This article was produced with AI assistance. Editorial standards apply — see our About editorial process.

AI Edited DeepSeek Harness: an engineer running a local plugin workspace beside a browser agent session
AI-edited illustration. View raw image.

Key takeaways

  • The README calls the project an open-source agent harness, MIT-licensed, in developer preview, with compatibility-breaking changes expected.
  • npx @deepseek-ai/dsh web serves a local Web UI at 127.0.0.1:3080. A saved model key does not require a restart.
  • Launch profiles in the architecture document: web, headless, sdk, sdk-minimal, and acp. The preview page names Standard, Code, Minimal, and Creator as working styles.
  • The safety notice says there has been no security audit. Prefer a disposable machine. The coding-agent alternative is a different page.

What the runtime is

DeepSeek Harness is DeepSeek AI’s open-source agent runtime: the project README defines it as an agent harness on an everything-is-a-plugin architecture, and the architecture document says the model adapter, the tool registry, the session log, and the agent loop are each replaceable from configuration.

The project README names the command dsh, points the plugin kernel at Cordis, and cites the paper A Programming Paradigm for Spatiotemporal Composability. Plugins contribute services, typed events, and reversible effects. Registrations unwind when a plugin unloads. The architecture document states there is no privileged core to patch.

The product page lists five jobs: everyday work with files, data, documents, and slides; coding; research with citations; background tasks; and plugins, including creation through chat in Creator mode. The same page marks Agent teams, Auto approval review, Scheduled tasks, and Voice input as experimental.

The category name for that wrapper is explained in the layer around the model. A reply box that never leaves the transcript is a different purchase, covered in agent versus chatbot.

PieceWhat the architecture document assigns
Session logAppend-only SessionEvent store at ctx.sessions. Model-visible context is derived from the log.
ToolsScoped registry and guarded pipeline at ctx.tools.
Agent handleInterface, live registry, and agent/* events at ctx.agents.
LoopDefault driver at ctx.agentLoop. The document says the loop itself is a plugin.
Model adapterMessage and stream vocabulary at ctx.llm.

How a local session starts

The README’s npm path is one command after Node.js is installed: npx @deepseek-ai/dsh web. That command starts the Web UI at http://127.0.0.1:3080 and opens the default browser for a local launch. An SSH launch only prints the host URL. --no-open keeps the server from opening a browser.

  1. Install Node.js for the npm path, or clone the repository and run pnpm install then pnpm run build.
  2. Start the Web UI with npx @deepseek-ai/dsh web or, from a built checkout, pnpm dsh web. The README says the built command uses existing artifacts and does not rebuild.
  3. Open Settings, then Models, enter a DeepSeek API key, and save. The Web UI guide says the route becomes usable without a server restart.
  4. Choose a workspace. The guide says the composer stays unavailable until a workspace is selected, even though the process uses its invoking directory as the default filesystem location.

The command name and the separate agent-registry package are unpacked in the command and agent registry. Parent cluster: Productivity & Workflows.

Plugins, profiles, and Creator mode

AI Edited A developer seating a replaceable plugin module into a local agent workspace
AI-edited illustration. View raw image.

The developer-preview page describes four working styles. Standard mode includes the full toolset. Code mode uses model-generated code to orchestrate multiple rounds of tool calls. Minimal mode keeps a shell tool and a file editor for benchmarking a model in a small environment. Creator mode inspects the current runtime, tests Cordis plugins in memory, and combines them into new modes.

The architecture document names the launch profiles separately: web, headless, sdk, sdk-minimal, and acp. dsh-base is the shared first layer of web, headless, sdk, and acp. sdk-minimal owns its own tree and does not apply that base layer. Layers apply in order: the profile’s bundles, its cordis.patch.yml, the home patch, then --patch overlays.

Creator mode, in the persistent-plugin guide, provides Plugin Manager and read-only runtime inspection. That guide’s MCP example asks the agent to write a configuration bundle that inserts @deepseek-ai/dsh-mcp-client. Skills, which are instructions, and MCP, which is a tool pipe, stay different layers on skills versus MCP tools.

Who should try it during preview

AI Edited An engineer reviewing an event log beside a sandboxed laptop in a disposable lab
AI-edited illustration. View raw image.

The safety notice is blunt. The software is experimental developer-preview code. It has not undergone a security audit. It can execute model-generated code and commands, load third-party plugins, and reach the network, processes, credentials, and files you expose. Sandboxes and approval prompts reduce risk and do not guarantee isolation. The notice says to prefer a disposable virtual machine, container, or dedicated environment, and it points at the MIT license’s lack of warranty.

That fits a team that wants to swap a model adapter or the loop from configuration, can tolerate compatibility-breaking changes, and will bring its own API key. It is a poor fit when the job is a stable coding agent across a terminal, an IDE, a desktop app, and a phone. That job split is the open runtime versus Claude Code. A cloud digital worker is a third shape, on Perplexity Computer versus other agents.

FAQ

Is the project open source?

The README’s license section points at MIT. Third-party dependency notices live in a separate file in the same repository.

Does saving an API key require a restart?

The Web UI guide says a saved DeepSeek key becomes usable immediately, without restarting the server.

Is the preview production-ready?

The safety notice says no. It says the project must not be treated as secure or production-ready.

What does Creator mode add?

The persistent-plugin guide says Creator mode provides Plugin Manager and read-only runtime inspection, and that plugin configuration belongs to the current profile.

Where do community plugins get discovered?

The README asks plugin authors to add the GitHub topic dsh-plugin so repositories can be found.

Sources

More guides like this appear when you search 'AI Agency Framework agent runtime' on Google.