New section: MCP servers for AI agents
Docs now cover the Ankr MCP servers: a read-only data plane that gives an AI agent chain data already decoded and compressed for a context window, and a control plane that manages your account, keys, usage, and notifications from an MCP client.
Docs
Two new pages under Node API:
- Agent RPC MCP: the data plane at
https://mcp.ankr.com/rpc. Connect it from Claude Code, Cursor, or VS Code with your Ankr API key and get seventeen read tools: transactions, logs, and blocks ABI-decoded, wallet and token data from the Advanced API, and a generic read escape hatch. Responses ask for TORPC tier 2, typically 25% to 58% fewer tokens on decode-heavy reads, and the page is explicit that the tier is negotiated per call: when a response is too large, or the method is one the proxy does not compress, it answers at tier 0 and_meta.tierreports the tier actually applied. - Management MCP: the control plane at
https://mcp.ankr.com/mcp. Keys and per-key security, usage and spending, balance and payments, notifications. Login is a real browser sign-in, and deleting a dedicated API key and revoking a Platform API key need an out-of-band human approval on a page that names the action, what it affects, and the account; the other writes execute on the first call, and each tool says which it is.
Worth checking on your side
- Nothing changed on the RPC service itself. No method limits, error codes, or rate limits moved. The MCP data plane calls the same endpoints with the same key, so your plan's limits and charging apply exactly as they do to your own calls, and there is no separate MCP quota.
- The data plane cannot write. Every tool reads. The generic
rpcCallis a write denylist: it refuses transaction broadcast, signing and node administration on every chain family, EVM and non-EVM alike, and forwards everything else for the chain's own schema and your tenant to answer. Signing stays with your own wallet.