DeepSeek TUI is a terminal-native coding agent built for DeepSeek V4 models. Written in Rust for speed, it provides a fast, keyboard-driven TUI that gives DeepSeek's frontier models (deepseek-v4-pro and deepseek-v4-flash) direct access to your workspace. Read and edit files, run shell commands, search the web, manage git operations, and orchestrate sub-agents — all without leaving your terminal.
The standout feature is native RLM (rlm_query), which fans out 1-16 cheap deepseek-v4-flash children in parallel for batched analysis, decomposition, or parallel reasoning tasks. This lets you solve complex problems faster by distributing work across multiple model instances simultaneously. Each child call runs independently and results are collected for synthesis.
DeepSeek TUI supports 1M-token context windows with automatic intelligent compaction when context fills up. The compaction algorithm preserves key information from earlier turns, so the model stays informed even during extended sessions. Thinking-mode streaming shows DeepSeek's chain-of-thought as it reasons about code, giving you visibility into the model's decision process and helping you understand why it makes specific changes.
Three interaction modes let you control autonomy: Plan mode is read-only exploration where the agent surveys your codebase without making changes. Agent mode requires approval before executing actions, giving you a safety net. YOLO mode auto-approves all actions for trusted workflows where speed matters more than oversight.
The tool includes workspace rollback with side-git snapshots, so you can restore previous states with /restore or revert an entire turn. Every turn gets a pre and post snapshot, making it easy to undo changes that didn't work out. Session save and resume lets you checkpoint long sessions and pick up exactly where you left off.
The HTTP/SSE runtime API enables headless agent workflows when you need to integrate DeepSeek TUI into automated pipelines or CI/CD systems. Run `deepseek serve --http` to start the API server and submit tasks programmatically.
MCP protocol support connects to Model Context Protocol servers for extended tooling, letting you integrate databases, APIs, and other external services. Live cost tracking shows per-turn and session-level token usage and cost estimates, so you always know how much your coding sessions cost.
Reasoning-effort tiers let you cycle through off, high, and max with Shift+Tab depending on task complexity. The skill system supports installable community skills from GitHub repos and local custom skills, extending the agent's capabilities beyond the built-in tool set.
DeepSeek TUI supports multiple providers beyond the official DeepSeek API: NVIDIA NIM, Fireworks AI, and self-hosted SGLang instances. UI localization covers Japanese, Chinese Simplified, Portuguese (Brazil), and English. The dark theme uses a DeepSeek-blue palette designed for long coding sessions.
Built with Rust (edition 2024), ratatui 0.29, tokio 1.49, and axum 0.8.4. The workspace includes 14 crates for modular architecture. Available via npm, cargo, or pre-built binaries for Linux, macOS, and Windows. Licensed under MIT.