The runtime behind Bitte's on-chain agents.
Bitte's production AI runtime turns natural language into signed on-chain transactions. I was the top human committer on it — streaming LLM agent loops, OpenAPI→tool conversion, and roughly twenty on-chain primitives across four chains, serving real users on Google Cloud Run.
An Express/TypeScript service on Google Cloud Run running streaming LLM agent loops: the model plans, calls tools, and streams intermediate state to the client while transactions are prepared, simulated, and signed.
user intent ──▶ LLM loop ──▶ tool call ──▶ on-chain primitive
▲ │ │
└──────── stream ◀────────────┘
tokens + tx state, live to the client
tools: OpenAPI specs ──▶ converted to function-calling schemas - streaming agent loops LLM plans and acts with live intermediate state to the client
- OpenAPI→tool conversion any spec becomes a callable agent tool, automatically
- on-chain primitives ~20 transaction builders across NEAR, EVM, SUI, and Cardano
- NEAR Intents cross-chain swaps routed through the Intents protocol
- production ops Express + TypeScript, deployed on Google Cloud Run
#1 committer (135 of 343 commits) on the agent SDK monorepo — @bitte-ai/chat, the make-agent CLI, and the OpenAPI function-calling manifests that let anyone stand up an on-chain agent.
#3 committer (249 commits) on Bitte Wallet — the production AI-agent wallet those SDK agents plug into.
Top human committer on the runtime that powers Bitte's agent economy — and the #1 committer on the SDK that lets external builders plug into it.