Connect your first agent

Install a Glide skill, connect it to Claude Desktop or ChatGPT, and watch the first tool call land in your audit feed.

The fastest path to seeing agent banking work end-to-end. Takes about 5 minutes.

1. Pick a skill

   Open the skills catalog and pick one. For your first run, the Trip budget skill is a low-stakes way to see the flow — it lets Claude track travel spend against a budget you set.

   Other options:

   • AP agent — let Claude draft vendor payments from QuickBooks invoices.
   • Treasury — let Claude allocate idle USDC across yield strategies.
   • Payroll cosigner — let Claude pre-validate payroll batches you’ll approve.

   See Skills catalog for the full list.

2. Install the skill

   Tap Install on the skill page. The install flow walks you through:

   1. Pick the entity (business) or sub-vault (personal) the skill operates on.
   2. Configure the policy envelope — per-transaction cap, daily cap, counterparty allowlist, step-up threshold. Defaults are sensible; you can override anything.
   3. OAuth handshake — you approve Glide as the authorization server for this skill.
   4. Sub-vault creation — Glide creates the scoped sub-vault the skill operates on.
   5. Grant issued — the agent gets its first short-lived token.

   The whole flow is on one screen and takes under a minute.

3. Add Glide to your AI runtime

   For Claude Desktop:

   1. Open Claude Desktop → Settings → Developer → Edit Config.
   2. Add the Glide MCP server:

   {
     "mcpServers": {
       "glide": {
         "url": "https://mcp.glide.co",
         "auth": {
           "clientId": "<from-skill-install-screen>",
           "clientSecret": "<from-skill-install-screen>"
         }
       }
     }
   }
   

   3. Restart Claude Desktop. You’ll see the Glide tool in the tools menu.

   For other runtimes (ChatGPT Apps, Google Vertex, OpenClaw, Hermes), see Runtimes.

4. Talk to the agent

   Ask Claude something the skill can do. For Trip budget:

   “I’m in Lisbon for the week with a $2,000 budget. How much have I spent?”

   Claude calls Glide’s MCP server to read the relevant balances and recent transactions, computes the answer, and replies. You see the call in your audit feed in real time.

5. Watch the audit feed

   From your Glide dashboard, open Agents → Activity feed. You’ll see every tool call: which agent, which tool, the redacted input and output digests, the risk verdict, the policy version that was in force.

   See Receipts for what each row means.

What just happened

You authorized a scoped sub-vault, installed a policy envelope, and gave Claude an OAuth token bound to that sub-vault. Every Claude tool call from here on:

• Validates the OAuth grant (signature, expiry, audience).
• Re-reads your tenant from the database (so a stolen grant on a stale tenant fails).
• Evaluates the policy envelope (caps, allowlist, step-up).
• Logs a receipt to your tamper-evident audit log.

If anything crosses a policy threshold, you get prompted to approve via step-up.

Test that the controls work

Try asking the agent to do something outside the envelope. For Trip budget:

“Send $5,000 to a counterparty I haven’t allowlisted.”

Claude will see the policy block, tell you it can’t do that, and explain why (per-transaction cap, counterparty not allowlisted, etc.). The block is enforced by Glide’s MCP server, not by Claude’s prompt — even a jailbroken agent can’t bypass it.

Next

• Skills catalog
• Policy envelope
• Step-up approvals
• Runtimes