Skip to main content

MCP for AI agents

Ishtaran has an official MCP (Model Context Protocol) server for AI agents. If you're building with Claude, Codex, Cursor, or any other MCP-compatible client, connect it before you write a single line of integration code — it teaches the agent Ishtaran's real capabilities, plans an integration, and validates a proposed plan, grounded in the actual public API contract and SDKs rather than the model's own guesses.

LLM thinks  ->  MCP teaches/plans/validates  ->  SDK executes the integration  ->  API executes the business

@ishtaran/mcp never executes financial operations — no deposits, settlements, payouts, withdrawals, or refunds. It only knows, plans, and validates. Your integration still runs on one of the official SDKs; the MCP just makes sure the agent building it doesn't hallucinate an endpoint, SDK method, error code, or webhook event that doesn't exist.

Install​

No API key or environment variable is required — every tool and resource works fully offline, right after install.

Claude Code​

Add to your project's .mcp.json (or run claude mcp add):

{
  "mcpServers": {
    "ishtaran": {
      "command": "npx",
      "args": ["-y", "@ishtaran/mcp"]
    }
  }
}

Cursor​

Settings → MCP → Add new MCP server, same command/args as above.

VS Code (or any other stdio-based MCP client)​

Point the client at the command npx -y @ishtaran/mcp with no arguments.

What it knows​

  • Capabilities and operations — every public API operation, with the real per-language SDK method for TypeScript, Java, Python, and Go, extracted directly from each SDK's source.
  • Errors — meaning, retryability, and remediation for every documented error code.
  • Webhooks — the full delivery contract (headers, HMAC signature formula, retry/ordering rules) and the real event catalog.
  • Recipes — full integration patterns (starting with the Mercatto marketplace example), including invariants and known anti-patterns.
  • Known gaps — what the platform genuinely doesn't support yet, so an agent doesn't invent a workaround for something that was never possible.

Ask it things like "How do I integrate Ishtaran?", "Build a marketplace using Ishtaran", "How do I verify an Ishtaran webhook?", or "Explain NETWORK_EXECUTION_FEE_INSUFFICIENT_BALANCE."

Staying current without reinstalling​

The MCP's own package version and its knowledge (docs, API surface, recipes, reference projects) are versioned independently. New docs, routes, capabilities, recipes, or reference projects reach an already-installed MCP through a small, signed, content-addressed knowledge update it checks for in the background — you don't need to run npx @ishtaran/mcp again or reinstall anything to get them. See the repository README for exactly how that works.

Source​

@ishtaran/mcp is Apache-2.0 and public: github.com/taylorjeftedasilva/ishtaran-mcp · npm