DeepSeek's Claude Code Rival Is MIT. It Also Runs Claude.

TL;DR

On August 13, DeepSeek open-sourced DeepSeek Harness (the CLI is dsh), an agent harness released under the MIT license as a v0.1 developer preview. The pitch fits in four words: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, the agent loop, the scheduler and the UI are all swappable components you configure rather than fork. The repo went from created to more than 55,000 stars in roughly a day, and the provider catalog it ships with includes first-class adapters for Anthropic and OpenAI.

What actually shipped

This is not a model announcement. It is a runnable stack, written in TypeScript, that you install with one command:

npx @deepseek-ai/dsh web

That starts a web UI on http://127.0.0.1:3080. Node 22.19 or later on the 22 line, or anything from 24 up, is the stated requirement. You can also clone the repo and build it with pnpm. Per the GitHub API, deepseek-ai/deepseek-harness was created on August 13 and, checked the following morning UTC, carried 55,035 stars and 4,481 forks.

For contrast, openai/codex is Apache 2.0, written in Rust, and took sixteen months to reach 105,770 stars. The anthropics/claude-code repo sits at 141,380 stars with no license field at all, because it is a tracker and docs repo rather than the product's source. DeepSeek shipped the source.

Everything is a plugin, and the plumbing is a surprise

The architecture rests on Cordis, which describes itself as a "Meta-Framework of Spatiotemporal Composability." Cordis does one job: load plugins, unload plugins, resolve the dependencies between them. Every concrete part of the harness, including the model adapter, the tool registry, the session log and the agent loop itself, is a separate Cordis plugin. The docs put the payoff plainly: you "can select, swap, or extend any capability in configuration without changing the DeepSeek Harness source code."

Most agent frameworks are a car where you get to change the tires and the stereo. This one is a car where the engine, the chassis and the steering column are all bolt-on parts, and Cordis is nothing but the bolt pattern everything shares.

The genuinely odd detail: Cordis is not a DeepSeek invention. It has been public since May 2022, it is MIT, it is TypeScript, and it carries fewer than 2,000 stars. A niche community plugin framework spent four years quietly existing and then woke up one Thursday as the substrate under the most-starred repo of the week.

Four modes, and one of them tells you who this is for

The harness ships four run modes, each loading a different default plugin set:

• Standard: "Full coding agent with file editing, shell, file and web search, skills, planning, goals, subagents, and workflows."
• Code: everything in Standard, but tools are exposed through a Code Mode SDK "so the model can combine multi-step operations in one TypeScript program" instead of a chain of separate tool calls.
• Minimal: "Two-tool coding agent with persistent bash and str_replace_editor."
• Creator: Standard plus runtime inspection, in-memory plugin experiments and preset-authoring guidance.

Minimal mode is the tell. A two-tool agent with a persistent shell and a string-replace editor is not a product decision, it is an evaluation rig: the smallest scaffold that still lets a model do real work, so you can measure the model rather than the harness around it. That is what a lab builds for itself. DeepSeek just shipped its internal measuring stick along with the product.

The provider list is the actual news

Open Settings, choose Add provider, and the installed catalog offers Anthropic, OpenAI, Bedrock, Vertex and Azure, plus a Codex entry that authenticates over OAuth, alongside DeepSeek's own endpoint. Anything missing goes in as a custom provider: lowercase provider ID, base URL, API protocol, credential, at least one model. Point it at vLLM on your own box and it does not care.

This is not charity, it is strategy. Model loyalty is a weekly decision now, and everyone in this market has watched it churn. Harness loyalty is sticky: it holds your presets, your skills, your session history and your muscle memory. DeepSeek is betting that owning where you work matters more than owning what you call, which is precisely the bet Anthropic made with Claude Code and OpenAI made with Codex, except DeepSeek is making it with the source in the open and a permissive license attached.

The session log is the part builders should read twice

From the docs: "Everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection." A Trajectory view lets you inspect that log by source, and resume, fork, search or replay it.

If you have ever tried to work out why an agent did something stupid four hours into a run, you know why this matters. Most harnesses hand you a reconstructed transcript, which is a summary of what happened written by the thing that did it. An append-only log of every injection, including the context you did not know was being added, is a different class of artifact. Forking a session at the exact turn where it went sideways, rather than starting over, is the feature that will quietly change how you debug.

Now the caveats, which are real

The README is blunt about status: this is a developer preview, and it warns in capitals that there will be compatibility-breaking changes. Treat any config you write this week as disposable.

The Hacker News thread hit 560 points and 246 comments, and the criticism is worth weighing. Multiple commenters objected to the install footprint, with one reporting a 47MB download that expanded to roughly 1.5GB after building. Others pushed back on TypeScript and Node for a CLI tool at all, and on npm as a distribution channel for something that will hold your API keys and run shell commands. A maintainer replied in the thread that it is an early developer preview and to expect rough edges, which is the correct answer and also not a fix.

The plugin ecosystem is the other thing to watch carefully. The dsh-plugin GitHub topic already lists 808 repositories. That is remarkable adoption and an unvetted supply chain arriving at the same time, in a system explicitly designed to let third-party code replace the sandbox and the tool registry. Read what you install.

Also note what did not ship: no benchmark table, no eval numbers, no claimed win over any other harness. For a company that usually leads with a scorecard, shipping a runtime with zero performance claims is a choice.

What to actually do

• Try Minimal mode first if you evaluate models. Two tools and a persistent shell is the cleanest apples-to-apples rig you will get without building one yourself.
• Point it at your own endpoint. The custom provider path takes a base URL and a protocol, so a local server behind an OpenAI-compatible gateway works without touching DeepSeek's API at all.
• Pin your versions and audit plugins. Breaking changes are promised, and 808 topic-tagged repos on day two is not a curated registry.
• Do not put production credentials in a v0.1 preview that runs shell commands and loads third-party plugins. Sandbox it or use a throwaway key.

Key Takeaways

• DeepSeek Harness v0.1 shipped August 13 under the MIT license, in TypeScript, as a developer preview with an explicit breaking-changes warning.
• The repo hit 55,035 stars and 4,481 forks in about a day, against 105,770 for the Apache-licensed Codex CLI over sixteen months.
• Everything is a Cordis plugin: models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI, all swappable in config without forking.
• The shipped provider catalog includes Anthropic, OpenAI, Bedrock, Vertex, Azure and Codex, so the harness is deliberately not a funnel to DeepSeek's API.
• Minimal mode, a two-tool agent with persistent bash and a string-replace editor, is a model-evaluation rig hiding in plain sight.
• The append-only session log with fork and replay is the most useful feature for anyone who debugs long agent runs.

Sources: DeepSeek Harness official page, deepseek-ai/deepseek-harness on GitHub, DeepSeek Harness provider docs, cordiverse/cordis, openai/codex, anthropics/claude-code, Hacker News discussion, VentureBeat, The New Stack