MCP Gateway — Stack

Technology landscape as of 2026-10-10. Knowledge cutoff May 2026.

Selection criteria

You are introducing a component that every tool call passes through, so the evaluation is dominated by what it costs you when it is wrong.

• Added latency. Milliseconds at the high percentiles for a pass-through call, measured with policy evaluation and logging switched on rather than in a bare proxy configuration.
• Blast radius and failure behaviour. Whether it can run active-active, whether a restart drops in-flight sessions, and what consumers experience when the policy store it depends on is unavailable.
• Catalogue curation. Whether offerings can be assigned to consumers selectively, renamed, and withheld, or whether aggregation is all-or-nothing.
• Policy expressiveness. Whether a decision can consider the caller, the specific offering, and the argument values, or only the first two. Argument-blind policy cannot distinguish reading one record from exporting the table.
• Credential brokering. Whether it can hold per-upstream credentials, exchange a consumer identity for an upstream-scoped one, and avoid forwarding an inbound token to a system it was not issued for.
• Error fidelity. Whether upstream error structure survives the crossing intact, which determines whether a model can still recover from a correctable failure.
• Protocol adaptation. Whether backends that do not speak the protocol can be exposed through the same policy surface, since most organisations have more of those than of native servers.
• Audit completeness. Whether the record covers who called what, with which arguments, under which policy decision, and what came back — and where that record is retained.

Options