![\"Agentplate\"](\"assets/agentplate-logo.svg\"){kind=logo}
\n\n \n
\n Self-improving multi-agent orchestration — from idea to deployed app.\n
\n\n\n ![\"npm](\"https://img.shields.io/npm/v/@ag-eco/agentplate?color=fb4b38&label=npm\")
![\"Runtime:](\"https://img.shields.io/badge/runtime-Bun-fb4b38\"){kind=icon}
![\"MIT](\"https://img.shields.io/npm/l/@ag-eco/agentplate?color=fb4b38\")
\n Spawn AI agent swarms in isolated git worktrees, coordinate them over a SQLite\n mail bus, and merge their work back. An interactive wizard picks your AI\n provider, a closed learning loop distills reusable skills as agents work, and a\n built-in pipeline takes you from build → configure CI/CD → deploy to the\n target of your choice — all observable from a CLI, a live TUI, and a web UI.\n
\n\n\n Quick start ·\n Features ·\n Commands ·\n Architecture ·\n Contributing\n
\n\n---\n\n> **v1.0.0 — first public release.** The orchestration engine, provider wizard,\n> self-improving skills, multi-runtime support, the web UI (`agentplate serve`),\n> and the build→deploy pipeline are all working and tested.\n\n## Why Agentplate\n\nMost \"AI coding agent\" tools give you *one* agent in *one* session. Agentplate turns a\nsingle command into a **team**: it spawns specialized worker agents in isolated git\nworktrees, coordinates them through a fast SQLite message bus, folds their work back\nwith tiered conflict resolution — and then **gets better the more you use it** by\ndistilling reusable skills from work that passed its quality gates.\n\n## Features\n\n- **🧙 Interactive setup wizard.** `agentplate setup` walks you through provider\n selection (Anthropic, OpenAI, OpenRouter, DeepSeek, Google, local/Ollama, or a\n custom OpenAI-compatible endpoint) and then **how to authenticate** — reuse an\n existing **subscription / CLI login** (e.g. a Claude Pro/Max session in Claude Code,\n `codex login`, or the Gemini CLI — no key stored), an **existing environment\n variable**, or **enter an API key** (saved to a **gitignored** secrets file). Switch\n any time with `agentplate model`. Models below a 64k context floor are rejected.\n- **🤖 Agent-swarm orchestration.** A `coordinator → lead → workers` hierarchy runs\n agents (scout, builder, reviewer, merger…) in isolated worktrees, coordinates them\n over `.agentplate/mail.db`, and merges branches back with a clean → auto-resolve tiered\n strategy. Headless **spawn-per-turn** execution — no tmux required.\n- **📈 Self-improving skills.** After an agent completes a task whose quality gates\n pass, Agentplate distills a reusable, versioned **skill** (a markdown playbook). Future\n agents retrieve the relevant skills into their context, apply them, and the skills\n earn or lose **confidence** from real outcomes — so the swarm compounds expertise.\n- **🚀 Build → CI/CD → Deploy.** A staged pipeline (architect → builder → devops →\n deployer → verifier) scaffolds an app, generates CI/CD + infrastructure, and ships\n it. Pluggable **deploy targets** (Docker + GitHub Actions today; PaaS, cloud,\n Kubernetes/Helm, and on-prem are adapter files). One command:\n `agentplate ship \"an idea\" --target docker-gha`.\n- **🛡️ Safe by construction.** Secrets live only in env-vars / a gitignored store and\n are injected solely into the deployer process — never committed, logged, or audited.\n Production deploys require an explicit gate (`--yes`); `--dry-run` generates config\n and plans with **zero** outward mutation; every deploy is recorded in an append-only\n audit log. Distilled skills are scrubbed for secrets and dangerous commands.\n- **🖥️ Three surfaces.** A CLI, a **3-pane `agentplate tui`** (active agents / live feed /\n tasks), and a modern **\"agent OS\" web UI** (`agentplate serve`) — an icon-rail shell with a\n Mission Control dashboard, an animated **3D Office** where agents sit and type, stand when\n idle, and walk to the whiteboard to talk when they hand off tasks (React Three Fiber), a real\n **System Monitor** (host CPU/RAM/disk/uptime), **Costs & Analytics** charts, **Tasks** and\n **Handoffs** views, a terminal-style live feed, per-agent detail drawers, a notification\n center, a **⌘K command palette**, and a chat to message the coordinator and submit tasks.\n Everything auto-refreshes every 5 seconds.\n- **💬 Conversational coordinator.** `agentplate coordinator start` opens an interactive\n Claude session primed as your orchestrator — describe what you want built and it dispatches\n the team.\n\n## Requirements\n\n- **[Bun](https://bun.sh) `>= 1.0`** — Agentplate runs on Bun (the CLI bin uses a\n `#!/usr/bin/env bun` shebang). Plain Node.js is **not** sufficient; install Bun first\n (`curl -fsSL https://bun.sh/install | bash`).\n- `git`\n- A coding-agent runtime — **Claude Code** (default), **OpenCode**, or **Codex** — and an API\n key (or subscription login) for your provider. Pick one in `agentplate setup`, or per\n command with `--runtime