AI agents that do real work, and the layer that keeps them honest.
Website · Bhai · MELRA · Product Hunt · LinkedIn · X / Twitter
Give everyone an AI agent they can actually trust with their life admin, and build the infrastructure that makes that trust earned rather than assumed.
AI can now browse, fill forms, send messages, book things and operate computers. The hard part is no longer what an agent can do. It is:
- What is it allowed to do?
- Did it actually happen?
- Can we prove it?
We work on both sides of that problem: Bhai, a personal AI agent for everyone, and MELRA, the open-source runtime that bounds, verifies and records what agents do.
The model does the thinking. MELRA does the allowing, the proving, and the undoing.
Bhai is a personal AI agent you message on WhatsApp, Telegram or email. Then you go touch grass.
It books the flight, fights for the refund, fills the boring form, hunts down the cheaper deal and remembers the birthday before you do.
Bhai is built on three things:
- Memory that makes you feel known: your people, your preferences, what you asked last week
- Reliability on real tasks: no "done" until the work is actually done, and an honest "I'm stuck" when it isn't
- Personality: the fun, smart, dependable big brother your to-do list always needed
Every message and file Bhai sends on your behalf runs through MELRA, so "sent" means the provider confirmed it, not that the agent said so.
→ Get early access (invite-only for now) · @oyybhai on X
MELRA is XAGI's open-source runtime for governed AI execution.
It sits between agents and the systems they act on, with one control path across:
- Files
- Terminal
- Browser
- Computer use
- Memory
- External tools
Every action passes through policy checks, scoped permissions, approvals where needed, evidence-based verification, receipts and execution certificates. Outcomes that can't be confirmed are parked for inspection instead of blindly retried.
MELRA is:
- Model-agnostic
- Agent-framework agnostic
- Designed to run inside your own environment
- Open source
Try MELRA: Watch the demo · With Node.js 22.13+ or 23.4+:
npx @melra/cli@alpha setup
Bhai is MELRA's toughest customer.
A personal agent that sends your messages, signs you up for things and touches your money cannot run on "trust me". Building Bhai forces MELRA to handle the messy real world, and MELRA lets Bhai take on real work without asking you to take its word for it.
Don't trust "success." Prove the effect.
If we haven't measured it, we don't claim it.
Every change to Bhai is measured on real tasks before it ships, and negative results and failed experiments are part of the record too.
Our work spans:
- Personal AI agents and long-running task reliability
- Agent memory and personalisation
- Computer-use and browser agents
- Policy-governed agent execution
- Verifiable execution and evidence
- AI permission and authority systems
- Computer-use safety
Infrastructure that asks you to trust it should be inspectable. MELRA is open source so developers, researchers, security teams and enterprises can read the assumptions instead of trusting our marketing.
Contributions to our public repositories are welcome.
We'd love to hear from:
- AI and agent engineers
- Security researchers and red-teamers
- Agent developers who want verified actions
- Early Bhai users with tasks they hate doing
- Open-source contributors
| Website | xagilab.com |
| Bhai | bhai.xagilab.com |
| MELRA | github.com/XAGI-Lab/melra |
| GitHub | github.com/XAGI-Lab |
| linkedin.com/company/xagi-labs | |
| X / Twitter | @xagilabs · @oyybhai |
| Founders | Dheeraj S and Bhavana Sanal |
| hello@xagilab.com | |
| HQ | Kerala, India |