deepseek-terminal-cli

An interactive computer agent for your terminal, powered by DeepSeek.

Run deepseek-cli and you're dropped into a shell where you type tasks. The agent runs commands, reads and writes files, and iterates until the task is done — like Claude Code, but for DeepSeek.

npm install -g deepseek-terminal-cli
deepseek-cli

DeepSeek Terminal CLI v1.2.0
cwd: /home/you/repo  •  model: deepseek-chat  •  /yes for auto-approve  •  /help for commands

> find every TODO in this repo and write them to TODOS.md
Searching the repo for TODOs.
⏺ run_shell(command="grep -rn TODO --include='*.js' .")
  ./src/index.js:42: // TODO: handle empty input
  ./src/auth.js:17:  // TODO: rotate keys
Writing the consolidated list now.
⏺ write_file(path="TODOS.md", content="# TODOs\n\n- src/index.js:42 — handle empty input\n- src/auth.j…")
  wrote 96 bytes to /home/you/repo/TODOS.md
✓ Wrote 2 TODOs from src/ into TODOS.md.

>

Install

npm install -g deepseek-terminal-cli

Requires Node.js 18+. After install, the deepseek-cli command is on your PATH.

How it works

deepseek-cli opens an interactive shell. You type a natural-language task; DeepSeek drives a tool-calling loop with full access to your computer:

• run_shell — execute any shell command (PowerShell on Windows, your $SHELL or /bin/sh on Mac/Linux).
• read_file — read a UTF-8 file.
• write_file — create or overwrite a file (creates parent dirs).
• append_file — append to a file.
• list_dir — list directory contents.
• finish — declare the task done.

By default each tool call is shown and you must approve it with y. Type /yes to auto-approve for the rest of the session, /no to re-enable confirmations.

Conversation history persists across turns: you can say "now make it async" and the agent remembers what "it" is.

Shell commands

| Command | Effect | | --- | --- | | /exit, /quit | Exit the shell | | /reset | Clear conversation history | | /yes | Auto-approve every tool call this session | | /no | Re-enable per-tool confirmations | | /cwd [path] | Show or change the working directory | | /help | Shell commands |

Launch flags

| Flag | Effect | | --- | --- | | --yes, -y | Start with auto-approve enabled | | --cwd <path> | Initial working directory | | --max-steps <n> | Cap on agent steps per turn (default 25) |

Non-interactive utility commands:

deepseek-cli help       # show help
deepseek-cli version    # print version

Authentication

The CLI ships with a built-in default DeepSeek API key, so it works out of the box. To use your own key, set DEEPSEEK_API_KEY — it overrides the default.

# macOS / Linux
export DEEPSEEK_API_KEY=sk-...

# Windows PowerShell
$env:DEEPSEEK_API_KEY = "sk-..."

Get a key at platform.deepseek.com. Auth is Authorization: Bearer <key> against the OpenAI-compatible DeepSeek API.

Environment variables

| Variable | Purpose | Default | | --- | --- | --- | | DEEPSEEK_API_KEY | Override the built-in default API key | (built-in) | | DEEPSEEK_BASE_URL | API base URL | https://api.deepseek.com | | DEEPSEEK_MODEL | Default model | deepseek-chat |

Safety

• Without /yes, the agent prompts before every shell command and every file write.
• The agent uses your full user privileges. There is no sandbox.
• Treat /yes like running a script you didn't read. Don't aim it at production systems.
• The CLI sends your task, tool output, and (when the agent reads files) file contents to the DeepSeek API.

License

MIT — see LICENSE.