> ## Documentation Index > Fetch the complete documentation index at: https://glide-9da73dea.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Agent banking > Let Claude, ChatGPT, or a Vertex agent operate a scoped Glide sub-vault. You set the policy; the agent works inside it. Agent banking lets you give an AI agent — Claude Desktop, ChatGPT Apps, Google Vertex, OpenClaw, Hermes — the ability to act on a small, scoped portion of your Glide account. The agent can see balances, simulate payments, draft transfers, and request your approval for things you've told it to ask about. You stay in control. The agent operates inside a **policy envelope** you configure once: per-transaction caps, daily caps, counterparty allowlists, step-up thresholds. Anything that crosses a threshold pauses for your explicit approval. ## Why this exists You've already got AI doing useful work in plenty of places. Drafting emails, summarizing documents, navigating spreadsheets. The frontier is letting AI act on real systems — but for money, "real" has a much higher bar. Glide's answer is a clean separation: 1. **The agent** sees what you've scoped it to see and proposes actions. 2. **The policy envelope** is a contract you've signed with yourself: this is what's allowed, this is what isn't. 3. **You** approve anything that crosses the line. The agent never has unbounded access. There's no "Claude has my bank password" failure mode. ## What you can do today Connect Claude Desktop to your Glide account in under 5 minutes. Pre-built skills for AP, treasury, trip budgets, and more. Caps, allowlists, step-up thresholds, kill-switch. Every tool call is signed and logged. Watch live or replay later. ## How it works under the hood Glide implements the **Model Context Protocol (MCP)** — the open standard Anthropic shipped for tool-using AI agents. When you install a skill, Glide sets up: * A **scoped sub-vault** — a piece of your account with the assets and limits the skill needs. * An **OAuth grant** — the agent gets a short-lived token (max 60 minutes) that's bound to that sub-vault. The token can't reach the rest of your account. * A **policy envelope** — the contract you signed. The MCP server enforces it on every tool call. * A **receipt** — every tool call appends a tamper-evident row to your audit log. The pieces are all open standards: OAuth 2.1 with RFC 7591 (dynamic client registration) + RFC 8707 (resource indicators), JWT grants, MCP for the wire protocol. The pattern works the same with Claude Desktop today and will work the same with any future MCP-compliant runtime. ## What this isn't * **Not an autonomous bot.** The agent doesn't move money on its own. Anything past your envelope thresholds asks you first. * **Not an open spigot.** The agent can't see beyond its sub-vault. It can't see other accounts, other skills, other tenants. * **Not a custody change.** Your money still lives in segregated accounts at our banking partners. The agent moves the same money you'd move; it just does it through a constrained API. ## Next * [Quickstart](/agents/quickstart) — connect Claude Desktop. * [Skills catalog](/agents/skills/catalog) — what's available now. * [Policy envelope](/agents/policy-envelope) — how the contract works.