
DeepSeek TUI
github.com/hmbown/deepseek-tui- Category
- AI Agents
- Rank
- No. 100Tools index
- Type
- APP
- Builder
- hmbown
- GitHub
- 40.9k stars
- Latest release
- v0.9.12
- Date
About
DeepSeek TUI is a terminal-native coding agent built around DeepSeek V4's 1M-token context and prefix cache. It ships as a single binary (no Node or Python runtime) with a built-in MCP client, sandbox, and durable task queue.
What it does
Codewhale takes a coding task, inspects the workspace, edits code, runs commands, and reports verification evidence. It supports interactive terminal work, headless automation, and a local browser view. Roles can use different providers, models, reasoning levels, instructions, and permission limits within one coordinated fleet.
Why it's ranked here
The strongest reason to consider it is control over the harness, not allegiance to one model. Explicit role routing, reviewable tool activity, resumable fleet ledgers, approval gates, and repository-level write holds make serious automation plausible. The breadth also raises setup and operating complexity, especially when mixing fleets, workflows, hooks, MCP servers, and permission postures.
What's good
Provider and model choices attach directly to saved roles, preventing accidental dependence on the currently selected provider. Plan mode blocks mutation, while stronger modes remain subject to approvals, sandboxing, and repository rules. Fleet work survives interruption through an append-only ledger. MCP supports local processes and remote servers, including OAuth, tool discovery, resources, prompts, and hot reload in the terminal interface.
Tradeoffs
The product exposes many overlapping concepts: modes, permission postures, fleets, workflows, roles, hooks, plugins, and memory. Teams must understand their boundaries before trusting automation. Linux command isolation requires opt-in bubblewrap, while macOS isolation depends on Seatbelt availability. Hooks only fire through the interactive terminal runtime. The browser client is loopback-only and is explicitly unsuitable for public hosting.
How to use it well
Use it for engineers who want one local coding harness across hosted and local models, particularly for reviewable repository work or durable multi-worker jobs. Start unfamiliar projects in read-only Plan mode, then grant narrower permissions after inspecting the proposed work. Use headless execution for scripts and CI. Choose another system when the main requirement is a remotely hosted collaborative browser workspace.
Technical notes+
Cargo.toml defines a Rust 2024 workspace with 20 members, Rust 1.88 as the minimum toolchain, crates/cli as the default member, and separate release and distribution profiles. package.json is a private npm workspace covering codewhale, the legacy deepseek-tui package, and a runtime SDK. docs/MCP.md specifies stdio and Streamable HTTP transports with legacy SSE fallback, OAuth support, and TUI hot reload, while headless runtimes require restart after MCP changes. docs/WEB.md documents a loopback-only Runtime API client with a one-time bootstrap capability exchanged for an HttpOnly session cookie. docs/FLEET.md describes JSONL ledger persistence, leases, heartbeats, retries, receipts, and resumable workers. docs/HOOKS.md limits lifecycle hooks to the interactive TUI. docs/MEMORY.md describes opt-in Markdown storage indexed locally with SQLite FTS5.
Observed
- License
- MIT
- Primary language
- Rust, using the 2024 edition and requiring Rust 1.88 or newer
- Packaging
- Available through npm, Cargo, Docker, Nix, Scoop, prebuilt archives, Android/Termux, and a CNB mirror
- Interfaces
- Interactive terminal UI, headless CLI execution, loopback browser client, MCP server mode, and Runtime HTTP/SSE API mode
- MCP transports
- Local stdio processes and remote Streamable HTTP servers with legacy SSE fallback
- Container platforms
- Published Linux images support amd64 and arm64
- Repository structure
- The Cargo workspace declares 20 member crates and uses the CLI crate as its default member
Read from README.md, Cargo.toml, package.json, docs/MCP.md, docs/WEB.md, docs/FLEET.md, docs/GUIDE.md, docs/HOOKS.md, docs/MODES.md, docs/VOICE.md, docs/DOCKER.md, docs/MCP.id.md, docs/MEMORY.md.
What it can do
Read and edit files in workspace
File paths and edit instructions → Modified files with changes applied
Execute shell commands
Shell command strings → Command execution results and output
Search and browse the web
Search queries or URLs → Web content and search results
Manage git operations
Git commands and repository actions → Git status updates and version control changes
Apply code patches
Patch files or diff content → Updated codebase with patches applied
Generate parallel analysis using multiple AI agents
Analysis tasks and code/data to examine → Batched analysis results from 1-16 parallel agents
Save and resume coding sessions
Current session state and workspace → Checkpointed session file for later resumption
Rollback workspace changes
Previous workspace snapshot or turn number → Restored workspace state without affecting main git repository
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.