# MCP gateway limits

> Argument and result sizes, time limits, per-key and per-server rate limits, sync intervals and the request allowance for tool calls through the MCP gateway.

Source: https://infro.io/docs/mcp-gateway/limits

> **Note:** INFRO is in private early access. Running these examples requires an early-access account, an active trial or subscription entitlement, and your own provider connection. Self-serve purchasing is not generally open. See [Early access](https://infro.io/early-access).

> **Note:** The MCP gateway is served at `https://mcp.infro.io/gateway/mcp` and `https://mcp.infro.io/gateway/<slug>/mcp`, on the Business plan and above.

| Limit | Value |
| --- | --- |
| Arguments of one call | 256 KB. Larger is refused with `invalid_tool_arguments`. |
| Result of one call | Per server, 1 KB to 5 MB; 1 MB unless an admin sets it. Larger is refused with `tool_result_too_large`. |
| Time limit | Per server, 1 to 300 seconds; 60 seconds unless an admin sets it. Past it, `upstream_timeout`. |
| Calls per key | The key's existing rate limit, shared with its model calls: one budget per key for `/v1` and the gateway. Past it, `429` with `Retry-After`. |
| Calls per server | Per server, 60 to 50,000 a minute; 300 unless an admin sets it. Counted across every key, to protect the server. Enforced at the nearest rate tier at or below the setting (60, 100, 200, 300, 500, 1,000, 2,000, 5,000, 10,000, 20,000, 50,000). Past it, `429 mcp_server_rate_limited` with `Retry-After`. |
| Credential | Up to 8 headers, each value up to 8 KB. |
| Tool list re-read | Every 15 minutes, or at the server's `ttlMs` when it asks for longer, at most every 6 hours; and on **Sync**. |
| Tool names | Exposed names up to 64 characters (see [naming](https://infro.io/docs/mcp-gateway#naming-on-the-aggregated-endpoint)); slugs 1 to 12 lower-case letters or digits. |
| Captured contents | Up to 256 KB per call when capture is on, kept for the organization's prompt-logging retention. |
| Request allowance | Tool calls do not count against it and are never priced. |

## Protocol

Both endpoints speak Streamable HTTP without sessions. They serve MCP revision 2026-07-28 (`server/discover`, per-request `_meta`) and the earlier `initialize` revisions, so current clients work either way. The tool list carries a 60-second `ttlMs` and is private to the key. When a server streams progress before its result, the gateway streams it to the agent and holds the final result until the result checks have run; an agent that disconnects cancels the call on the server. Toward your servers INFRO speaks either the current revision or the 2025 revisions over Streamable HTTP, detected when the server is registered and on each sync.