Configuration — DeepSeek TUI Wiki

DeepSeek TUI reads config.toml for everyday defaults—model choice, safety, reasoning depth, hooks, and optional provider profiles.

Where the file lives

# macOS / Linux
~/.deepseek/config.toml

# Windows
%USERPROFILE%\.deepseek\config.toml

Starter example

[api]
key = "sk-..."
base_url = "https://api.deepseek.com"
model = "deepseek-v4-flash"

[ui]
default_mode = "agent"
approval_mode = "suggest"
reasoning_effort = "high"

[mcp]
enabled = true

Exact keys evolve with releases—when in doubt, mirror the sample from the official repository wiki bundled with your installed version.

Approval modes

Modes (Plan / Agent / YOLO) describe how autonomous the agent feels. Approval settings further gate tool calls—common values include suggest, auto, and never. Treat them as independent axes: you can be in Agent mode while still tightening approvals.

Reasoning effort

Reasoning controls how hard the model tries on supported endpoints—trade latency and tokens against depth. Cycle shortcuts are documented in upstream release notes.

Providers and API keys

The upstream docs list multiple providers. Typical IDs include:

Provider	ID	Env var (typical)
DeepSeek	`deepseek`	`DEEPSEEK_API_KEY`
NVIDIA NIM	`nvidia-nim`	`NVIDIA_API_KEY`
OpenAI	`openai`	`OPENAI_API_KEY`
OpenRouter	`openrouter`	`OPENROUTER_API_KEY`
Novita	`novita`	`NOVITA_API_KEY`
Fireworks	`fireworks`	`FIREWORKS_API_KEY`
SGLang	`sglang`	`SGLANG_API_KEY`
vLLM	`vllm`	`VLLM_API_KEY`
Ollama	`ollama`	`OLLAMA_API_KEY`

Use deepseek auth set --provider <id> to align CLI auth with the backend you chose.

Hooks

Hooks let you run commands around lifecycle events—session start, before/after tool calls, and more. Enable them when you want logging, metrics, or lightweight automation tied to agent behavior.

MCP config file

MCP servers commonly live in ~/.deepseek/mcp.json. See MCP & skills for examples.

Suggested profiles

Budget-friendly

[api]
model = "deepseek-v4-flash"

[ui]
default_mode = "agent"
approval_mode = "suggest"
reasoning_effort = "off"

Balanced

[api]
model = "deepseek-v4-flash"

[ui]
default_mode = "agent"
approval_mode = "suggest"
reasoning_effort = "high"

Heavy reasoning

[api]
model = "deepseek-v4-pro"

[ui]
default_mode = "agent"
approval_mode = "suggest"
reasoning_effort = "max"

Configuration guide