# Refunds for AI agents over MCP > How AI assistants can look up refund cases and propose refunds over the Model Context Protocol while policy and people keep the final say. - Canonical URL: https://lumtry.com/guides/mcp-refunds-for-ai-agents - Last updated: 2026-09-24 Lumtry exposes refund and dispute cases to AI agents through a Model Context Protocol server. Agents can read cases with a scoped API key and, where enabled, file proposals such as a decision or dispute evidence. Agents never move money: every proposal waits for a person to approve it in the dashboard, in Slack or by phone. ## What can an agent do through Lumtry's MCP server? Read and propose. Read tools list and fetch refund cases and dispute cases. Proposal tools let an agent suggest a decision, attach dispute evidence or suggest cancelling a case. Agents cannot open new cases. ## Can an agent issue a refund? No. Every write is a proposal that a person decides in the dashboard, in Slack or by phone. An agent can never set where money goes, override an amount or mark a manual resolution complete, and those requests are refused outright. ## How do agents sign in? Read access uses a scoped API key. Proposal access uses OAuth 2.1 with PKCE, exact redirect URIs and tokens bound to both the client and the person who delegated access. An agent's scopes never exceed that person's role, and revoking the client stops it on its next request. ## How is agent text kept from steering the system? Anything an agent writes is stored as untrusted data attributed to that agent, and it is never read as an instruction. Every tool call and every rejected sign-in leaves an audit row. ## Connect an agent in five steps 1. **Check your plan**: Reading cases with an API key works on any plan. Proposal tools need the External Agent Access add-on on Starter or Growth, or the Scale plan, and are turned on for your workspace on request. 2. **Create a key or an agent client**: For read access, create an API key with only the scopes the agent needs. For proposals, add the agent under Settings, Agent clients, with its exact redirect URI. 3. **Add Lumtry to the agent's MCP configuration**: Point the agent's MCP client at Lumtry's server and sign in with the API key or the OAuth client you created. 4. **Let the agent file proposals**: The agent reads a case and files a proposal with its reasoning. Approvers see the proposal attributed to the agent that made it. 5. **Decide and follow up**: A person approves or rejects the proposal in the dashboard, in Slack or by phone. The agent can poll the task status, or receive a signed callback on a host you allow. ## Frequently asked questions ### Can I limit how often an agent calls Lumtry? Yes. Each agent client has its own rate limit, and the lower of the client limit and your plan limit applies. ### Can an agent's proposal skip my refund policy? No. An approved proposal runs through the same workflow as any other decision, with the case pinned to its policy version and re-checked against its live state before anything executes. ### Can I see what an agent did? Yes. Every tool call leaves an audit row naming the agent client, and each proposal shows who approved or rejected it. ### Which agents can connect? Any MCP client that can sign in with an API key, or with OAuth 2.1 and PKCE for proposals. The server does not depend on a particular AI vendor. ## Related - [Home](https://lumtry.com/index.md): Decide refunds by a versioned policy, send the exceptions to your team in Slack or the dashboard, and keep every step on the record. Start free. - [Guides](https://lumtry.com/guides.md): Answer-first guides to refunds, returns, chargebacks, approvals and audit trails for Shopify, WooCommerce and BigCommerce merchants. - [Slack approvals for Stripe](https://lumtry.com/guides/slack-refund-approvals-for-stripe.md): How to route Stripe refunds that need a person to Slack, approve them from the message, and keep the dashboard as the source of truth. - [Voice approvals](https://lumtry.com/guides/voice-refund-approvals.md): How outbound phone calls let an approver confirm a held refund when they are away from Slack and the dashboard, with the answer recorded. - [Return abuse screening](https://lumtry.com/guides/return-abuse-screening.md): How to score refund requests on claim history, missing returns and unusual values, and hold risky cases for a person without slowing honest customers. - [Chargeback evidence](https://lumtry.com/guides/chargeback-evidence-automation.md): How to assemble dispute evidence from orders, policy and the refund audit trail so every chargeback response starts from a complete file. - [Refund audit trail](https://lumtry.com/guides/refund-audit-trail.md): What a refund audit trail should record, why it must be append-only, and five steps to build one that holds up in disputes and audits. - [Retention offers](https://lumtry.com/guides/retention-offers-before-refund.md): When to offer store credit, an exchange or a partial refund before paying money back, and how a policy decides which offer fits a case. - [Marketplace refunds](https://lumtry.com/guides/multichannel-marketplace-refunds.md): How to bring store and marketplace refunds from Amazon and eBay into one queue, with a policy per channel and every decision in one audit trail. - [Pricing](https://lumtry.com/pricing.md): Lumtry plans by case volume and seats, from Free to Enterprise, with add-ons and capped outcome fees. Start free or hear when billing opens. - [MCP server](https://lumtry.com/developers/mcp.md): Connect an AI assistant or your own agent to your refund workspace through the Model Context Protocol. Read cases and propose actions, never move money. - [Trust](https://lumtry.com/trust.md): How Lumtry secures merchant data, keeps AI inside a deterministic refund policy, lists every sub-processor and explains data flow and retention. Review it. - [llms.txt](https://lumtry.com/llms.txt)