Floe Labs
The layer that meters what an AI agent spends, and stops it before it spends more than you allowed.
- Role
- Founding Software Engineer
- Timeline
- Aug 2025 to Present
- Live site
- Visit the live site
- users onboarded
- 400
- embedded wallets, no seed phrase
- 746
- models routed across 20 providers
- 125
- MCP tools in the official registry
- 80
Overview
Payments and metering infrastructure for AI agents. The API meters every call an agent makes and enforces its spend limits before money moves, and an OpenAI-compatible gateway in front of it routes 125 models across 20 providers. I built the API, the console, the gateway, the voice platform, the CLI, and the deploy pipeline.
The problem
An AI agent that calls paid APIs has no way to pay for them and nobody watching what it spends. Every provider bills separately, so a team running agents across twenty vendors reconciles twenty invoices by hand and learns what a run cost only after the money is gone.
What I built
Architected the core payments and metering API in TypeScript, Node, and Postgres: 313 REST endpoints under 2,193 automated tests, with spend limits checked before funds move. In front of it sits an OpenAI-compatible LLM and voice gateway routing 125 models across 20 providers and 6 modalities, which collapses twenty vendor invoices into one real-time bill. A Next.js console adds 39 routes of agent management, spend policy, billing, and webhooks. Signup runs on Privy and provisions embedded wallets behind email or Google login, so nobody handles a seed phrase. Real-time voice runs on a Node WebSocket media gateway that meters each call leg and halts a call breaching budget mid-conversation. Stripe carries the monetization: four subscription tiers, Connect-based client invoicing, and daily usage reporting into Billing Meters.
Results
400 users and 746 wallets onboarded without anyone touching a private key. Four developer packages shipped to npm and PyPI, including a zero-dependency CLI of 33 commands and an MCP server exposing 80 LLM-callable tools, listed in the official MCP Registry. Six CI/CD workflows deploy the production estate with post-deploy verification that blocks a bad release.
Decisions
Check the limit before money moves
An agent with nobody watching learns what a run cost only after the money is gone. Checking first means a call over budget never goes through.
One gateway, one bill
Every provider bills separately. Putting all 20 behind one OpenAI-compatible gateway turns twenty vendor invoices into a single real-time bill.
No seed phrases
Signup provisions embedded wallets behind email or Google login, so nobody has to handle a private key.
Deploys that check themselves
Post-deploy verification blocks a bad release instead of reporting false success.
Stack
- TypeScript
- Node.js
- Next.js
- PostgreSQL
- Stripe
- Privy
- WebSockets
- MCP
- Docker
- GCP
This is the work behind the Founding Software Engineer role at Floe Labs. Built by Raj Bhensdadiya, an AI & full-stack engineer in Montreal. The full work history and the other projects have more, and email is the fastest way to reach me.