DeepSeek also has its own exclusive Coding Agent. The name is simple and crude, just called DeepSeek-TUI. The author calls himself a DeepSeek enthusiast who is a "brother of whales". Just now, the number of stars for this project suddenly started to increase sharply, reaching 2.3k, and it also appeared on the GitHub hot list.
This is a TUI programming tool written in Rust language. It runs in the terminal like Claude Code, but is optimized and adapted specifically for DeepSeek.
In order to promote his work to domestic netizens, the author Hunter Bown also specially used DeepSeek to translate the promotional tweets into Chinese.
When DeepSeek -After TUI became popular on GitHub as he wished, Hunter posted a picture and bluntly said that these were the craziest two days of his life, and expressed his gratitude to the "whale brothers" in Chinese.
“DeepSeek Edition Claude Code"
DeepSeek-TUI is a programming agent that lives in the terminal. To understand it more simply, it is the "DeepSeek version of Claude Code".
It was initiated by American independent developer Hunter Bown in January this year. It is written in Rust language and is open source under the MIT license. However, it has been tepid until the release of DeepSeek-V4 and Hunter’s Chinese promotion. This project began to explode during this May Day holiday.
Such as reading and writing files, executing Shell, searching web pages, managing Git, scheduling sub-Agents, connecting to MCP servers...these Claude It can basically do everything Code can do, and it also supports the installation of Skills, but it uses DeepSeek V4 to run behind it.
The entire tool, from design logic to functional details, revolves around the features of DeepSeek.
The most direct one is the thinking chain.
DeepSeek-TUI streams the model's reasoning process directly to the terminal - how the model analyzed the problem, which path it took, and whether it changed its mind midway, all are visible in real time.
Then there’s context. V4 supports a context window of 1 million tokens, which is fully used by the project by default. You don’t have to worry about memory gaps when running complex tasks from beginning to end.
When the context is almost full, TUI will automatically compress the content, or it can be triggered manually/compact.
The compression strategy specifically takes into account DeepSeek's prefix caching mechanism - try to keep the previous stable part so that the cache can continue to hit.
This TUI also has a design called RLM, and the idea is "very DeepSeek" - since DeepSeek is cheap enough to be used in large quantities, this tool directly uses this feature.
In RLM mode, a main model directs up to 16 V4 Flash subtasks to run simultaneously for batch analysis or task disassembly. The output price of Flash is about one-third of that of Pro. By giving it subtasks that do not require strong reasoning, the overall cost can be reduced a lot.
Model switching has also been specially handled. In addition to the official DeepSeek API, it also supports NVIDIA NIM, Fireworks, and self-hosted SGLang paths.
There are three operating modes:
Plan is a read-only exploration, I will give you a plan first;
Agent is the default file, and every tool call requires you to nod;
YOLO, as the name suggests, is fully automatic, so just open it if you don’t want to be interrupted. Sessions can be saved and restored, and the workspace has an independent Git snapshot. Rolling back in rounds will not affect the original warehouse, so you don’t panic if it overturns.
However, one thing to note is that if too many sub-agents are opened, the cache hit rate is difficult to guarantee.
You should know that the price of a missed token is 10 times that of a hit. There is a round-by-round cost display on the project interface. It is recommended to pay attention to it if you are running a long session, and don’t be surprised when the bill is finished.
Installation, Linux, macOS, and Windows have precompiled binaries, just npm install -g deepseek-tui can be done with one command.
In addition, the author has also prepared a special Chinese version of README document and special configuration path for domestic users, which supports TUNA Cargo mirroring. The release package can also be hosted on Alibaba Cloud OSS or Tencent Cloud COS. The
project was established on January 19. It has been less than 4 months since it has been iterated to v0.8.8 and 37 versions have been released. The pace is not slow.
Judging from the update records, it is roughly divided into several stages.
The early version mainly built the skeleton-tool calling, session management, and basic Git snapshots. Getting the Agent running is the first priority.
v0.7.x stage began to polish the details, adding multi-language interface support (v0.7.6), TUI prompts, help text, and status bars in Chinese and other languages began to be localized. This is also a step to adapt to domestic users.
v0.8.x is the main spindle of recent versions, focusing on stability and experience polishing.
v0.8.2 specifically fixes file handle leaks in long sessions Leakage problem;
v0.8.6/v0.8.7 up A number of interactive functions have been added, including displaying countdown retry banners when current limiting or server errors are reported, input history search, and running message queue visualization;
v0.8.8 has made a round of closure on this basis, and at the same time, Linux ARM64 precompiled binaries added.
Looking at the overall rhythm, this iteration path has intensive feature updates, but each version basically has clear problems to be solved.
"Musician who loves science"
In fact, Hunter has always been an avid fan of DeepSeek. Since the release of V4, he has sent many tweets praising it.
At the same time, he also likes other Chinese models and has participated in Xiaomi’s One Billion Token Creator Incentive Program.
Hunter Bown’s starting point was actually music. He once aspired to be a band conductor.
He first studied music education at the University of North Texas. After graduation, he continued his education and obtained a master's degree in music education from Southern Methodist University.
After graduating with a master’s degree, Hunter worked as a band conductor for three years as he wished.
Later, he got an MBA from the University of Texas at Dallas, and then returned to his previous alma mater, SMU, and entered the law school specializing in patent law.
As for coding, it is even more of a "midway monk" choice.
But this "halfway" is not a career change, it is more like several lines finally coming together.
When he was studying vocal music science, he came across a concept called "missing fundamental" - the human ear can reconstruct a pitch that does not physically exist from overtones.
He later discovered that this directly corresponds to information theory. You don't need to give all the information explicitly, the system itself will complete it.
This intuition from music became a key for him to understand the AI system.
Last year, he founded a studio for himself called Shannon Labs, which is positioned as "the next Bell Labs in the AGI era."
DeepSeek-TUI is just one of many research projects for him. There are 65 public repositories on his GitHub, including the same terminal Agent NeMoCode for NVIDIA Nemotron, as well as the MLX kernel toolkit and so on.
The projects under Shannon Labs have a wider span.
Hegelion is a dialectical reasoning engine that walks It is the circular logic of "thesis → antithesis → synthesis";
Al eph is an MCP server that focuses on high-capacity context with zero token cost;
Heliosinger converts solar wind data into sound in real time, spanning from AI infrastructure to space acoustics.
He also built three software architectures (SCU, Driftlock, Hegelion) and a hardware solution (Driftlock Choir), in his view, these are put together to build infrastructure for the AGI era.
can put these directions together, and it is also related to his family story.
His great-grandfather, Ralph Bown Sr., was the vice president of research at Bell Labs and a radio pioneer. In his spare time, he liked to make homemade wax cylinders and go to Carnegie Hall to record.
Hunter realized in the patent law class that he was on a journey A road that intersects with this ancestor -
brings the musician's way of perception into technical research to discover those "ideas that have been ignored because researchers without this background".
He compared himself with his great-grandfather on his personal website, "He is a scientist and loves music; I am a musician and love science."
One More Thing
In the contributor list of DeepSeek-TUI, we can also see some familiar shadows.
It includes a series of AI models such as Claude, Gemini, and Qwen, and programming tools such as Cursor and GitHub Copilot.
Detailed records show that most of the code was submitted directly by Hunter, and more than 150 commits were made by Claude. In addition, some real contributors submitted a small number of commits.
A programmer who became a monk on the way, uses AI-assisted programming to write an auxiliary programming framework for AI. This workflow is also closed-loop (manual dog-head).
GitHub address:
https://github.com/Hmbown/DeepSeek-TUI