nex-t1.ai
Production-oriented monorepo for Nex Copilot (https://nex-t1.ai): a governed AI copilot where users search, analyze, automate, and execute across wallets and 300+ tools — with read-only / approval / execution modes, spending limits, contract allowlists, human-in-the-loop approvals, and machine-readable execution receipts.
AI copilots that touch wallets, DeFi tools, and external connectors fail quickly when they treat every action as a chat completion. The hard problem is controlled delegation: users need useful autonomy without giving models direct authority over tokens, private connector credentials, or irreversible transactions.
Nex Copilot separates clients, API/BFF, agent runtime, MCP tool gateway, sandbox execution, on-chain intents, policy checks, approvals, and audit receipts. The user can stay in read-only, approval, or execution mode while every risky action passes through deterministic policy and human-in-the-loop gates.
The system has to coordinate streaming UX, resumable agent sessions, connector vaulting, sandbox artifacts, transaction simulation, signing isolation, and event ledgers without leaking secrets into model context or letting retries create duplicate runs.
The architecture treats on-chain execution as typed intent orchestration rather than raw transaction generation. It combines OpenAI Agents SDK, AI SDK streams, MCP routing, E2B-style isolation, and Supabase-backed approvals into one governed execution plane.
This page shows Stefan's depth across agent platform architecture, Web3 risk controls, MCP connector boundaries, Supabase/RLS data design, SSE run streaming, policy packages, and audit-ready execution receipts.
Nex Copilot is an execution-control system disguised as a copilot. The technical depth is in separating what the model can propose, what policy can approve, what infrastructure can execute, and what the user can audit after the fact.
Clients never get direct model, connector, or signer authority. They speak to a BFF that creates runs, streams events, and exposes approval surfaces while connector credentials and signing paths remain isolated behind server-side boundaries.
UI chat streams and multi-agent tool execution are separate concerns. AI SDK can serve app-facing streaming while OpenAI Agents SDK handles handoffs, sessions, interruptions, and MCP tool loops inside the governed runtime.
The important product primitive is the typed intent, not the generated answer. Intents pass spending caps, contract allowlists, simulation, approval, and signing gates before a receipt records wallet, fees, risk flags, approvals, and transaction hashes.
The system is designed for retry safety, run idempotency, event ledgers, sandbox artifacts, and reviewable tool traces. That makes it closer to a financial control plane than a consumer chatbot.
Primary web surface for chat, tool cards, streaming runs, and approval UI over the BFF contract.
Postgres + RLS for agent_runs, events, approvals, and artifact metadata — one persistence boundary.
Multi-agent handoffs, sessions, HITL interruptions, and MCP tool loops in agent-runtime.
UI-facing chat streams in React/Next — kept separate from the multi-agent runtime per platform boundary.
300+ tools route through tool-gateway — connectors never called directly from clients or LLM-ad-hoc fetch.
Supabase JWT + API key auth; creates agent_runs; streams agent_run_events over SSE.
Pauses risky tool or on-chain actions until explicit human approval.
OpenAI Agents SDK ManagerAgent with Postgres session adapter and structured handoffs.
MCP/function/internal tool registry with Auth0 Token Vault connector boundary.
E2B-isolated filesystem/shell/repo execution with artifact upload.
Typed intents → simulation → policy → approval → isolated signing facade.
Deterministic spending caps, contract allowlists, redaction, and kill switches.
Human-readable proof: wallet, fees, risk flags, approvals, transaction hashes.