A native, terminal-first coding agent for the DeepSeek V4 model family — built to match the look-and-feel of Claude Code, command-for-command, while running entirely on your DeepSeek API key.

There are two great DeepSeek terminal projects already (Hmbown/DeepSeek-TUI in Rust, CJackHwang/ds2api in Go), and both inspired this one. DeepSeek-CLI is different in three ways:

1. Identical UX to Claude Code. Slash commands, tool cards, permission prompts, status bar, splash screen — all match the Claude reference, so muscle memory carries over.
2. TypeScript + Ink stack. Same stack Anthropic uses for Claude Code itself; visual fidelity is highest here.
3. Native API + reverse-proxy login. Officially supports DEEPSEEK_API_KEY, with optional ds2api-style web-session reverse proxy for users who prefer browser auth.

Roadmap: see TODO.md. M2 (this release) adds the login proxy, plan/agent/yolo modes, reasoning tiers, and the two-pane splash. M3–M5 add subagents, MCP/Skills/Hooks, and binary distribution.

M5 will ship a one-liner installer. For now, install from source:

git clone git@github.com:yinshuo-thu/deepseek-cli.git
cd deepseek-cli
npm install
npm run build
npm link        # exposes `deepseek` and `ds` globally

Dev mode (no build step):

npm run dev      # runs src/cli.tsx via tsx

export DEEPSEEK_API_KEY=sk-...
deepseek

First run with no key shows a wizard that prompts for one and writes it to ~/.deepseek/config.json (chmod 600).

One-shot mode for scripting:

deepseek --print "summarise this repo's README" < README.md

Switch in-session: /model deepseek-v4-pro. When /reasoning max is set, the agent transparently routes to deepseek-reasoner for that turn.

Shipped through M2.0:

• Slash commands: /help /clear /model /mode /reasoning /config /cost /resume /cwd /login /logout /whoami /exit
• Tools: Read · Write · Edit · Bash · Glob · Grep · list_dir · apply_patch · WebFetch · WebSearch
• Modes: Plan / Agent / YOLO — Tab cycles through them with the input box empty
• Reasoning effort: off / high / max — Shift+Tab cycles. max routes the turn to deepseek-reasoner
• Streaming agent loop with permission prompts (allow once / always / deny) and Read-before-Write safety
• Per-project session persistence at ~/.deepseek/projects/<slug>/sessions/ plus a /resume picker
• First-run API-key wizard and ds2api-style /login flow with a local OpenAI-compatible reverse proxy that wires the real DeepSeek-web format end-to-end (M2.1a) — see the /login caveats below
• Two-pane Claude-Code-style splash with Recent activity + What's new
• DeepSeek-blue theme and whale brand art; Esc cancels the stream, Ctrl+C twice to exit

The /login flow speaks the reverse-engineered chat.deepseek.com web protocol — it is not an officially supported DeepSeek API surface. It works the same way ds2api does: paste the Cookie header from your logged-in browser tab, and a local OpenAI-compatible proxy rewrites every /v1/chat/completions request into a DeepSeek-web chat_session/create + chat/completion SSE handshake (with a pure-JS DeepSeekHashV1 solver for the proof-of-work challenge when DeepSeek demands one).

Best-effort, may break, may violate DeepSeek's Terms of Service. Use at your own risk. The official API-key flavor (DEEPSEEK_API_KEY) remains the default and recommended path; /login is for users who prefer browser auth. DeepSeek can change the wire format, tighten Cloudflare gating, or revoke web sessions at any time, and we will not always patch the same week.

1. Open https://chat.deepseek.com in Chrome/Edge/Firefox and log in normally.
2. Open DevTools (F12) → Network tab → reload the page.
3. Click any request to chat.deepseek.com/api/v0/... → Headers → Request Headers.
4. Copy the entire value of the Cookie: header (everything after Cookie: , all on one line — typically several KB).
5. Run /login in deepseek-cli, paste the cookie into the local browser page that opens, and submit.

If validation fails with "session expired" your cookie is stale (re-login at chat.deepseek.com). If it fails with "DDoS-Guard challenge" Cloudflare is gating that IP — pass a fresh challenge in your browser first, then retry.

src/
├── cli.tsx            # entry: yargs, first-run wizard, render <App/>
├── App.tsx            # main TUI — state + agent loop driver
├── api/
│   ├── client.ts      # streaming OpenAI-compatible client (SSE)
│   └── types.ts       # ChatMessage, ToolCall, StreamEvent
├── agents/
│   └── loop.ts        # stream → accumulate tool_calls → run → loop
├── auth/              # /login proxy server + session store
├── tools/             # Read · Write · Edit · Bash · Glob · Grep · list_dir · apply_patch · WebFetch · WebSearch
├── commands/          # slash-command registry
├── ui/                # Splash, Message, StatusBar, Permission, ResumePicker, theme, markdown
├── session/           # per-project persistent message log
└── config/            # ~/.deepseek/config.json + project key hashing

This project is itself built by a small fleet of role-specific Claude Code agents driven by the orchestrator in .claude/:

• .claude/agents/ — agent definition files. Each is a Markdown front-matter spec for a role (plan, dev, eval, summary, github) plus its tools and behavioural rules.
• .claude/runs/ — a chronological log of every round. Files like 2026-05-02-r03-dev-m20-skeleton.md capture the prompt, the agent's output, and the evaluator's findings.

The typical loop is Plan → DEV → Eval → Summary per milestone, with the GitHub agent handling commits and PRs. New contributors can read runs/ to see exactly how each feature got built and why each design call was made.

DeepSeek-CLI is designed as a feature-parallel open-source alternative to Claude Code, running on DeepSeek models instead of Claude. The table below tracks parity:

Feature	DeepSeek-CLI	Claude Code
Core UX
Streaming chat + Markdown rendering	✅	✅
LaTeX math display (inline & block)	✅	✅
Reasoning / thinking blocks	✅	✅
Cost & token tracking (/cost)	✅	✅
Plan / Agent / YOLO modes	✅	✅
Permission prompts (once / always / deny)	✅	✅
Tools
Read · Write · Edit · Bash · Glob · Grep	✅	✅
WebFetch · WebSearch	✅	✅
apply_patch	✅	✅
Multi-agent (Agent / SendMessage / Monitor)	✅	✅
Multimodal / image input	❌	✅
Slash Commands
/help · /clear · /model · /mode · /cost	✅	✅
/resume + per-project session persistence	✅	✅
/compact (context compression)	✅	✅
/doctor · /init · /bug	✅	✅
/login (browser OAuth / reverse proxy)	✅	❌
/reasoning effort tiers (off / high / max)	✅	❌
Extensions
MCP servers	✅	✅
Skills (slash commands loaded from .md)	✅	✅
Hooks (PreToolUse / PostToolUse / Stop / UserPromptSubmit)	✅	✅
Auth & Distribution
API key auth (DEEPSEEK_API_KEY)	✅	✅
Browser-session login proxy (ds2api-style)	✅	❌
npm link global install	✅	✅
IDE integration (VS Code / JetBrains)	❌	✅
One-line binary installer	🔜 M5	✅
Advanced
Memory system (project-level CLAUDE.md)	❌	✅
Background / scheduled agents	❌	✅
Git-aware operations (native git tool)	❌	✅

~/.deepseek/config.json:

{
  "baseUrl": "https://api.deepseek.com",
  "model": "deepseek-v4-flash",
  "theme": "dark",
  "telemetry": false,
  "apiFlavor": "openai",
  "permissionMode": "agent",
  "reasoningEffort": "off"
}

Override per run with --api-key, --base-url, --model or env vars.

Issues and PRs welcome. The fastest way to make a change is npm run dev then iterate on src/. Before opening a PR, run npm run typecheck.

MIT — see LICENSE.

DeepSeek-CLI is an independent project. Not affiliated with DeepSeek or Anthropic.