🐈 mewagent - a personal AI agent on Cloudflare Workers

Introduction

I wanted a personal AI agent I could message from my phone: something that remembers things about me, can fetch a page, update a daily note on this site, and otherwise sit quietly until I need it.

There are already projects in this space (OpenClaw, Hermes, and a pile of "personal agent" scaffolds). I like the shape: channels in, tools + memory in the middle, model out. What I didn't want was a VPS to babysit, multi-tenant auth, or a bot that talks to whoever finds its username.

So I built mewagent: an OpenClaw/Hermes-style agent on a Cloudflare Worker + Telegram webhook, OpenRouter free models, long-term memory in Workers KV.

Why bother building another one?

The laptop is not a server

Most personal agents assume a process on your machine or a cheap VPS: long-poll Telegram, keep the daemon alive, hope the laptop is open when you message from the bus. Fine for hacking; a poor always-on story.

Cloudflare Workers give you an HTTPS endpoint, secrets, and KV without provisioning a box. Telegram posts to the Worker; the agent runs; you get a reply. No long-poll. No "is my gateway still running?"

Free models are good enough to start

LLM calls go through OpenRouter, defaulting to free models (openrouter/free and other :free variants). For notes, context, and light tool use, that's often enough. When it isn't, you change a model string, not an architecture.

Optionally route through Cloudflare AI Gateway for analytics, caching, and rate limits. Same code; different base URL.

The shape of the system

Telegram ──POST /telegram──► Worker ──► agent ──► [AI Gateway?] ──► OpenRouter
                               │
                               └── MEMORY (KV): notes + chat sessions

The agent loop is simple: load session, build a system prompt (persona + memories + optional project playbooks), call the model with tools, run tools for a few rounds, save the reply. Slash commands (/help, /memory, /reset, /whoami, /model) skip the LLM.

Allowlisting happens at the edge. Worker tools stay Workers-safe: memory, web fetch, datetime, optional GitHub. No shell, no local filesystem.

Why Telegram

For a personal agent, messaging friction matters more than feature checklists.

Telegram gives you a real webhook, private chats, and a phone UI everyone already has. Groups are ignored. BotFather can disable groups so the bot never lands in a public chat by accident.

Security

Allowlist

Worker + KV for memory and sessions. Secrets: OpenRouter, Telegram bot token, webhook secret.

With ALLOWED_TELEGRAM_IDS empty, the bot only answers commands like /whoami - no agent, no tools. DM it, copy your numeric id, set the allowlist, redeploy. After that, only listed ids get the agent.

Webhook authenticity

TELEGRAM_WEBHOOK_SECRET is required. Deploy the Worker, then Telegram's setWebhook to /telegram with the same secret_token. Missing secret → /telegram returns 503. Timing-safe compare. Random POSTs to a guessed path don't drive the agent.

GitHub with least privilege

Optional: fine-grained PAT + GITHUB_REPOS allowlist so the agent can edit repos via the GitHub Contents API. PAT scoped to Contents only on those repos; the allowlist is a hard check in code as a second belt.

For this site: I message the bot to add a daily entry, it reads src/daily/YYYY-MM-DD.org, appends a bullet, commits, replies with the URL. CI builds; Cloudflare Pages deploys. A system-prompt playbook keeps org-mode style.

What I use it for

• Talk to a model from Telegram without opening a browser
• Durable notes across sessions (preferences, project context) in KV
• Append to the public daily log without a laptop editor
• Occasional web fetch for a quick summary

Not a desktop coding agent. The always-reachable assistant in my pocket.

Tradeoffs vs OpenClaw / Hermes-style daemons

Smaller on purpose. Want multi-channel daemons, skills marketplaces, rich memory graphs? Use a bigger project. Want a single-owner bot on the edge that can edit your git-backed site? This shape has been pleasant.

Limits worth knowing

• Free models + Workers free tier are fine for light use. Tool loops waiting on LLMs benefit from a paid Workers plan (CPU / wall time). Mid-reply deaths can trigger Telegram webhook retries.
• KV is not a query engine. Note and session keys are enough for now; D1 if list/query needs grow.
• Cron, Discord, Durable Objects per user, skills-as-SKILL.md: roadmap, not promises.

Closing

Personal agents are having a moment. The interesting work is less "which model" than where the loop runs, who can talk to it, and what tools you trust on that runtime.

mewagent is my answer for now: Workers for presence, Telegram for reach, OpenRouter for models, KV for memory, GitHub for the few write paths I care about, fail-closed so a discoverable bot username is not an open invitation.

If you build something similar, or still prefer a VPS daemon, I'd like to hear about it.