Say it once.Your agents take it from there.

Tiny Agents is a messenger where every contact is an AI agent. They remember you, work in their own browser and computer, and ask before anything important.

Self-hosted · Node and SQLite · Your model keys stay in your gateway

50 seconds: memory, browser and computer use, approvals, permissions, a terminal, MCP, alerts and a team of agents.

How it works

One message. The work gets done.

Pick an example. You say what you need in plain words; the agent plans, uses the tools you've allowed, and checks with you before anything important.

Chief of Staff Online
Every weekday at 9, brief me on my calendar and anything urgent in email.
Created a routine · Weekdays at 9:00 Reads your Calendar and Email over MCP
Monday, 9:00 AM
Good morning. Three meetings today; the 2 PM design review moved to 4. One urgent email: the venue wants the deposit by Wednesday. Want me to draft a reply?
Features

Built to act, and to ask first.

Each agent has its own persona, model and tools. You decide what each one may do on its own.

Memory

Remembers your preferences across chats. You can read and edit everything it knows.

Browser use

Opens real websites in its own browser. Watch it live, or take over with one click.

Computer use

Each agent gets its own desktop on a sandbox machine, never yours, and can run commands there.

Approvals

Reading is automatic. Anything that spends money or changes something waits for your yes.

Vault

Passwords, keys and cards go straight into the page or request. The agent never sees them.

Permissions

Set every tool to Allow, Ask or Off, per agent.

MCP

Connect calendars, email, docs or any remote MCP server, and choose which agents use it.

Routines

Set up recurring work in plain words. Agents message you first when something changes.

A team

Agents hand work to each other. Mention one with @ and it picks up the job.

Terminal

Open a shell on the same machine your agents use. Tabs keep running when you leave.

Alerts

Push notifications when an agent needs approval, has a question or finishes a routine.

Skills

Reusable instructions any agent loads when a task needs them.

Self-host

Runs on your server.

One Node process and two SQLite files. Conversations are durable: refresh mid-reply, restart the server or redeploy, and the reply picks up where it left off.

  • RuntimeNode 24, one process
  • StorageSQLite: app.db and agent.db
  • ModelsClaude, GPT and more via Cloudflare AI Gateway
  • Provider keysStay in the gateway
  • Sign-inPassword or passkey
$ pnpm install
$ pnpm dev
# open http://localhost:3000

# real inference: point it at your gateway
$ export CLOUDFLARE_ACCOUNT_ID=…
$ export CLOUDFLARE_GATEWAY_ID=…
$ export CLOUDFLARE_API_KEY=…

Your team of always-on AI agents.

Message them like a friend. They take it from there.

Watch the film