Go rewrite of a terminal agent obsessed with DeepSeek's prefix cache

What it does Reasonix is a terminal-based AI coding agent rewritten in Go that edits files, runs shell commands, and fetches context through a bubbletea TUI. All behavior is driven by reasonix.toml: providers, built-in tools, MCP plugins, and permission rules are declared there rather than hardcoded.

The interesting bit The entire design is built around prefix-cache economics. It optionally splits work between two models—say, a cheap flash executor and a slower pro planner—running in separate cache-stable sessions so long conversations don’t repeatedly re-bill the same system context. Under the hood, providers and built-in tools self-register via Go init() functions, while external tools speak newline-delimited JSON-RPC over stdio using the MCP protocol.

Key highlights

• Ships as a single static Go binary with no CGO; cross-compiles to six OS/arch targets
• Native MCP client: consumes both local stdio subprocesses and remote HTTP servers, and reads existing .mcp.json files without conversion
• Two-model collaboration: executor and planner run in separate sessions, with subagent tasks dispatchable to yet another configured model
• Permission layers: policy-based deny/ask/allow gating for tool calls, plus a sandbox that confines file writers to workspace_root
• Custom slash commands defined by dropping Markdown prompt templates into .reasonix/commands/

Caveats

• Version 1.0.0 is not yet published to npm, so the prebuilt binary must be built from source for now
• Bash sandboxing only enforces restrictions on macOS via Seatbelt; Linux and Windows currently run unconfined

Verdict Worth a look if you want a hackable, terminal-first agent with granular cost control and MCP extensibility. Skip it if you need a polished GUI or fully cross-platform sandboxing today.