DeepSeek-Reasonix: a DeepSeek-native terminal coding agent with cache-aware context maintenance and MCP-compatible plugins

What it solves

Reasonix is a terminal-based AI coding agent designed to reduce token costs and friction during long development sessions. It provides a structured harness for AI models to interact with a local codebase, allowing developers to automate tasks like implementing TODOs or adding unit tests directly from the command line.

How it works

Built as a single static Go binary, Reasonix uses a config-driven approach where providers, models, and tools are defined in a reasonix.toml file. It is specifically tuned for DeepSeek's prefix cache to maintain low costs across sessions. The agent can use built-in tools or external plugins that run as subprocesses via stdio JSON-RPC (MCP-compatible). It also manages context by injecting environment summaries and pruning stale tool output to keep the conversation efficient.

Who it’s for

It is designed for developers who want a high-performance, low-overhead AI coding assistant integrated into their terminal, particularly those using DeepSeek or other OpenAI-compatible models.

Highlights

• Config-driven architecture: No hardcoded models; everything from providers to plugins is declared in TOML.
• Cache-aware context: Optimized for prefix caching to minimize token usage during long sessions.
• MCP-compatible plugins: Supports external tools running as subprocesses over JSON-RPC.
• Zero-friction distribution: Distributed as a single Go binary with cross-platform support.
• Multi-model composition: Ability to run a planner and an executor model simultaneously in separate sessions.