DeepSeek-TUI: Run a DeepSeek Coding Agent in Your Terminal

DeepSeek-TUI is an AI coding agent that runs in the terminal. It is built around DeepSeek V4 models and starts from the deepseek command. Inside a keyboard-driven TUI, it can read and edit files, run shell commands, search the web, manage git, connect to MCP servers, and coordinate sub-agents.

It is closer to a terminal workbench than a simple chat CLI. The goal is not only to send a question to a model, but to combine code reading, file edits, commands, diagnostics, session recovery, and workspace rollback in one local workflow.

The repository is mainly written in Rust and uses the MIT license. Its GitHub description is: “Coding agent for DeepSeek models that runs in your terminal.”

Who It Is For

DeepSeek-TUI fits developers who prefer terminal workflows and want to use DeepSeek models for real local development tasks.

It is useful when you want to:

Use DeepSeek models for code changes and project analysis.
Work without opening a full IDE.
Let an AI tool read and modify a local workspace.
Switch between Plan, Agent, and YOLO modes.
Save sessions and resume long tasks.
Roll back workspace changes.
Connect MCP, LSP diagnostics, HTTP/SSE runtime APIs, and skills.

If you only need simple Q&A, a web client or lightweight CLI is enough. DeepSeek-TUI is better when the model should become part of your local development loop.

Installation

DeepSeek-TUI ships Rust binaries. The common entry command is deepseek, and the companion TUI binary is deepseek-tui.

Using npm:

1
2
3


npm install -g deepseek-tui
deepseek --version
deepseek --model auto

The npm package is an installer and wrapper that downloads prebuilt Rust binaries. It requires Node.js >=18.

Using Cargo:

1
2


cargo install deepseek-tui-cli --locked
cargo install deepseek-tui --locked

Using Homebrew:

1
2


brew tap Hmbown/deepseek-tui
brew install deepseek-tui

You can also download prebuilt binaries from GitHub Releases. The README lists Linux x64/ARM64, macOS x64/ARM64, and Windows x64 builds.

Docker:

1
2
3
4


docker run --rm -it \
  -e DEEPSEEK_API_KEY \
  -v "$PWD:/workspace" \
  ghcr.io/hmbown/deepseek-tui:latest

In mainland China, use npm or Cargo mirrors, or download release binaries manually.

Configure the API Key

On first launch, DeepSeek-TUI asks for your DeepSeek API key and saves it to:

1

~/.deepseek/config.toml

You can also configure it explicitly:

1
2


deepseek auth set --provider deepseek
deepseek auth status

Or use an environment variable:

1
2


export DEEPSEEK_API_KEY="YOUR_KEY"
deepseek

Check the setup:

1

deepseek doctor

If the wrong key source is used, run deepseek auth status. Saved config keys take precedence over the keyring and environment variables.

Clear a saved key:

1

deepseek auth clear --provider deepseek

Auto Mode

Auto mode:

1

deepseek --model auto

Inside the TUI:

1

/model auto

Auto mode chooses both:

Model: deepseek-v4-flash or deepseek-v4-pro
Thinking: off, high, or max

Before the real request, DeepSeek-TUI makes a small routing call to analyze the latest request and recent context. Simple tasks can stay on Flash with thinking off; coding, debugging, architecture, release, security review, or ambiguous multi-step work can move to Pro or higher thinking.

auto is local to DeepSeek-TUI. The upstream API receives the concrete model and thinking setting chosen for that turn. For benchmarks, strict cost control, or fixed behavior, use an explicit model.

Modes

Mode	Use
Plan	Read-only exploration and planning
Agent	Default interactive mode with approval gates
YOLO	Auto-approve tools in a trusted workspace

Plan is for investigation. Agent is safer for everyday coding. YOLO is fast but risky and should only be used in trusted workspaces and low-risk branches.

Tooling

The README lists a broad tool set:

File reads/writes and apply patch.
Shell execution.
Git operations.
Web search and browse.
Sub-agents.
MCP servers.
LSP diagnostics.
Session save/resume.
Workspace rollback.
Durable task queue.
HTTP/SSE runtime API.
Skills system.

LSP diagnostics are especially useful because errors from rust-analyzer, pyright, typescript-language-server, gopls, clangd, and similar tools can be fed back after edits.

Workspace rollback uses side-git snapshots and provides /restore and revert_turn. It does not touch the repository’s own .git, but normal git commits are still the safest baseline.

Common Commands

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19


deepseek
deepseek "explain this function"
deepseek --model deepseek-v4-flash "summarize"
deepseek --model auto "fix this bug"
deepseek --yolo
deepseek auth set --provider deepseek
deepseek doctor
deepseek doctor --json
deepseek models
deepseek sessions
deepseek resume --last
deepseek resume <SESSION_ID>
deepseek fork <SESSION_ID>
deepseek serve --http
deepseek serve --acp
deepseek pr <N>
deepseek mcp list
deepseek mcp validate
deepseek update

Zed and ACP

DeepSeek-TUI can run as an Agent Client Protocol server:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10


{
  "agent_servers": {
    "DeepSeek": {
      "type": "custom",
      "command": "deepseek",
      "args": ["serve", "--acp"],
      "env": {}
    }
  }
}

The README notes that the first ACP slice supports new sessions and prompt responses, while tool-backed editing and checkpoint replay are not exposed through ACP yet.

Configuration and Providers

User config:

1

~/.deepseek/config.toml

Workspace overlay:

1

<workspace>/.deepseek/config.toml

Some sensitive fields, such as api_key, base_url, provider, and mcp_config_path, are denied in workspace overlays.

DeepSeek-TUI supports the default deepseek provider plus NVIDIA NIM, Fireworks, OpenAI-compatible endpoints, SGLang, vLLM, Ollama, and others.

OpenAI-compatible example:

1
2


deepseek auth set --provider openai --api-key "YOUR_OPENAI_COMPATIBLE_API_KEY"
OPENAI_BASE_URL="https://openai-compatible.example/v4" deepseek --provider openai --model glm-5

Ollama example:

1
2


ollama pull deepseek-coder:1.3b
deepseek --provider ollama --model deepseek-coder:1.3b

Cost and Context

DeepSeek-TUI is designed around DeepSeek V4. The README mentions deepseek-v4-pro and deepseek-v4-flash, 1M-token context windows, token usage, cost estimates, and prefix-cache telemetry.

For light tasks, deepseek-v4-flash or auto mode may be enough. For complex refactors, long-context debugging, and architecture work, use higher thinking or Pro.

Pricing and discounts change, so check DeepSeek’s official pricing page and the current TUI cost estimates before relying on numbers.

Suggested Workflow

Start safely:

Try it in a small test repository.
Run deepseek doctor.
Use Plan mode for read-only exploration.
Use Agent mode for small edits.
Review changes with git diff and tests.
Learn /restore and session recovery.
Use YOLO only in trusted temporary branches.

Do not store API keys in project files. For company code, confirm provider, logging, web search, and compliance rules first.

Summary

DeepSeek-TUI is a full terminal AI coding agent. It brings DeepSeek V4, TUI interaction, tool calls, LSP diagnostics, session recovery, rollback, MCP, and skills into one Rust-based workflow.

It is not the lightest DeepSeek client, but its strength is moving from chat to executable local development.