AI-Powered, Non-Intrusive Terminal Assistant embedded in tmux sessions that observes multi-pane context and provides intelligent command suggestions with execution.
TmuxAI is an AI terminal assistant deeply embedded in tmux sessions, adopting a three-zone separation architecture — Chat Pane / Exec Pane / Read-Only Panes — to deliver a zero-workflow-switch "AI pair programming" experience. Core capabilities include: Pane Observe Mode (reading commands, output, shell/OS info from all visible panes as context), Prepare Mode (precisely tracking command execution history and exit codes), and Watch Mode (continuously monitoring terminal activity and proactively offering suggestions). For command execution, AI-suggested commands run in a dedicated Exec Pane with whitelist/blacklist pattern matching and three-tier risk indicators (✓/?/!).
The knowledge system features Knowledge Base for loading Markdown predefined context files, and a Skills system using frontmatter metadata for L1 discovery blocks with on-demand lazy loading, supporting auto-matching based on word frequency and description relevance. Context management via Context Squashing automatically triggers summary compression at 80% threshold, enabling sustained long sessions.
Model integration covers OpenAI (including Responses API), OpenRouter, Azure OpenAI, Google Gemini (go-genai SDK), AWS Bedrock (Converse API), and GitHub Copilot (official Go SDK), with multi-model concurrent configuration and runtime switching. External information retrieval integrates Brave Search / SearXNG search and web content fetching (with Wayback Machine / Google Cache fallback chains). Through Model Context Protocol, it supports stdio / SSE / streamable-http transports for connecting external MCP servers. Built on Go 1.25.0 with cobra (CLI), viper (config), chroma (syntax highlighting), go-readline-ny (REPL), and distributed via GoReleaser as precompiled binaries. Licensed under Apache-2.0.