Tratopedia
繁中
Settings

Text size

Language

Theme

High contrast

Version

v1.230.2

The release this page was built from. It is what the service worker caches under.

APIs · Model Context Protocol · AI Agents · Software Engineering · Web Infrastructure · Comparison BriefThe five as they stand on 27 October 2026

Five ways software talks to software

REST, SOAP, gRPC, GraphQL and MCP: pick by who calls

REST, SOAP, gRPC and GraphQL answer one question: how does one program call another? MCP, the Model Context Protocol, answers a different one: how does an AI model find a capability and use it safely? The first four are rivals only in part, since each suits a different caller. MCP usually sits on top of them rather than replacing them. Its newest specification, dated 28 July 2026, made the protocol stateless.

Where each of the five stands Latest specification and steward, October 2026

  • 2026-07-28MCP’s current revision: stateless, governed by the Agentic AI Foundation under the Linux Foundation
  • 3.2.0OpenAPI, the usual contract for REST, dated 19 September 2025
  • 2025GraphQL’s September edition, the first since October 2021
  • 2017gRPC joined CNCF as Incubating, and it is still listed there
  • 2007SOAP 1.2 Second Edition, a W3C Recommendation and still the current version

Confirmed

The MCP 2026-07-28 changes come from its maintainers’ post and changelog. The handshake and session header are gone, each request describes itself, method and tool names travel in headers, list results carry cache hints, and Roots, Sampling and Logging are deprecated with at least twelve months’ notice. Each version and date above comes from its own steward.

Confirmed, but the sources disagree

How widely MCP is used. On the same day, the MCP blog said the Tier 1 SDKs see “close to half-a-billion downloads a month”, while Anthropic said MCP had “surpassed 400M monthly SDK downloads”. Both are promotional figures and neither was checked independently.

Proposed, not standard

MCP over gRPC. In January 2026 Google Cloud proposed gRPC as a pluggable MCP transport and said the core maintainers had agreed to support pluggable transports in the SDKs. Spotify says it runs it experimentally in-house. The official transports are still stdio and Streamable HTTP.

From XML envelopes to a plug for AI agents Dates as each steward or source gives them

  1. Jun 1998XML-RPC appears; SOAP grows out of it (Wikipedia).
  2. 2000Roy Fielding names and defines REST in his doctoral dissertation (Wikipedia).
  3. 27 Apr 2007SOAP 1.2 Second Edition becomes a W3C Recommendation (W3C).
  4. Mar 2015Google decides to build the successor to Stubby, its internal RPC system since about 2001, as open source; the result is gRPC (Wikipedia).
  5. 14 Sep 2015Facebook releases GraphQL publicly, after developing it since 2012 (Wikipedia).
  6. 16 Feb 2017CNCF accepts gRPC at the Incubating level (CNCF).
  7. Jun 2022RFC 9110 restates HTTP’s semantics, the ground REST stands on (RFC Editor).
  8. Nov 2024Anthropic introduces MCP (Wikipedia).
  9. 15 Aug 2025AWS introduces AgentCore Gateway, which turns OpenAPI and Smithy APIs and Lambda functions into MCP tools (AWS).
  10. 8 Sep 2025GraphQL’s September 2025 edition is announced, the first since October 2021 (GraphQL).
  11. 19 Sep 2025OpenAPI 3.2.0 is published (OpenAPI Initiative).
  12. 4 Nov 2025Anthropic shows one MCP workflow falling from 150,000 to 2,000 tokens when the agent writes code instead of loading every tool definition (Anthropic).
  13. 9 Dec 2025Anthropic donates MCP to the Agentic AI Foundation, co-founded with Block and OpenAI under the Linux Foundation; the maintainers stay the same (Anthropic; MCP blog).
  14. 14 Jan 2026Google Cloud proposes gRPC as a pluggable transport for MCP (Google Cloud).
  15. 29 Apr 2026A Tech Insider benchmark puts gRPC at 2.3 ms against REST’s 10.1 ms on small payloads, with a much smaller gap on large ones (Tech Insider).
  16. 28 Jul 2026MCP 2026-07-28 ships, with a stateless core, header routing, cacheable lists and a deprecation policy; Anthropic calls it the fifth spec release (MCP blog; Anthropic).

Choose by caller, then put MCP on top What the specifications and their stewards support, and what complicates it

Each style suits one caller. REST rides on HTTP’s own semantics, so a GET can be cached by browsers, proxies and CDNs, and RFC 9110 calls the conditional GET “the most efficient mechanism for HTTP cache updates”. That suits outside developers. gRPC needs HTTP/2 and binary Protocol Buffers, and Microsoft’s architecture guide says it “should be the primary choice for direct synchronous communication between microservices”. A browser, though, needs gRPC-Web or a JSON gateway, and gRPC-Web cannot do client or bidirectional streaming. GraphQL lets the client name the fields it wants from one endpoint, which suits front ends with many screens. SOAP’s WSDL contracts and message-level signing keep it alive in older enterprise links. MCP is a different kind of layer: a model, not a compiler, reads its contract, and the specification warns that “Tools represent arbitrary code execution”. Gateways from Microsoft, AWS and Apollo already turn REST, OpenAPI and GraphQL into MCP tools without rebuilding the API underneath.

  • 0Sessions an MCP 2026-07-28 server must keep: any request can land on any instance
  • 98.7%Token saving in one Anthropic MCP example, from 150,000 to 2,000
  • 77%gRPC’s lower latency than REST on small payloads, in one benchmark

What complicates it

Speed figures do not travel. The same Tech Insider benchmark that gives gRPC 77% lower latency on small payloads gives it only 15% on large ones. The 98.7% token saving is Anthropic’s own best case, and running code that way needs a secure sandbox. MCP also deprecates things quickly: Roots, Sampling and Logging were deprecated less than two years after the first release, though they keep working for at least twelve months.

Securing each one, and wrapping them for agents Each item as the named source gives it

StyleWhat to watchSource
GraphQLQuery depth and amount “can both be unlimited which may lead to a DoS”. Set limits, cost analysis and timeouts, and watch for batching, where one request carries many operations.OWASP
MCPToken passthrough, where a server forwards a client’s token downstream without checking that it was issued to that server, is named an anti-pattern. Tool descriptions count as untrusted unless the server is trusted, and hosts need explicit user consent before calling a tool.MCP specification
gRPCSet a deadline on every call. In a browser, gRPC-Web cannot call client-streaming or bidirectional methods.gRPC; Microsoft
ProtobufMoving from proto2 or proto3 to editions such as 2024 changes the generated code, not the wire format.Protocol Buffers
  • Gateways

    REST and OpenAPI into MCP

    Microsoft; AWS

    • Azure API Management can expose a managed REST API as a remote MCP server. It offers tools only, not resources or prompts.
    • AWS’s AgentCore Gateway converts OpenAPI and Smithy APIs and Lambda functions into MCP tools.
  • Graph and RPC

    GraphQL and gRPC into MCP

    Apollo; Google Cloud

    • Apollo MCP Server exposes GraphQL operations as tools, and its mutation_mode defaults to none, so nothing can write until someone allows it.
    • For gRPC, Google Cloud’s transport is a proposal, so a translating gateway is still the usual route.

Which one suits which caller Each card: who it is for, its strongest point, and the trade-off

  • REST

    Outside developers and the web

    With OpenAPI 3.2

    • Best at: the lowest barrier, the most tools, and caching by browsers and CDNs.
    • Trade-off: over-fetching and extra round trips.
  • gRPC

    Your own services, talking to each other

    HTTP/2 and Protocol Buffers

    • Best at: typed contracts, small binary messages, deadlines and four streaming modes.
    • Trade-off: no HTTP caching, and browsers need gRPC-Web or a JSON gateway.
  • GraphQL

    Front ends with many screens

    One endpoint, one typed schema

    • Best at: letting the client ask for exactly the fields it needs.
    • Trade-off: caching needs extra work, and servers need depth and cost limits.
  • SOAP

    Existing bank, insurance and government links

    WSDL and XML

    • Best at: security on the message itself, which survives every hop between systems.
    • Trade-off: verbose and heavy; wrap it behind REST rather than spread it.
  • MCP

    AI agents using your tools and data

    On top of the other four

    • Best at: one integration that any compatible AI app can find and call.
    • Trade-off: the risks lie in meaning, such as prompt injection and token passthrough, and the spec still moves fast.
  • Ask who calls before asking what is fastest. A single benchmark is not your workload.
  • Build new MCP servers for 2026-07-28. Keep no session, let gateways route on the headers, and leave Roots, Sampling, Logging and HTTP+SSE alone.
  • Treat every MCP tool as code that runs. Start writes switched off, ask a person before sensitive actions, and never pass a client’s token through.

Quick notes

  1. API guide · 27 October 2026

    Five ways software talks to software

    REST, SOAP, gRPC, GraphQL and MCP, explained for beginners and professionals.

    REST · SOAP · gRPC · GraphQL · MCP1 / 15

  2. The basics

    What is an API?

    An API is an agreed way for one program to ask another for something and get a predictable answer. Think of ordering in a restaurant.

    You

    The client: an app, a website or an AI agent that wants something.

    The waiter

    The API: the menu and the rules for ordering.

    The kitchen

    The server: the system that holds the data and does the work.

    REST · SOAP · gRPC · GraphQL · MCP2 / 15

  3. The basics

    The five at a glance

    NameWhat it isThink of it as
    RESTWeb addresses plus standard actionsA library catalogue
    SOAPFormal XML messages with strict contractsA registered letter
    gRPCFast binary calls between servicesA direct line between colleagues
    GraphQLOne endpoint; you ask for exactly the fields you needA custom order form
    MCPA standard plug for AI agents to use tools and dataA universal adapter

    REST · SOAP · gRPC · GraphQL · MCP3 / 15

  4. The five · 1 of 5

    REST: resources at web addresses

    Best for: public web APIs, websites and mobile apps.

    StepExample
    RequestGET /orders/42
    Response200 OK, with the order as JSON
    • Everything is a resource with its own address, such as an order, a user or a photo.
    • Standard verbs: GET reads, POST creates, PUT or PATCH updates, DELETE removes.
    • Cache-friendly: browsers and CDNs can store GET results.

    REST · SOAP · gRPC · GraphQL · MCP4 / 15

  5. The five · 2 of 5

    SOAP: the formal envelope

    Best for: connecting existing bank, insurance and government systems. Avoid it for new APIs.

    PartHolds
    Envelope (XML)The whole message
    HeaderSecurity, routing and extras
    BodyThe actual request or reply
    • Strict contract: a WSDL file and XML schemas describe every message.
    • Message-level security: WS-Security can sign and encrypt the message itself.
    • Heavy: XML is verbose, so it is slow and awkward on mobile.

    REST · SOAP · gRPC · GraphQL · MCP5 / 15

  6. The five · 3 of 5

    gRPC: fast calls between services

    Best for: internal microservices, low latency and live streams.

    In a .proto contract
    rpc GetOrder(OrderId) returns (Order);
    • Contract first: one .proto file generates client and server code in many languages.
    • Compact and quick: binary messages over HTTP/2, with four streaming modes.
    • Browser limits: web pages need gRPC-Web or a JSON gateway.

    REST · SOAP · gRPC · GraphQL · MCP6 / 15

  7. The five · 4 of 5

    GraphQL: ask for exactly what you need

    Best for: apps with many screens or clients sharing one data graph.

    One query, one endpoint
    Ask for order(id: 42), and from it only status and each item’s name
    • The client picks the fields, so there is no over-fetching and fewer round trips.
    • Typed schema that tools can explore and check automatically.
    • Extra care: caching needs extra work, and servers need depth and cost limits.

    REST · SOAP · gRPC · GraphQL · MCP7 / 15

  8. The five · 5 of 5

    MCP: a standard plug for AI agents

    Build one MCP server and any compatible AI app can use it, instead of one custom connector per product.

    Tools

    Actions the AI can run, such as creating a ticket.

    Resources

    Data the app can read, such as a file or a record.

    Prompts

    Ready-made templates a person can choose.

    • Runs over stdio (local) or Streamable HTTP (remote), using JSON-RPC 2.0.

    REST · SOAP · gRPC · GraphQL · MCP8 / 15

  9. How they fit together

    MCP sits on top of the others

    MCP does not replace them. It wraps existing APIs as tools, so an AI agent can discover and call them in a standard way. Cloud gateways already turn REST, OpenAPI and GraphQL services into MCP tools.

    LayerWhat it is
    TopAn AI agent or chat app
    MiddleAn MCP server: tools · resources · prompts
    BottomREST · GraphQL · gRPC · SOAP

    REST · SOAP · gRPC · GraphQL · MCP9 / 15

  10. Comparing

    Side by side

    RESTSOAPgRPCGraphQLMCP
    Message formatJSON, mostlyXMLProtobuf (binary)JSONJSON-RPC 2.0
    TransportHTTPMostly HTTPHTTP/2HTTP, usually POSTstdio or Streamable HTTP
    ContractOpenAPIWSDL and XSD.proto fileSchema (SDL)JSON Schema per tool
    StreamingLimitedNoneFour modesSubscriptionsServer events (SSE)
    HTTP cachingBuilt inAlmost noneNoneNeeds extra workCacheable lists since 2026-07-28
    Main callerAny developerEnterprise systemsOther servicesFront-end appsAI agents

    REST · SOAP · gRPC · GraphQL · MCP10 / 15

  11. Comparing

    Speed depends on the workload

    • 77%Lower latency for gRPC than REST on small payloads, 2.3 ms against 10.1 ms
    • 15%Lower latency on large payloads
    • The gap shrinks as messages grow. This is one 2026 benchmark, so test your own workload before choosing a protocol for speed.
    • For MCP, the usual bottleneck is tokens in the AI’s context window, not the wire format.

    REST · SOAP · gRPC · GraphQL · MCP11 / 15

  12. Comparing

    Security: what protects you, what to watch

    MCP’s biggest risks sit in meaning, not transport: treat tool output as untrusted.

    OptionMain protectionCheck for
    RESTTLS, OAuth 2.x, API keysAccess checks on every object, rate limits
    SOAPWS-Security signing and encryptionXML parser hardening
    gRPCTLS or mutual TLS, per-call tokensA deadline on every call
    GraphQLDepth and cost limits, timeoutsCostly queries and batching abuse
    MCPOAuth 2.1, least privilege, approvalsPrompt injection, token passthrough

    REST · SOAP · gRPC · GraphQL · MCP12 / 15

  13. Deciding

    Which one should I use?

    SituationFirst choiceWhy
    Public API for outside developersREST + OpenAPIEasiest to adopt and cache
    Internal service-to-service callsgRPCFast and strongly typed
    An app with many different screensGraphQLFetch only needed fields
    Legacy bank or government linkSOAPWrap it behind REST
    Let an AI agent use your toolsMCPOne plug for many AI apps

    REST · SOAP · gRPC · GraphQL · MCP13 / 15

  14. Deciding

    One company, several styles

    StyleServes
    RESTPartners and third-party developers
    GraphQLYour own web and mobile front ends
    gRPCMicroservices talking to each other
    SOAPLegacy systems, isolated behind an integration layer
    MCPAI agents, through a gateway that handles identity, policy and audit

    REST · SOAP · gRPC · GraphQL · MCP14 / 15

  15. Takeaways

    Three things to remember

    • Pick by who calls you: outside developers, your own services, your own front ends or an AI agent.
    • MCP sits on top of the other four. It does not replace them.
    • Test with your own workload, and secure every layer.

    REST · SOAP · gRPC · GraphQL · MCP15 / 15

  • David Soria Parra and Den Delimarsky, “The 2026-07-28 Specification”, MCP blog, 28 Jul 2026
  • Model Context Protocol, 2026-07-28 changelog, specification, versioning, security best practices and Streamable HTTP pages
  • David Soria Parra, “MCP joins the Agentic AI Foundation”, 9 Dec 2025
  • Anthropic, “Donating MCP to the Agentic AI Foundation”, 9 Dec 2025
  • Adam Jones and Conor Kelly, “Code execution with MCP for efficient agents”, Anthropic, 4 Nov 2025
  • Anthropic, “Bringing MCP 2026-07-28 to Claude”, 28 Jul 2026
  • Chris Wood, “Announcing OpenAPI v3.2”, 23 Sep 2025
  • OpenAPI Specification v3.2.0, 19 Sep 2025
  • CNCF, gRPC project page
  • Lee Byron, “Announcing the September 2025 Edition of the GraphQL Specification”, 8 Sep 2025
  • W3C, SOAP Version 1.2 Part 1 (Second Edition), 27 Apr 2007
  • Fielding, Nottingham and Reschke, RFC 9110, Jun 2022
  • Wikipedia, “REST”, “GraphQL”, “SOAP”, “gRPC” and “Model Context Protocol”
  • Nadia Dubois, “gRPC vs REST 2026”, Tech Insider, 29 Apr 2026
  • Microsoft Learn, gRPC in the .NET cloud-native guide, gRPC-Web in ASP.NET Core, Compare gRPC services with HTTP APIs, and Expose REST API as MCP server
  • Protocol Buffers, Editions Overview
  • gRPC, Core concepts
  • OWASP, GraphQL Cheat Sheet
  • Patel, Rustagi and Liu, “Introducing Amazon Bedrock AgentCore Gateway”, AWS, 15 Aug 2025
  • Apollo GraphQL, Apollo MCP Server config file reference
  • Victor Moreno and Mark D. Roth, “A gRPC transport for the Model Context Protocol”, Google Cloud, 14 Jan 2026. All read 27 Oct 2026.

Versions

This document is rewritten when what it says has to change. Every version stays published at its own address.

  1. v0001 current

The current version is also at latest/.