guide to OpenClaw). For a side-by-side look at dozens of AI agents, see our AI Agent comparison page. Pick dsh for maximum features at zero cost, Hermes for stability, Claude for polish.
Pricing: The Tool Is Free — the Model Isn’t
The tool itself costs nothing: free and open source under MIT, no paid tiers, no gated features (GitHub repository). What you pay for is the model, billed by API tokens — text chunks, roughly word-sized (DeepSeek API pricing). Rates changed around mid-August 2026, so don’t trust secondhand numbers — check the official page above. Because dsh speaks OpenAI-compatible endpoints, other providers’ keys work too (documentation).
How We Tested
This review is based on official documentation, pricing pages, and published third-party reports — we did not run the tool hands-on. All product facts were verified against official sources on September 2, 2026.
Who Should Use DeepSeek Harness (and Who Shouldn’t)
Try it if:
- You like tinkering and learn by doing.
- You’re OK with occasional breakage and re-installs.
- You want a free, model-agnostic runtime with a visible trajectory of what the model did.
- You’ll run it on a disposable machine.
Wait if:
- You need a dependable daily tool for real work.
- You’re not comfortable with the command line yet.
- You can’t tolerate breaking changes or an absent security audit.
- You want one-click setup and polished support.
The Bottom Line
DeepSeek Harness is the most feature-generous free, open-source agent harness around; 7.6/10 reflects real strengths alongside dev-preview risk. Try it on a throwaway setup if you’re curious; wait for a stable release if you need reliability. Beginners should lean on the official documentation, work in a disposable folder, and keep backups. Prefer proven software? Our Hermes Agent review covers a steadier option; we also track tools like this on our roadmap. This review came from ToolBrain’s automated pipeline (how it works).
Back to all posts
