A whole AI-native app that improves itself, for $5 a month.

The whole app in one deploy unit.

Serves the site, the MCP server and the webhooks from one fetch handler, plus scheduled for cron and queue for consumers. Static assets ride along for free.

export default { fetch: app.fetch, scheduled, queue }

$5: 10M requests · 30M CPU-ms · assets unlimited

Multi-step agent runs that survive failures.

Each step is checkpointed and retried on its own, so a report that makes five LLM calls repeats only the one that failed. Put one paid call per step.

await step.do("summarize", () => llm(page))

$5: 500k steps · 1 GB state

Ack the webhook now, think later.

WhatsApp gives you 20 seconds. Push the message on a queue, answer 200, and let a consumer run the slow agent loop with retries and batching.

await env.JOBS.send({ phone, text })

$5: 1M operations · then $0.40/M

The agent's heartbeat.

Build expensive snapshots ahead of visitors, prune analytics, send reminders. Same Worker, same bindings, one line of config.

"triggers": { "crons": ["*/5 * * * *"] }

$5: included

Relational data, analytics and the agent board.

SQLite with numbered migrations. Holds users, events, goals, hypotheses and tasks. Costs nothing while idle, so one database per customer is a real option.

await env.DB.prepare("SELECT * FROM goals").all()

$5: 25B rows read · 50M written · 5 GB · 50k databases

Chat threads, snapshots, dedupe keys.

Read-heavy and global. Conversation turns keyed by user and day with a TTL give an assistant long-term memory with no schema at all.

await env.THREADS.put(key, turns, { expirationTtl })

$5: 10M reads · 1M writes · 1 GB

Files, audio, screenshots, the RAG corpus.

S3-compatible storage with no egress fee, ever. Voice notes, generated images and the markdown your AI Search index reads all live here.

await env.FILES.put(`audio/${id}.ogg`, req.body)

Free tier: 10 GB · then $0.015/GB · egress $0

One live object per room, agent or document.

A single-threaded instance with its own SQLite, addressed by name. Chat rooms, presence, WebSocket hubs and per-agent sessions, hibernating when idle.

env.ROOM.get(env.ROOM.idFromName(roomId))

$5: 1M requests · 400k GB-s · 5 GB SQLite

Every model call goes through one URL.

Logs, caching, rate limits, retries and fallbacks in front of OpenRouter, Anthropic or OpenAI. Store provider keys in the gateway and the Worker holds no LLM secrets. Tag calls with metadata and the dashboard becomes your token analytics.

gateway.ai.cloudflare.com/v1/{acct}/{gw}/openrouter

Free on every plan

Speech, embeddings and small models, no keys.

Transcribe a WhatsApp voice note with Whisper, embed text with bge-m3, classify or describe an image. Runs next to your Worker, billed in neurons.

await env.AI.run("@cf/openai/whisper-large-v3-turbo", { audio })

10k neurons/day free · then $0.011 per 1k

Point it at an R2 bucket, get retrieval.

Chunking, embedding, indexing and re-syncing are managed. Upload markdown, then search from the Worker. No vector database to run.

await env.AI.autorag("docs").search({ query })

Built on R2 + Vectorize + Workers AI

Your own embeddings, when you need control.

Similarity search over vectors you compute yourself: product matching, memory recall, dedupe. Metadata filters included.

await env.VECTORS.query(embedding, { topK: 5 })

$5: 50M queried dims · 10M stored dims

Agents that can see the web.

Headless Chrome driven by Playwright or Puppeteer from inside the Worker. Screenshots, scraping and audits, results saved to R2.

const browser = await launch(env.BROWSER)

$5: 10 browser-hours · 10 concurrent

The app's API is an agent's toolbox.

A ~150-line JSON-RPC server at /mcp exposes the same functions the UI calls. Claude, ChatGPT and deco studio can operate the app, and its UI renders inside them.

{ "method": "tools/call", "params": { "name": "get_briefing" } }

Plain code in the Worker

A per-caller budget in front of every paid call.

Key it on the visitor's IP or user id. Anything that triggers an LLM call gets a limit before it reaches the model.

const { success } = await env.LIMIT.limit({ key: ip })

Binding, no extra service

Bot checks without CAPTCHAs, admin pages without auth code.

Turnstile verifies a human before a vote or a signup. Access puts Google login in front of any route; the Worker verifies the signed token.

await verifyAccessJwt(req, env.ACCESS_AUD)

Turnstile free · Access free to 50 users

Push is deploy, every PR gets a URL.

Connect the GitHub repo once. main goes to production in about 30 seconds; branches get preview URLs; rollback is one click.

git push origin main

Included

Structured logs and first-party analytics.

Workers Logs for every invocation, and a track() helper writing events to D1 so agents can query real usage over MCP. No third-party analytics script.

"observability": { "enabled": true }

$5: 20M log events