Repository navigation
The first public release of AgentForEach: the open-source brain for personal AI agents. It runs one agent for each of your users, on a hyperscale serverless architecture where an idle user costs only storage.
users.forEach(user => agent(user))
Highlights
- One agent per user with long-term memory, reminders and heartbeats, human-in-the-loop forms and approvals, skills, and an optional private code sandbox
- Channels: web and apps over Web PubSub, Telegram and WhatsApp, with identity linking and pairing
- Serverless on Azure: Functions (Flex Consumption), Durable Functions, Cosmos DB and Web PubSub, created by one Pulumi program
- Models: OpenAI (Responses API), Azure OpenAI, Anthropic and OpenAI-compatible providers, with failover
Measured
On a fresh stack in Central India (method and raw results):
- 3,000 of 3,000 turns completed with 1,000 users arriving over 3 minutes; a message accepted in 0.12 s p50
- 4,799 of 4,800 turns completed with GPT-5.6 Luna; about $1.72–2.19 of model cost per 1,000 turns
- Not yet measured: sandboxes, channels, more than a few hundred users active at once, multi-region
What it costs
About 1–2¢ of platform cost per user per month, with model tokens most of the bill. The estimate for 10,000 to 1,000,000 users and a script to run your own numbers: What it costs.
Before you deploy
This is a preview. Read SECURITY.md and Getting started: a fresh stack returns 401 until you configure how users sign in.
Created by Mohit Garg. Apache-2.0.