MCP Server
Let Claude, Cursor, and other MCP clients work against your live Snippe account
If you work in an MCP client — Claude Code, Claude Desktop, Cursor, VS Code — you can connect the Snippe MCP server and let the agent work against your live account: check whether a payment cleared, look up a reference, read your balance, or hand a customer a payment link, without leaving the chat.
This is the runtime counterpart to the Agent Skill. The skill teaches an agent how to write a Snippe integration. The MCP server lets it call Snippe on your behalf.
Connect a Client
Keys, headers, and config for Claude Code, Claude Desktop, and Cursor
Tools & Prompts
Everything the agent can call, and the playbooks that guide it
Self-hosting
Install, run, and put it behind a proxy
Console and Checkout
One process serves two MCP servers, split by audience:
| Server | Endpoint | Audience | Tools | Prompts |
|---|---|---|---|---|
| Snippe Console | /console/mcp | You — manage the whole account | 10 | 6 |
| Snippe Checkout | /checkout/mcp | Your customers — collect payments | 6 | 4 |
The split is the point. An agent pointed at Checkout can create payments and links but can never see your balance, your payment history, or your session list — so it's safe to put in front of customers. You don't run two deployments; you pick the endpoint that matches what the agent is for.
What you can do
Once connected, your client can:
- Check the status of a payment by reference, or browse recent ones.
- Read your available balance before you commit to a payout.
- Create hosted checkout sessions — fixed amount, itemized cart, or pay-what-you-want — and hand back a payment link.
- Look up a session, or cancel one that hasn't been paid.
- Charge a customer by mobile money, so they get a USSD push to approve on their phone.
The server also ships MCP prompts — short playbooks your client can invoke directly, so common jobs don't need re-explaining. See Tools & Prompts for the full list.
Disbursements aren't exposed over MCP. Use the Disbursements API or an SDK to send money out.
Keeping it safe
Some of these tools move real money. Treat the key you give the server as production credentials.
- Match the server to the job. Use Checkout wherever you only need payment flows — it can't read your account.
- Use a link key for customer-facing agents. A
plk_key can create links and nothing else, so a prompt-injected agent has nowhere to go. - Keep approval prompts on. Don't allowlist tools that create payments for automatic execution.
- Don't commit keys. Read them from the environment in any config file that lives in Git.
- Keep it off the public internet. Bind localhost, or put it behind a proxy that authenticates.
Requests are rate limited at the MCP edge to 10 per second, burst 20, bucketed per key — the key is hashed, so it never appears as a bucket id or in an error. Underneath, calls still count against the Snippe API's 60 requests per minute. create_mobile_payment generates its own idempotency key per call. Request logs record the key's type — api, link, anonymous — never the key itself.
Skill or server?
Both, usually — they do different jobs.
| Agent Skill | MCP Server | |
|---|---|---|
| Gives the agent | Knowledge of how Snippe works | Access to your Snippe account |
| Reach for it when | You're writing integration code | You're checking or moving real money |
| Touches live data | No | Yes |
If you only want to query your account by hand, the Dashboard is faster. If you want to build against Snippe in code, start with an SDK.