BePay AI is the operating layer that lets autonomous agents run real payment operations — reconciliation, FX, settlement — inside hard policy guardrails, with every decision written to a tamper-evident audit chain. Agents can also mint real wallets and raise real payment intents on BePay Money test rails through the same MCP connection. Humans: open the app for your own team of agents with free models included. Software: connect any agent over MCP.
The BePay AI app is your personal workspace: create bots, give them real work, connect MCP servers and chat apps, attach computers, and run on free models included with every account — no API key needed to start. Every user gets their own private space: their own bots, chats, files, memories and model connections, isolated from everyone else. Below is the live demo of the engine — the app is where it becomes yours.
https://sidai.live/bepay/app/Private space · first bot on signup · free models · MCP connectors · computersSign up → first bot is created for you → chat immediatelyThe loop is not a diagram — it is the actual execution path of every run on this page. The ring above is driven by the same live state. Each card below shows what the agent does, which system does it, and what evidence lands in the audit chain.
When a draft crosses the $10,000 auto-settlement envelope — or any hold rule fires — the entire loop pauses. A human decision is then written to both layers: the payment rail itself (approve / reject on the draft) and the control-plane mission (approve_mission). Nothing proceeds on one signature alone.
The agent reasons and decides how; the policy engine — six declarative rules with hard thresholds — decides whether. Rule evaluation is deterministic code, not model output. That is the separation that makes agent autonomy auditable.
A credible agent-economy demo has to say exactly what is real. Here is the honest split — the same honesty an operator needs before trusting an agent with money.
Evaluated in resolution order BLOCK > HOLD > FLAG > PASS — worst verdict wins.
Everything the agent does is appended to a hash chain: hashn = SHA-256(eventn + hashn-1). The numbers below are read live from this server — and you can recompute them yourself via the /api/audit/verify endpoint or the bepay.get_audit_proof MCP tool.
BePay AI exposes its governed operations as real MCP tools — JSON-RPC 2.0 over Streamable HTTP. Claude, GPT agents, Cursor, or your own runtime can start governed runs, watch policy decide, record approvals and pull cryptographic proof. No client-side keys: the reasoning layer is managed server-side. Human looking for the app? It is right above — log in there.
https://sidai.live/bepay/mcphttps://mcp.sidai.live/mcphttps://sidai.live/bepay/.well-known/agent-card.jsonStreamable HTTP · JSON-RPC 2.0 · session id issued on initializeEvery tool wraps a live gateway function. Sensitive tools are rate-limited per client.
A built-in MCP client. You are acting as the external agent: pick a tool, send a real JSON-RPC call to /mcp, inspect the raw response.
// pick a tool and send — this talks to the same MCP endpoint external agents use
Pick a scenario and press run. The reasoning layer analyzes, the policy engine governs, missions are admitted on the control plane, money drafts on the rails — and if policy holds it, you are the human gate. Every event streams live and lands in the chain.
Understand → Plan → Govern → Act → Approve → Execute → Prove.
Every reasoning call, policy decision, money movement and approval lands in a tamper-evident SHA-256 audit chain.