Skip to main content
@ovra/mcp is the canonical way to give an LLM payment capabilities. It implements the Model Context Protocol and exposes Ovra’s full surface — 19 tools organized in three groups — over Streamable HTTP and stdio. Tools are designed for LLMs, not humans. Each tool uses an action parameter to multiplex operations (e.g. ovra_card { action: "issue", ... }), descriptions are written for model consumption, and required-field gating uses JSON Schema if/then so the model gets clear errors.
Sandbox-only today. Use a sk_sandbox_* or sk_test_* key. Live mode opens in v1.3+.

​
Transports

Streamable HTTP

Remote, no install needed. Best for production. https://api.getovra.com/api/mcp

Stdio

Local subprocess via npx @ovra/mcp@latest. Best for desktop clients.

​
Supported clients

Anything that speaks MCP: Claude Desktop, Claude Code, Cursor, Kiro, OpenAI Agents SDK, LangGraph (via langchain-mcp-adapters), CrewAI, custom clients.

​
The 19 tools

​
Agent tools (9) — what an agent calls directly

​
Money-movement (6) — wallets, vendors, workflows

​
Admin (4) — configuration, cross-cutting

See tool reference for every action and parameter.

​
Two-mode payment flow

For fine-grained control, use ovra_intent + ovra_credential directly.

​
Security

  • Response redaction. Every tool response goes through a redact pass — PAN, CVV, API keys, webhook secrets are masked before they enter model context.
  • Confirmation gates. Destructive actions (card.close, gdpr_delete, token_revoke) require explicit confirm: true.
  • Input validation. Required parameters enforced via if/then. IDs sanitized against path traversal.
  • URL validation. Webhook and execute URLs must be HTTPS, SSRF-checked.
  • Tool annotations. Every tool carries readOnlyHint, destructiveHint, idempotentHint metadata so MCP-aware clients can warn users.

​
Sacred guarantees through MCP

  • PAN/CVV physically cannot reach the model — server-side redaction plus the credential boundary.
  • requireIntent enforced server-side; an ovra_pay call without an approved intent returns a structured error, never a charge.
  • Strict agent isolation — at_* tokens see only their own agent’s data.

​
Next

Setup

Install and configure for your client.

Tool reference

Every tool, every action, every parameter.

Claude Desktop

The simplest possible client.

OpenAI Agents

Hosted MCP integration.