“DeepSeek Version of Claude Code” Gains 2.3k Stars on GitHub
DeepSeek now has its own dedicated CodingAgent called DeepSeek-TUI, created by a DeepSeek enthusiast who calls himself a “whale brother.” The project recently experienced a surge in stars, reaching 2.3k and topping GitHub's trending list.
This is a TUI (Text User Interface) programming tool written in Rust, running in the terminal like Claude Code, but specifically optimized and adapted for DeepSeek.
To promote his work to domestic users, author Hunter Bown used DeepSeek to translate the promotional post into Chinese.
After DeepSeek-TUI went viral on GitHub as he hoped, Hunter posted a picture, saying these past two days have been the craziest of his life, and expressed his gratitude to the “whale brothers” in Chinese.
“DeepSeek Version of Claude Code”
DeepSeek-TUI is essentially a programming Agent living in the terminal – or, simply put, the “DeepSeek version of Claude Code.”
It was initiated in January of this year by American independent developer Hunter Bown, written in Rust, and open-sourced under the MIT license. However, it remained relatively unknown until the release of DeepSeek-V4 and Hunter’s Chinese promotion, after which the project exploded in popularity during the May Day holiday.
Like reading and writing files, executing Shell commands, searching the web, managing Git, scheduling sub-agents, and connecting to MCP servers – it can basically do everything Claude Code can do. It also supports installing Skills, but now runs on DeepSeek V4 in the background.
The entire tool, from its design logic to its functional details, revolves around the characteristics of DeepSeek.
The most direct aspect is the chain of thought.
DeepSeek-TUI streams the model’s reasoning process directly to the terminal – how the model analyzes the problem, which path it takes, and whether it changes its mind mid-way, all are visible in real-time.
Then there’s the context window. V4 supports a 1 million token context window, and the project defaults to using it fully, so you don’t have to worry about memory loss when running complex tasks from start to finish.
When the context is nearly full, the TUI automatically compresses the content, or you can manually trigger /compact.
The compression strategy specifically considers DeepSeek’s prefix caching mechanism – preserving the stable front part as much as possible to allow the cache to continue hitting.
This TUI also has a design called RLM, with an approach that is “very DeepSeek” – since DeepSeek is cheap enough to be stacked in quantity, this tool directly incorporates that feature.
In RLM mode, a master model directs up to 16 V4 Flash sub-tasks to run simultaneously, for batch analysis or task decomposition. Flash output is approximately one-third the price of Pro, and delegating sub-tasks that don’t require strong reasoning to it can significantly reduce overall costs.
Model switching has also been specifically handled, supporting not only the official DeepSeek API, but also NVIDIA NIM, Fireworks, and self-hosted SGLang.
There are three operating modes:
Plan is read-only exploration, first giving you a proposed solution;
Agent is the default mode, requiring your approval for each tool call;
YOLO, as the name suggests, is fully automatic, allowing everything to proceed without interruption. Sessions can be saved and restored, the workspace has independent Git snapshots as a fallback, rolling back rounds without affecting the original repository, so you don’t have to worry about failures.
However, it’s important to note that opening too many sub-agents can make it difficult to guarantee cache hit rates.
Remember, the price of an unhit token is 10 times that of a hit token. The project interface displays the cost per round, so it’s recommended to pay attention when running long sessions, so you don’t get a shocking bill.
For installation, pre-compiled binaries are available for Linux, macOS, and Windows. A single command, `npm install -g deepseek-tui`, does the trick.
In addition, the author has prepared a dedicated Chinese README document and dedicated configuration paths for domestic users, supporting TUNA Cargo mirrors, and release packages can be hosted on Alibaba Cloud OSS or Tencent Cloud COS.
The project was created on January 19th and has already iterated to v0.8.8 in less than 4 months, with 37 releases, at a steady pace.
From the update log, it can be roughly divided into several stages.
Early versions were mainly focused on building the framework – tool calls, session management, basic Git snapshots, and getting the Agent running was the top priority.
The v0.7.x stage began to refine the details, adding multi-language interface support (v0.7.6), TUI prompts, help text, and status bar localization in Chinese and other languages, which is a step towards adapting to domestic users.
v0.8.x is the main focus of recent versions, with a focus on stability and experience refinement.
v0.8.2 specifically fixed a file handle leak issue in long sessions;
v0.8.6/v0.8.7 added a batch of interactive features, including a countdown retry banner when rate-limited or the server errors, input history search, and a visualization of the running message queue;
v0.8.8 built on this with a round of cleanup, and also supplemented the pre-compiled binary for Linux ARM64.
Overall, the iteration path shows dense feature updates, but each version basically has a clear problem to solve.
“A Scientific Musician”
In fact, Hunter has always been a passionate DeepSeek fan, and has posted many tweets praising it since V4 came out.
He also enjoys other Chinese models and participated in Xiaomi’s trillion-token creator incentive program.
Hunter Bown’s starting point, however, was music. He once aspired to be a band conductor.
He first studied music education at the University of North Texas, and continued his studies after graduation, earning a Master of Music Education from Southern Methodist University.
After graduating with his master’s degree, Hunter fulfilled his dream of becoming a band conductor for 3 years.
Later, he earned an MBA from the University of Texas at Dallas, and then returned to his alma mater SMU, entering law school specializing in patent law.
As for coding, he chose it as a “late bloomer.”
But this “late bloomer” wasn’t a career change, but rather a convergence of several lines.
While studying vocal science, he encountered a concept called “missing fundamental” – the human ear can reconstruct a physically non-existent pitch from harmonics.
He later discovered that this directly corresponds to information theory, you don’t need to explicitly give all the information, the system itself will fill it in.
This intuition from music became a key to his understanding of AI systems.
Last year, he founded a studio called Shannon Labs, positioning it as “the next Bell Labs of the AGI era.”
DeepSeek-TUI is just one of many research projects here. He has 65 public repositories on GitHub, including NeMoCode, a similar terminal Agent for NVIDIA Nemotron, and the MLX kernel toolkit, among others.
Shannon Labs’ projects span a wider range.
Hegelion is a dialectical reasoning engine, following the cycle of “thesis → antithesis → synthesis”;
Aleph is an MCP server, focusing on zero-token-cost large-capacity context;
Heliosinger converts solar wind data into sound in real-time, crossing the boundary from AI infrastructure to space acoustics.
He also built three software architectures (SCU, Driftlock, Hegelion) and a hardware solution (Driftlock Choir). He believes that these combined are building infrastructure for the AGI era.
His family history is also closely related to his ability to put these directions together.
His great-grandfather, Ralph Bown Sr., was a research vice president at Bell Labs and a pioneer in radio, and in his spare time, he liked to make wax cylinders and record at Carnegie Hall.
Hunter realized in patent law class that he was walking a path that converged with his ancestor –
Bringing the perceptual way of a musician into technical research, to discover those “ideas that were overlooked because researchers lacked this background.”
He compared himself to his great-grandfather on his personal website, “He was a scientist who loved music; I am a musician who loves science.”
One More Thing
In the contributor list of DeepSeek-TUI, you can also see some familiar faces.
These include a series of AI models such as Claude, Gemini, and Qwen, as well as programming tools such as Cursor and GitHub Copilot.
Detailed records show that most of the code was submitted directly by Hunter, and Claude made over 150 commits, along with some minor commits from other human contributors.
A programmer who started out late, using AI to assist programming to write an AI-assisted programming framework – this workflow is a closed loop (manual dog head).
GitHub address: