Cloudflare Workers에서 실행되는 터미널 스타일 AI 코딩 에이전트

A terminal-style AI coding agent running fully on Cloudflare Workers. It combines: - a browser terminal UI ( ghostty-web ) - a SQLite-backed Durable Object for persistent sessions and files - pi's worker-friendly coding-agent runtime - Dynamic Worker Loader sandboxes for code execution - session-scoped published Workers - Durable Object alarm-powered cron jobs Live deployment in this repo currently targets: Note: this example requires a CF_GATEWAY_TOKEN secret, so you may still need to finish a small amount of post-deploy setup in Cloudflare/Wrangler. Inside a session, the agent can: - read, write, edit, and list files in a persistent filesystem - execute JavaScript in an isolated worker sandbox - import local files and many packages through esm.sh - publish a file as a Worker endpoint at /w// - create recurring cron jobs that send prompts back into the session examples/terminal-agent/ ├── src/ │ ├── index.ts # main Worker + Durable Object + loopback outbound entrypoint │ ├── tui-session.ts # pi agent session wiring │ ├── frontend.ts # browser terminal UI │ ├── sqlite-tools.ts # SQLite-backed file tools │ ├── published-workers.ts # publish file-backed Workers │ ├── cron-tools.ts # cron job tools + schedule parsing │ └── pi-fork/ # worker-specific terminal/runtime shims ├── wrangler.jsonc ├── vitest.config.ts └── wrangler.test.jsonc Routes: / → creates a new session and redirects to/s/ /s/ → serves the terminal web UI/ws/ → websocket for terminal I/O/w// → published session Worker endpoint Each session lives in a SQLite-backed Durable Object. It stores: - message history - pi session state - persistent files - published Worker metadata - cron job metadata The execute tool uses Cloudflare's Dynamic Worker Loader to run user code in isolated workers. Supports: - local relative imports from the session filesystem - package imports resolved through esm.sh - outbound fetch() through a loopback outbound entrypoint in the same Worker The agent can expose a file as a real HTTP Worker endpoint. Example route: /w// Published Workers are cached by dependency version, so they are not rebuilt on every request. Cron jobs are stored in the Durable Object and scheduled with alarms. Each run: - sends the stored prompt back into the same session - computes the next matching run time - re-arms the next earliest alarm Current minimum interval: - every 10 minutes From the repo root: npm install Create: examples/terminal-agent/.dev.vars With at least: CF_GATEWAY_TOKEN=your_cloudflare_ai_gateway_token From examples/terminal-agent : npm run dev The predev script builds the local worker packages first. From examples/terminal-agent : npm run deploy Main config file: examples/terminal-agent/wrangler.jsonc Important values: CF_ACCOUNT_ID CF_GATEWAY_NAME AI_GATEWAY_MODEL - custom domain / routes - Durable Object migrations - worker loader binding From examples/terminal-agent : npm test There is worker-pool coverage for published Workers and dynamic loader behavior. Note: depending on local environment, the Cloudflare Vitest worker pool may emit an upstream node:os warning even when assertions pass. Ask the agent to run a file like: examples/import-test.js Create a file that exports either: export default async function(request, env, ctx) { return new Response("hello"); } or: export default { async fetch(request, env, ctx) { return new Response("hello"); }, }; Then publish it with the worker publishing tool. Example schedule: */10 * * * * This runs every 10 minutes in UTC. - Sessions are persistent across reconnects. - Published Workers are session-scoped. - Cron jobs are session-scoped. bash is intentionally not exposed in this environment.- The frontend branding and Open Graph metadata live in src/frontend.ts . This example is actively experimental, but it is real and deployable.