DeepSeek open-sources its agent harness under MIT

DeepSeek has open sourced DeepSeek Harness, a Node.js agent runtime that picked up more than 33,000 GitHub stars within a few hours of Thursday's release. The project is available on GitHub as a developer preview under the MIT license, as reported by The New Stack.

At a glance

• The model adapter, the tool registry, the session log and the agent loop are each implemented as replaceable plugins, and the documentation states there is no privileged core to patch when extending the runtime.
• Four presets ship with the harness: Standard, Minimal with only bash and str_replace_editor, Code, which generates a TypeScript SDK so five round trips collapse into one call, and Creator.
• The provider catalog covers Anthropic, OpenAI, AWS Bedrock, Microsoft Azure, Google's Gemini Enterprise Agent Platform and DeepSeek's own endpoint, so teams can run the runtime without touching a DeepSeek model.

Why it matters. Coding agents have become standard issue at every major lab, which pushes differentiation toward how far a runtime can be taken apart. DeepSeek's bet appears to be that developers want substitution rather than configuration, and a harness whose agent loop and tool registry are swappable sits closer to a framework than to a product. Refusing external pull requests likely pushes that work outward, into plugins the project itself does not maintain.

The architecture rests on Cordis, described as a meta-framework for spatiotemporal composability. The underlying idea is that modern harnesses need dynamic composition, meaning components can be added or removed without affecting the rest of the system and can see how they depend on each other. A recent paper by three researchers from Peking University and DeepSeek sets out the approach.

Minimal ships two tools while Code mode collapses five round trips into one call

Four presets ship with the harness, covering the range from a fully equipped agent to a two-tool setup. Standard mode provides the full coding agent, with filesystem tools, shell access, web search, subagents and a plan mode. Minimal strips the toolset down to two entries, bash and str_replace_editor.

Code mode changes how tools reach the model: rather than exposing them as individual function calls, it generates a TypeScript SDK and has the model write a program against it, so a sequence that would otherwise take five round trips runs as a single call. Creator mode, meant for developers building custom presets, inherits Standard and adds runtime inspection, plugin experiments and preset-authoring guidance.

The session log, the tool registry and the agent loop are all replaceable plugins

Everything in the harness is a plugin, and DeepSeek applies that literally: the model adapter, the tool registry, the session log and the agent loop itself can each be swapped out. The documentation states there is no privileged core to patch, so extending the runtime means mounting a plugin beside the existing ones.

The harness keeps an append-only session log, and anything reaching a model request must be reconstructable from it. Resume, fork, replay, transcripts, telemetry and the web UI read from that one event stream, and any new model-visible input requires a new session event. Sandboxing is handled by the local backend, which wraps subprocesses in Linux Landlock through a Node addon DeepSeek wrote, macOS Seatbelt, or a Windows ACL restricted-token runner.

Two subagent providers hand work to Claude Code and Codex, both off by default

The provider catalog covers Anthropic, OpenAI, AWS Bedrock, Microsoft Azure and Google's Gemini Enterprise Agent Platform, which the documentation still refers to as Vertex, alongside DeepSeek's own endpoint. Custom OpenAI-compatible gateways can be added for other inference providers. The harness also includes an MCP client, Agent Client Protocol support, and can read AGENTS.md and CLAUDE.md files.

Two subagent providers delegate work directly to Anthropic's Claude Code and OpenAI's Codex, resolving each product's binary from the host PATH so the user supplies the install and the login. Both are switched off by default. DeepSeek also ships bridges that run an existing hooks.json from either product against the harness's own interception points, which the README describes as a compatibility path rather than the better design.

Coding agent harnesses have proliferated across labs. Among the Chinese labs, Alibaba's Qwen Code and ByteDance's Trae Agent both launched in mid-2025, Zhipu's ZCode arrived in July and Moonshot's Kimi CLI in October, and most of that competition is open source.

What's next. DeepSeek is not accepting external pull requests at present and directs would-be contributors to GitHub Discussions and to writing plugins instead. It adds that packages in the official repository are not inherently more important than community ones, and describes the repo as "an idea, an official showcase, and a source of inspiration, but not a mandate from us." No timeline has been given for when the developer preview becomes a stable release.