# DeepSeek Harness

DeepSeek's open-source, plugin-based harness for running AI coding agents locally

**Verdict: caution.** DeepSeek Harness is a free, MIT-licensed way to run a coding agent from a local web UI, reading and editing files and running commands with DeepSeek, Anthropic, OpenAI, Kimi, GLM or custom compatible models. Caution because DeepSeek's own safety notice calls it unaudited and not secure or production-ready, and a critical flaw, CVE-2026-82533, let a sandboxed agent switch off its own sandbox on unpatched installs. Right for experimenters in a disposable VM; wrong for machines holding credentials or production code. Last verified 2026-11-12.

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI, MIT licensed and published as a developer preview whose README warns of compatibility-breaking changes. It is built on the Cordis framework around an everything-is-a-plugin design, and its architecture docs name model adapters, tools, persistence, sandbox and approval policy as parts of the base layer. The quickest start is npx @deepseek-ai/dsh web, which serves a Web UI at http://127.0.0.1:3080; it can also be started with pnpm dsh web, and the architecture docs describe an Electron desktop application. In the Web UI you enter a DeepSeek API key and choose a workspace, and the agent can read and edit workspace files, run commands, delegate work and maintain a plan, asking before operations that need approval under the active permission policy. Built-in provider ids include anthropic, openai, moonshotai for Kimi and zai for GLM, and custom providers such as company gateways or self-hosted servers can use the openai-completions, openai-responses or anthropic-messages protocols. OX Research describes it as local-first and says agent commands run inside a bubblewrap, Landlock or Seatbelt sandbox depending on the platform. The Hacker News reports that the sandbox restricted file writes only, not reads or network access. DeepSeek's SAFETY.md says it can execute model-generated code and commands and access the network, processes, credentials and files made available to it, and that it has not undergone a security audit and must not be treated as secure or production-ready. CVE-2026-82533, rated 9.4 under CVSS 4.0 and published by VulnCheck on September 8, 2026, let a sandboxed agent use the unauthenticated local API to set its session to danger-full-access and bypass approval prompts, according to The Hacker News. VulnCheck names 0.1.2-alpha.1 as the fixed version, and The Hacker News reports that the first fixed release published to npm was 0.1.2-alpha.2. The releases page read on November 12, 2026 lists only pre-releases, with dsh-v0.2.2-rc.2 as the newest tag after dsh-v0.2.2-rc.1 and dsh-v0.2.2-alpha.3, and none carries a security note; the page's date labels did not render reliably, so we cite tags only.

## Strengths

- Running a coding agent from a local browser UI: npx @deepseek-ai/dsh web starts it at http://127.0.0.1:3080, and the agent can read and edit workspace files, run commands, delegate work and keep a plan
- Choosing your model: besides a DeepSeek API key, built-in provider ids include anthropic, openai, moonshotai for Kimi and zai for GLM, and custom providers can speak openai-completions, openai-responses or anthropic-messages
- Developers who want to rework the agent itself, since everything is a plugin on the Cordis framework, including model adapters, tools, persistence, sandbox and approval policy
- Zero software cost under the MIT licence
- Production or sensitive machines: DeepSeek's SAFETY.md says it has not undergone a security audit, must not be treated as secure or production-ready, and recommends a disposable virtual machine, container or dedicated environment
- Stable integrations, since the README warns of compatibility-breaking changes in the developer preview
- Agents that read untrusted content on unpatched installs: under CVE-2026-82533, attacker-supplied text the agent read could get it to disable its own sandbox with one shell command, on shipped defaults and with no network exposure
- Trusting the CVE record's fix version on npm: The Hacker News reports 0.1.2-alpha.1 was fixed on GitHub only and never published to npm, so npm users need 0.1.2-alpha.2 or later

## Pricing

- Model: open-source
- Open source (MIT): $0

Website: https://github.com/deepseek-ai/deepseek-harness

## Sources

- [DeepSeek Harness docs, architecture](https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/architecture.md)
- [VulnCheck advisory, CVE-2026-82533](https://vulncheck.com/advisories/deepseek-harness-alpha-1-authentication-bypass-via-host-header-spoofing)
- [deepseek-harness GitHub repository and SAFETY.md](https://github.com/deepseek-ai/deepseek-harness)
- [OX Research: CVE-2026-82533](https://www.ox.security/blog/cve-2026-82533-deepseek-harness-ai-agent-sandbox-escape/)
- [DeepSeek Harness docs, model providers](https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/providers.md)
- [DeepSeek Harness docs, Web UI guide](https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/docs/user/guide/index.md)
- [DeepSeek Harness SAFETY.md](https://raw.githubusercontent.com/deepseek-ai/deepseek-harness/master/SAFETY.md)
- [DeepSeek Harness releases](https://github.com/deepseek-ai/deepseek-harness/releases)
- [The Hacker News: DeepSeek Harness flaw let AI agents disable their own file sandbox (Sep 9, 2026)](https://thehackernews.com/2026/09/deepseek-harness-flaw-let-ai-agents.html)

Canonical: https://neomanex.com/directory/deepseek-harness
