Guides

Writing playbooks

Create repeatable Ramp agent workflows and share useful ones with the community.

Playbooks are reusable instructions for agents. A good playbook tells the agent what Ramp surface to use, what information it needs from the user, which safety checks to run, and when to stop for human confirmation.

Use playbooks when a workflow is valuable enough to repeat: month-end spend review, vendor renewal checks, policy questions, trip cleanup, or an Agent Card purchase.

What makes a good playbook

  • It has one job. "Review SaaS spend for renewal risk" is better than "manage finance."
  • It names the Ramp surface. Tell the agent whether to use MCP, the Ramp CLI, Agent Cards, or a specific connector.
  • It declares prerequisites. Include required permissions, scopes, funds, connector setup, CLI auth, and whether demo data is acceptable.
  • It asks for missing inputs. Vendor, amount, time range, fund, billing address, memo, category, approver, or any other workflow-specific detail.
  • It includes stop conditions. Agents should pause before payments, irreversible writes, unsupported flows, or unclear policy decisions.
  • It produces a useful artifact. Ask for a summary, decision, receipt plan, command output, or list of follow-up actions.

Template

Start with this shape and adapt it to your runtime.

# [Playbook name]
 
## Goal
[One sentence describing the outcome.]
 
## Runtime
- Ramp surface: [MCP | Ramp CLI | Agent Cards | connector name]
- Environment: [demo | sandbox | production]
- Required setup: [connector installed, `ramp auth login`, active fund, scopes]
 
## Inputs
- [Input the user must provide]
- [Input the agent should ask for if missing]
 
## Instructions
1. Confirm the authenticated Ramp user and environment.
2. Gather the required inputs.
3. Use Ramp to retrieve only the data needed for the task.
4. Summarize what you found before taking any write or payment action.
5. Stop for confirmation when the workflow crosses a safety boundary.
6. Complete the workflow only after confirmation.
 
## Safety checks
- Do not expose secrets, card numbers, tokens, or private employee data.
- Do not invent Ramp tool names. Use available MCP tools or CLI help.
- Do not continue if the user lacks permission or the policy answer is unclear.
- Do not make purchases that require unsupported flows like 3DS or card-on-file.
 
## Output
[Specify the final answer format.]

Example: vendor spend review

Use Ramp to review spend for [VENDOR] over [TIME RANGE].
 
First confirm which Ramp account and environment you are using. Then find recent
transactions, cards, funds, and vendor records related to the vendor. Summarize
total spend, owners, unusual changes, missing receipts or memos, and any renewal
risk.
 
Do not make changes in Ramp. If you find missing context, ask me before
continuing.
 
Return:
- Spend summary
- Owners or likely owners
- Risks or anomalies
- Recommended next actions

Example: Agent Card purchase

Use Ramp to make this purchase with an Agent Card:
 
Vendor: [VENDOR]
Item: [WHAT TO BUY]
Expected total: [AMOUNT]
Fund: [FUND NAME]
Memo: [BUSINESS PURPOSE]
Billing address: [ADDRESS]
 
Before generating credentials, confirm the fund, amount cap, merchant, and
business purpose with me. Stop if the merchant requires 3DS, card-on-file billing,
or another unsupported flow. After checkout, summarize the result and tell me how
receipt handling should be completed for this runtime.

Verify before sharing

  • Test with https://demo-mcp.ramp.com/mcp or Sandbox when possible.
  • Check exact MCP tool names against the MCP tools reference or use ramp --help for CLI workflows.
  • Remove secrets, real card data, customer data, and internal-only details.
  • Include the permissions and scopes the workflow needs.
  • Make sure the playbook tells the agent when to ask for confirmation.

Share it in Slack

Built a Ramp agent workflow that others should copy? Join the Ramp developer community and share it there.

The most useful posts include the playbook, the runtime you used, what it helped you do, and any safety checks you added.