zot

zot is a minimal terminal coding agent shipped as a single static Go binary, created by patriceckhart and released under the MIT license. It connects to over 25 LLM providers — including Anthropic, OpenAI, DeepSeek, Google Gemini, GitHub Copilot, and local Ollama models — and exposes a clean agent loop with file read/write/edit and bash execution tools. The project describes itself as "yet another coding agent harness, lightweight and written (vibe-slopped) in go," and is self-described as "in beta forever."

What It Is

zot is a CLI-first coding agent harness that wraps a broad LLM provider catalog behind a single binary with no runtime dependencies, no Docker requirement, and no dedicated plugin package manager. It sits in the category of terminal-based AI coding assistants and agent harnesses — tools that let developers drive large language models to read, edit, and run code directly in their working directory. The core job is to give developers a fast, portable agent loop they can drop on any machine and immediately use with their existing API keys or subscriptions.

Provider and Model Coverage

The built-in provider catalog is one of zot's most distinctive features. It covers:

• Subscription-capable: Anthropic Claude Pro/Max, OpenAI Codex/ChatGPT Plus/Pro, Kimi Code, GitHub Copilot
• Direct API: Anthropic, OpenAI Chat Completions/Responses, DeepSeek, Google Gemini/Vertex, Groq, Cerebras, xAI, Together AI, Hugging Face Router, OpenRouter, Mistral, MiniMax, Fireworks, Vercel AI Gateway, Cloudflare AI, and more
• Local/compatible: Ollama and any OpenAI-compatible endpoint via --base-url

Model discovery merges a baked-in catalog with live /v1/models results (cached 6 hours) and speculative entries. Custom models can be added via a models.json file in $ZOT_HOME.

Run Modes and Embedding

zot supports four distinct run modes: an interactive TUI with streaming output and slash commands; a print mode (-p) that writes only the final assistant text to stdout; a JSON mode (--json) that emits newline-delimited JSON events for scripts and CI; and an RPC mode (zot rpc) that runs as a long-lived subprocess accepting NDJSON commands on stdin. Two embedding paths exist for developers: a Go in-process SDK (packages/agent/sdk) and the out-of-process RPC interface with reference clients in Python, Node, shell, and Go.

TUI Features and Slash Commands

The interactive TUI includes a rich set of slash commands covering session management, model switching, side-chat (/btw), background subagents (/swarm), skill discovery (/skills), context compaction (/compact), sandbox mode (/jail), and a Telegram bridge (/telegram). The /swarm command spawns parallel background subagents — each a separate zot subprocess with its own model loop and session file — that share the host's working directory. An auto-swarm setting lets the main agent fork subagents autonomously when a task splits into independent parallel work. Sessions are stored as JSONL transcripts and support export, import, forking, and branching via /session.

Extensions and Skills

zot's extension system uses a subprocess + JSON-RPC protocol, allowing extensions written in any language to register slash commands, expose new tools to the model, intercept tool calls, and open interactive panels inside the TUI. Extensions are opt-in and managed via zot ext install/list/enable/disable/remove. Skills are per-folder SKILL.md files with YAML frontmatter that zot discovers at startup and surfaces to the model on demand, keeping context lean by loading only the instructions needed.

Update: v0.2.3

The latest release is v0.2.3, published on 2026-05-29, with the repository last pushed on the same date. The project was created in April 2026 and has been actively releasing since. The GitHub repository shows 65 stars and 3 forks with zero open issues at the time of data collection. The project's self-described "beta forever" status signals ongoing, iterative development rather than a stable production release cadence.