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

AI Edited DeepSeek Harness vs Claude Code: two developers comparing an open plugin chassis with a finished coding laptop
AI-edited illustration. View raw image.

Key takeaways

  • README: MIT license, developer preview, compatibility-breaking changes. Safety notice: not production-ready.
  • Claude Code overview: terminal, IDE, desktop, and browser on one engine. Most surfaces need a Claude subscription or Console account.
  • Open runtime extension is configuration, including the loop. Claude Code’s documented extension surface is instructions, skills, hooks, MCP, subagents, and the Agent SDK.
  • Choose the preview to own the runtime. Choose Claude Code to ship in a codebase today.

The job split

DeepSeek Harness vs Claude Code is a job split: one side is an MIT-licensed developer preview you compose from plugins, and the other is Anthropic’s coding agent you install and use across several surfaces.

The repository README states the license is MIT, marks the project a developer preview, and warns that compatibility-breaking changes will happen. Anthropic’s Claude Code overview describes an agentic coding tool that reads a codebase, edits files, runs commands, and connects to development tools, available in the terminal, IDE, desktop app, and browser.

Product shape for the open side is on the open agent runtime guide. This page stays on the choice.

Open runtimeClaude Code
License and statusMIT. Developer preview. Safety notice: no security audit, not production-ready.Anthropic product. Overview does not grant an MIT license. Native installs update in the background.
First commandnpx @deepseek-ai/dsh web on 127.0.0.1:3080.Install, then claude inside a project directory.
Documented surfacesProfiles web, headless, sdk, sdk-minimal, acp, plus an Electron desktop app.Terminal, VS Code, JetBrains, desktop app, browser, CI, Slack, Chrome.
Model routeSettings key. Provider guide lists built-in ids and custom OpenAI-compatible or Anthropic-protocol endpoints.Most surfaces need a Claude subscription or Console account. Terminal, VS Code, and JetBrains support third-party providers.
Extension surface the vendor documentsPlugins, including the agent loop, via configuration.CLAUDE.md, skills, hooks, MCP, subagents, background agents, Agent SDK.

Where each one runs

AI Edited A developer choosing between a local browser runtime and a coding session that moves across a phone and a laptop
AI-edited illustration. View raw image.

The architecture document launches Node applications through named profiles. web adds the browser application. headless is a one-shot runner with no server. sdk adds a JSON-RPC server. acp adds an automation-only server. The desktop app ships an Electron host that loads the same Web application and keeps a reserved desktop profile. The default Web UI port in the README is 3080 on localhost.

Claude Code’s overview lists a terminal CLI, a VS Code extension with inline diffs, a JetBrains plugin, a desktop app that can run sessions side by side, and a browser app at claude.ai/code with no local setup. The same page documents moving a session: Remote Control, claude --teleport into a terminal, and /desktop from the terminal into the desktop app on macOS and x64 Windows. Desktop scheduled tasks run on the machine. Cloud routines keep running when the computer is off.

Editor products that already sit in this cluster are compared in Cursor versus Copilot. The open runtime is not an editor fork and not an IDE overlay.

How you extend each one

AI Edited An engineer attaching skills and hooks to a finished coding bench while a second chassis has every board removable
AI-edited illustration. View raw image.

Anthropic’s overview says CLAUDE.md is read at the start of every session, skills package repeatable workflows such as a review or a deploy, and hooks run shell commands before or after actions. MCP connects Claude Code to external data. A lead agent can spawn subagents. The Agent SDK builds custom agents on Claude Code’s tools, with control over orchestration, tool access, and permissions. The overview does not document replacing the agent loop as a configuration row.

The architecture document says the opposite arrangement for the open runtime: the model adapter, the tool registry, the session log, and the agent loop are plugins, so each is replaceable from configuration. A provider guide lets you add Anthropic, OpenAI, Moonshot (Kimi), Z.ai (GLM), or a custom endpoint that speaks OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages. Instruction packs versus tool pipes are separated on skills versus MCP tools.

Which job to assign

Pick Claude Code when the work is today’s codebase: the overview’s install paths, the shared engine across terminal and IDE and desktop and browser, and a paid desktop subscription are the product Anthropic documents. The overview says most surfaces require a Claude subscription or an Anthropic Console account.

Pick the open runtime when the work is the runtime: swap a provider, keep an append-only session log, or replace the loop without forking the repository. The safety notice still applies. Budget a disposable environment. Do not treat the preview as the sole control for untrusted workloads.

A cloud worker that deploys subagents across connectors is a third purchase, on Perplexity Computer versus other agents. Hub: Legal & Professional AI Tools. Silo: AI Tools & Software.

FAQ

Can the open runtime call Claude models?

The provider guide lists an Anthropic provider id and an Anthropic Messages protocol for custom endpoints. That is a model route inside the open runtime, not a switch into Claude Code’s product.

Can Claude Code use a non-Claude model?

The overview says the Terminal CLI, VS Code, and JetBrains support third-party providers. It says most surfaces require a Claude subscription or Console account, and the desktop app requires a paid subscription.

Which one is production-ready?

The open runtime’s safety notice says it is not. Claude Code’s overview documents native background updates and several shipping surfaces. Maturity still depends on the surface you install.

Do they cost the same?

The open runtime’s software license is MIT. You still pay the model provider whose key you save. Claude Code’s overview ties most surfaces to a subscription or Console account and states that the desktop app requires a paid subscription. This page does not invent dollar prices.

Sources

More guides like this appear when you search 'AI Agency Framework coding agents' on Google.