今天 GitHub Trending 第一是 DeepSeek-TUI,2434 星------一个在终端里运行的 AI 编程代理。我花了 15 秒装好,配了 API Key,让它生成了代码,也踩了几个坑。这篇文章是我从零到跑通的全过程记录。
环境
Windows 11,Node.js v22.22.1,DeepSeek API Key(platform.deepseek.com 申请)。
第一步:安装(15 秒)
bash
npm install -g deepseek-tui --registry=https://registry.npmmirror.com
npmmirror 镜像对国内用户更快。输出只有一行:
go
added 1 package in 15s
npm 包本质是个安装器------它自动从 GitHub Releases 拉匹配的 Rust 预编译二进制。装完 deepseek 和 deepseek-tui 两个命令直接可用。
bash
$ deepseek --version
deepseek 0.8.14
第二步:跑诊断
任何新工具我习惯先跑诊断命令,看看它到底有什么功能、缺什么配置。
bash
$ deepseek doctor
完整输出比较长,我挑关键的说:
yaml
Version: deepseek-tui 0.8.14
Configuration:
! config.toml not found (using defaults/env)
API Keys:
· deepseek: env=no, config=no
... (6 个 provider 全未配置)
MCP Servers:
✓ MCP feature flag enabled
Skills:
· skills dir not found (3 个位置都未创建)
Platform:
OS: windows, Arch: x86_64
✓ sandbox available
诊断暴露了几个信息:API Key 没配、Skills 目录没建、MCP 和沙箱功能已启用但没配置。这些卡点正好是接下来要解决的。
第三步:配 API Key
三种方式,我选了最省事的------保存到配置文件,一次配置全局生效:
bash
$ deepseek auth set --provider deepseek
# 粘贴 API Key,回车。输入时不回显。
saved API key for deepseek to ~\.deepseek\config.toml
另外两种方式:环境变量 export DEEPSEEK_API_KEY=sk-xxx,或者手动写 ~\.deepseek\config.toml 加 api_key = "sk-xxx"。
配完再跑一次 doctor,变化明显:
ini
Configuration:
✓ config.toml found
API Keys:
✓ deepseek: env=yes, config=yes
✓ active provider key resolved from config.toml
API Connectivity:
✓ API connection successful (model: deepseek-v4-pro)
第四步:看它支持什么模型
bash
$ deepseek models
Available models (default: deepseek-v4-pro)
deepseek-v4-flash (deepseek)
* deepseek-v4-pro (deepseek)
默认是 Pro,还有一个 Flash------Pro 做深度推理,Flash 做快速任务。在交互模式里可以用 /model auto 让它自动选。
第五步:查看功能开关
bash
$ deepseek features list
feature stage enabled
shell_tool stable true
subagents experimental true
web_search experimental true
apply_patch experimental true
mcp experimental true
exec_policy experimental true
6 个 feature,1 个稳定版(shell_tool),5 个实验版。subagents 支持子 Agent 协调,mcp 支持 Model Context Protocol,web_search 支持联网搜索。这已经不是"对话工具"了------是一个完整的 Agent 运行环境。
第六步:让它写代码
非交互模式用 deepseek exec,适合在脚本里调用:
bash
$ deepseek exec "Create a Python file that prints a greeting and calculates fibonacci numbers"
输出:
python
# hello_agent.py
def fibonacci(n: int) -> int:
"""Return the n-th Fibonacci number (0-indexed)."""
if n < 0:
raise ValueError("Input must be a non-negative integer.")
if n == 0:
return 0
if n == 1:
return 1
a, b = 0, 1
for _ in range(2, n + 1):
a, b = b, a + b
return b
if __name__ == "__main__":
print("DeepSeek TUI Agent says hello!")
for i in range(10):
print(f"fibonacci({i}) = {fibonacci(i)}")
代码可以直接用------有类型标注、有边界检查、有迭代而非递归(避免大数栈溢出)。这说明 DeepSeek V4 Pro 的代码生成品质是可以的。
第七步:项目初始化
bash
$ deepseek init
✓ Created AGENTS.md
生成的是一个模板文件,包含构建命令、架构概述、数据流等占位区域。编辑它之后,Agent 在项目里运行时会自动加载这些指令------相当于给 Agent 配置了项目上下文。
第八步:让它解释概念
bash
$ deepseek exec "DeepSeek-V4-Pro 和 Flash 在推理能力上有什么区别?用一句话回答"
输出:
DeepSeek-V4-Pro excels at deeper, multi-step logical reasoning for complex problems, while DeepSeek-V4-Flash sacrifices some reasoning depth for significantly faster, more efficient inference on simpler tasks.
一句话讲清楚:Pro 做深度推理,Flash 追求速度。
踩到的坑
坑 1:exec 模式不写文件。 deepseek exec 生成代码是在响应里输出的,不会自动创建文件。如果要让它真正操作文件,得进交互模式(直接跑 deepseek 进 TUI)------交互模式里 Agent 模式需要逐次审批文件操作,YOLO 模式自动放行但不建议在正式项目里用。
坑 2:Windows 下没有 Rust 工具链,doctor 显示 rust: unknown。 不影响使用,但如果要 cargo 安装需要先装 Rust。
坑 3:国内访问 DeepSeek API 偶尔超时。 可以在 config.toml 里把 base_url 改成 https://api.deepseeki.com(国内镜像)。
跟同类工具的比较
跟 Claude Code 比:DeepSeek-TUI 的优势是绑定 DeepSeek 模型------1M token 上下文 + 流式推理,处理大代码库能力强。劣势是生态没 Claude Code 成熟,Skills 和 MCP 的社区内容少得多。
跟 Codex CLI 比:Codex 更专注 OpenAI 生态,DeepSeek-TUI 支持多 Provider 切换(deepseek/nvidia-nim/openrouter/fireworks/sglang),选模型更灵活。
结论
DeepSeek-TUI 0.8.14 是一个工具链完整的终端 AI 编程代理。安装零门槛(npm 一行命令),诊断报告清晰(doctor 直接告诉你缺什么),API Key 配置简单(三种方式),代码生成品质达到了可用的标准。
适合谁:已经在用 DeepSeek API 做开发的人,或者想找一个终端原生 AI 编程工具的人。
不适合谁:重度依赖 Claude Code Skills 生态的人(迁移有成本),或者对 experimental 功能零容忍的生产环境。
我接下来准备做的:把 AGENTS.md 按照 SeedScribe 项目的规则写好,然后在交互模式里跑一次实际的 Bug 修复。
安装环境:Windows 11, Node.js v22.22.1, deepseek-tui 0.8.14 API: DeepSeek V4 Pro 时间:2026-05-06 实测