Overview
DeepSeek-Reasonix is a config-driven, plugin-driven AI coding agent packaged as a single static Go binary. With 30,000+ stars, it is tuned around DeepSeek's prefix cache for low token costs across long sessions and is available as a CLI/TUI, desktop app, and VS Code extension.
Features
- ✓Config-driven via reasonix.toml
- ✓Multi-model and composable
- ✓MCP-compatible plugin system
- ✓Prefix-cache optimization for low costs
- ✓Single static binary distribution
- ✓CLI/TUI, desktop, and VS Code modes
Installation
npm i -g reasonixPros
- +Single binary, minimal dependencies
- +Prefix-cache optimization reduces token costs
- +Fully configurable via TOML
- +30,000+ stars, active community
- +MIT license
Cons
- −Requires DeepSeek or OpenAI-compatible API key
- −Optimized primarily for DeepSeek prefix cache
- −Relatively new, features still maturing
- −Go toolchain needed for source builds
Alternatives
Documentation
DeepSeek-Reasonix
Overview
DeepSeek-Reasonix is a DeepSeek-native AI coding agent for the terminal, packaged as a single static Go binary. With 30,000+ GitHub stars, Reasonix is tuned around DeepSeek's prefix cache so token costs stay low across long sessions — a significant advantage for developers who keep an agent running throughout a workday.
Reasonix is config-driven and plugin-driven: providers, the agent, enabled tools, and plugins are all declared in a reasonix.toml file, with no hardcoded models. It can run any OpenAI-compatible endpoint, but ships with DeepSeek as the optimized default. The project offers three access modes from a single engine: an interactive CLI/TUI, a desktop application, and a VS Code extension.
Features
- Config-driven: providers, agent, tools, and plugins all declared in
reasonix.toml— no hardcoded models - Multi-model and composable: DeepSeek ships as default preset, but any OpenAI-compatible endpoint works; optionally run two models (executor + planner) in separate cache-stable sessions
- Plugin-driven: external tools run as subprocesses over stdio JSON-RPC (MCP-compatible); built-in tools self-register at compile time
- Cache-aware context maintenance: startup injects a small stable environment summary; stale tool output is pruned before summary compaction
- Prefix-cache optimization: tuned for DeepSeek's prefix cache to minimize token costs across long-running sessions
- Single binary distribution:
CGO_ENABLED=0static binary, cross-compile to six targets with one command - Three access modes: CLI/TUI, desktop app, and VS Code extension all share the same engine
Installation
# CLI / TUI via npm
npm i -g reasonix
# CLI / TUI via Homebrew (macOS)
brew install esengine/reasonix/reasonix
# VS Code extension
# Search "SivanLiu.reasonix-agent" in VS Code Marketplace or Open VSX
# Desktop app
# Download from reasonix.io
Build from source:
git clone https://github.com/esengine/DeepSeek-Reasonix
cd DeepSeek-Reasonix
make build
Prebuilt binaries are available for darwin, linux, and windows across amd64 and arm64 architectures.
Quick Start
# Configure a provider and model
reasonix setup
# Start an interactive terminal session
reasonix
# Run a single command
reasonix run "implement the TODOs in main.go"
# In an interactive session, create project instructions
/init
Core Concepts
- Config over code: the entire agent configuration lives in a TOML file — switching providers or models requires no code changes
- MCP-compatible plugin system: external tools communicate over stdio JSON-RPC, the same protocol used by MCP servers
- Cache stability: the engine maintains a small, stable context prefix so DeepSeek's prefix cache delivers consistent token savings across sessions
- Single binary: no Python runtime, no Node.js dependency — the agent is a portable executable
Advanced Features
- Dual-model mode: run an executor model and a planner model in parallel, each in its own cache-stable session
- Context pruning: stale tool output is automatically snipped before summary compaction, keeping context windows efficient
- Custom plugin development: build your own tools as stdio JSON-RPC subprocesses and register them in the config
- VS Code integration: editor context, tool-call approvals, and model selection all work within the IDE
Examples
| Scenario | Description |
|---|---|
| Interactive TUI coding | Open a terminal and converse with an AI agent on any codebase |
| Persistent background agent | Keep Reasonix running across a full day of development |
| VS Code chat | Get AI chat with full editor context and inline tool-call approvals |
| Desktop app workflow | Use a GUI-based agent for developers who prefer a visual interface |
| Custom plugin development | Extend with domain-specific tools via the MCP-compatible interface |
Pros
- ✅ Single binary with minimal dependencies — easy to distribute and deploy
- ✅ Prefix-cache optimization keeps token costs down over long sessions
- ✅ Fully configurable via TOML; no code changes needed to swap models
- ✅ Three access modes (CLI, desktop, VS Code) share one engine
- ✅ Large active community (30,000+ stars)
- ✅ MCP-compatible plugin system for extensibility
- ✅ MIT license
Cons
- ❌ Requires a DeepSeek (or OpenAI-compatible) API provider and API key
- ❌ Primarily optimized for DeepSeek's prefix caching — other providers may not benefit equally
- ❌ Relatively new and evolving — some features still maturing
- ❌ Go toolchain needed for building from source
When to Use
Use DeepSeek-Reasonix when you want a terminal-based AI coding companion with low token costs from prefix caching. It is ideal for developers working primarily in the terminal who want a persistent agent that stays cheap across long sessions. If you need multi-agent orchestration, chat-only workflows, or a no-server local model, frameworks like CrewAI, LangGraph, or AirLLM are better fits.