GitHub Trending Top Spot: An AI Coding Agent in the Terminal
Today’s GitHub Trending chart is dominated by an open-source project — DeepSeek-TUI, which gained 2,434 stars in a single day, pushing its total past 11.3k and making it the hottest AI tool project on GitHub right now.
What Is It?
DeepSeek-TUI is an AI coding agent that runs in your terminal, built in Rust and directly interfacing with DeepSeek V4 series models (deepseek-v4-pro / deepseek-v4-flash). It’s not just a code completion tool — it’s a full-featured programming assistant that can read files, write code, execute shell commands, search the web, manage Git, and coordinate sub-agents.
In one sentence: Run a Claude Code / Cursor-level AI agent in your terminal, completely free, open-source, and with zero backend service dependencies.
Core Capabilities
Auto Mode: Let the Agent Pick Its Own Model
--model auto is DeepSeek-TUI’s most distinctive feature. Instead of manually selecting a model, the agent automatically routes each task to the appropriate model and reasoning level based on complexity:
- Simple questions →
deepseek-v4-flash, reasoning off, fast response - Coding, debugging, architecture design → automatically upgrades to
deepseek-v4-prowithhighormaxreasoning
The routing decision is made by a lightweight deepseek-v4-flash call. The upstream API always receives a concrete model ID — it never sees "auto".
Streaming Reasoning Blocks: Watch AI Think in Real Time
DeepSeek V4 supports thinking mode, and DeepSeek-TUI streams the model’s reasoning process live in the terminal. You can see how the AI thinks step by step before writing code — far more valuable than just seeing the final output.
Three Work Modes
| Mode | Behavior |
|---|---|
| Plan 🔍 | Read-only investigation — the model explores the codebase and proposes a plan without modifying files |
| Agent 🤖 | Interactive mode with multi-step tool use, each operation requiring approval |
| YOLO ⚡ | Auto-approve all tool calls, ideal for trusted workspaces |
Full Toolchain
File operations, shell execution, Git management, web search, patch application, sub-agent coordination, MCP protocol support — DeepSeek-TUI’s tool registry covers nearly every development scenario. Notable highlights:
- MCP Protocol: Connect to Model Context Protocol servers to extend the tool ecosystem
- LSP Diagnostics: After every edit, diagnostics are automatically injected via rust-analyzer, pyright, typescript-language-server, etc., feeding real compiler feedback into the AI’s next reasoning step
- Skills System: Composable, installable skill packs — one-click install from GitHub community, no backend service needed
- Session Save/Resume: Long-running tasks can be checkpointed and resumed — no fear of closing the terminal
- Workspace Rollback: Side-git-based edit snapshots,
/restoreto revert without touching your repo’s.git
Extremely Low Cost
DeepSeek V4’s pricing is crushing the current market:
deepseek-v4-flash: Input $0.14/1M tokens (cache hit $0.0028), Output $0.28/1Mdeepseek-v4-pro: Input $0.435/1M tokens (cache hit $0.003625), Output $0.87/1M
The Pro model currently has a 75% limited-time discount, valid until May 31, 2026. DeepSeek-TUI includes real-time cost tracking — token usage and costs per turn are visible at a glance.
Multi-Model Support
Not limited to the official DeepSeek API — also supports:
- NVIDIA NIM: via
--provider nvidia-nim - Fireworks: via
--provider fireworks - Self-hosted SGLang: just set the
SGLANG_BASE_URLenvironment variable - Self-hosted vLLM: just set the
VLLM_BASE_URLenvironment variable
This means you can deploy DeepSeek V4 locally and use the agent completely offline.
Zed Editor Integration
DeepSeek-TUI supports ACP (Agent Client Protocol) and can be used directly as a custom agent in Zed editor:
{
"agent_servers": {
"DeepSeek": {
"type": "custom",
"command": "deepseek",
"args": ["serve", "--acp"]
}
}
}v0.8.14 Update Highlights
The latest v0.8.14 release is a stabilization version with key improvements:
- Auto mode restored:
--model auto,/model auto, config items, and sub-agents all support automatic routing - Per-turn cost accounting fix: V4 reasoning tokens are now correctly counted as billable output
- First-run setup optimization: Missing config files now automatically guide users through API key setup
- vLLM Provider support: Self-hosted vLLM endpoints can connect directly
Installation
# npm (easiest, auto-downloads prebuilt binaries)
npm install -g deepseek-tui
# Cargo (for Rust users)
cargo install deepseek-tui-cli --locked # deepseek entry point
cargo install deepseek-tui --locked # deepseek-tui interface
# Homebrew (macOS)
brew tap Hmbown/deepseek-tui
brew install deepseek-tui
# Mirror-accelerated (China)
npm install -g deepseek-tui --registry=https://registry.npmmirror.comWhy It Matters
DeepSeek-TUI’s explosive growth signals several deeper trends:
- Terminal agents are a real need: Developers don’t want to leave the terminal for AI coding. DeepSeek-TUI proves the “AI agent in terminal” paradigm has a massive market
- DeepSeek ecosystem is maturing fast: The third-party tool ecosystem around DeepSeek models is expanding rapidly — from TUI to MCP to Skills, its completeness rivals the OpenAI ecosystem
- Open-source alternatives are rising: As Cursor and Claude Code get more expensive, a free, open-source, Rust-built high-performance agent is becoming developers’ go-to choice
- China developer-friendly: Built-in Chinese localization, npm/Cargo domestic mirror support, and DeepSeek API’s low cost — this is practically tailor-made for Chinese developers
Conclusion
DeepSeek-TUI is not another wrapped chat tool — it’s a genuine terminal coding agent. Its 11.3k GitHub stars and 2,434-star single-day growth show the market is voting with its feet: the combination of terminal AI agents, open-source, and DeepSeek models is becoming the new standard for developer tools in 2026.
If you’re still using Claude Code or Cursor, DeepSeek-TUI is well worth trying. One command — npm install -g deepseek-tui — gets you started, and DeepSeek API pricing may be just one-tenth of your current costs.