zot: Yet Another Coding Agent Harness

My notes

Summary

zot is a single-binary coding agent written in Go that competes with Claude Code, Codex CLI, and similar tools. It targets developers who want a lightweight, provider-agnostic alternative with no runtime dependencies, just drop it on $PATH and run. Its differentiator is intentional minimalism: four core tools, clean session control, and an extension model rather than a built-in plugin ecosystem.

Key Insight

• Architecture choice matters. Single Go binary = zero installation friction, no Docker, no Node, no Python env. Meaningful advantage for teams doing remote/server-side agent work.
• Four tools only at core: read, write, edit, bash. This matches what most coding agents actually use 90% of the time; the constraint is a feature, not a limitation.
• Broad provider support out of the box: Anthropic, OpenAI/Codex, Gemini, Kimi, DeepSeek, Bedrock, Vertex, Copilot, OpenRouter, local models. One agent loop, any model, useful for cost-switching or model comparison without changing workflow.
• Swarm mode allows background subagents each running their own model loop in the same repo, parallel workstreams without a separate orchestrator.
• Session control is first-class: resume, fork, branch, compact, export portable transcripts. Auto-compacts before hitting context window limits, a practical safeguard missing from many alternatives.
• Telegram bridge lets you drive the agent remotely via DM, same read/write/edit/bash capabilities as the TUI. Useful for monitoring long-running tasks from mobile.
• No mandatory MCP is an explicit design decision. Extensions use subprocess + JSON-RPC instead, keeping the binary lean. MCP can be added via extension if needed.
• Config philosophy: single config.json, override system prompt with one file, no sprawl. Low cognitive overhead for multi-machine setups.