9 min read

MCP 2026-07-28 Spec: Stateless Servers and Your Data

The MCP 2026-07-28 spec drops sessions and the initialize handshake, adds cacheable lists and header routing. What changed and how to format tool results.

By json2toon.co

MCP 2026-07-28 makes the Model Context Protocol stateless. It removes the initialize handshake and the Mcp-Session-Id header, puts protocol metadata on every request, and adds routing headers, cacheable list results and multi round-trip requests. Roots, Sampling and Logging are deprecated. Tool result types stay the same, but each result now has to stand alone.

What changed in MCP 2026-07-28?

The release was announced on July 28, 2026 by lead maintainers David Soria Parra and Den Delimarsky on the official MCP blog. Soria Parra called it "MCP's most important since remote MCP first launched over a year ago." Here is what changed:

AreaBeforeIn 2026-07-28
Connection setupinitialize / initialized handshake, Mcp-Session-Id headerBoth retired; each request carries protocol version, client info and capabilities in _meta
RoutingGateways parse the JSON bodyMcp-Method and Mcp-Name headers; x-mcp-header mirrors chosen tool parameters into Mcp-Param-* headers
Server asks client for inputServer-initiated requests over an open streamMulti Round-Trip Requests: result type input_required, client retries with inputResponses
List resultsNo caching hintsttlMs and cacheScope on tools, prompts and resources lists
AuthorizationDynamic Client RegistrationDCR deprecated in favor of Client ID Metadata Documents; RFC 9207 issuer validation required
TasksExperimental core featureio.modelcontextprotocol/tasks extension, alongside MCP Apps and Enterprise Managed Authorization
Roots, Sampling, Logging, HTTP+SSESupportedDeprecated, with a minimum twelve-month window

Why did MCP go stateless?

Scale. With sessions, every request from a client had to reach the server instance that held its session state, which means sticky load balancing and fragile reconnects. Without sessions, any instance behind a load balancer can answer any request, and gateways can route and meter on headers instead of parsing JSON bodies.

Usage numbers explain the pressure. Monthly npm downloads of the TypeScript SDK went from about 50.2M in January 2026 to about 219.4M in September 2026. The MCP team reports "close to half-a-billion downloads a month" across its Tier 1 SDKs (TypeScript, Python, Go and C#).

What does a stateless MCP request look like?

Information that used to be exchanged once during the handshake now travels with every call. The spec's tools page says every request MUST include the _meta fields for protocol version, client info and client capabilities:

POST /mcp
Mcp-Method: tools/call
Mcp-Name: list_inventory

{
  "jsonrpc": "2.0",
  "id": 7,
  "method": "tools/call",
  "params": {
    "name": "list_inventory",
    "arguments": { "warehouse": "AMS" },
    "_meta": {
      "io.modelcontextprotocol/protocolVersion": "2026-07-28",
      "io.modelcontextprotocol/clientInfo": { "name": "my-agent", "version": "1.4.0" },
      "io.modelcontextprotocol/clientCapabilities": {}
    }
  }
}

List responses now say how long they can be cached. The spec's own tools/list example returns "ttlMs": 300000 and "cacheScope": "public", so a client can reuse the tool list for five minutes.

What does stateless mean for tool results and data?

The tool result format itself did not change. Results can contain text, images, audio, resource links, embedded resources and an optional structuredContent value. What changed is the context around them, in three ways.

State becomes explicit. The spec says "MCP has no protocol-level session", so a server that needs state, such as a shopping cart or a database transaction, should return a handle from one tool and accept it as an argument on later calls. That handle is now data the model has to carry forward in its context.

Results must be self-contained. A server cannot assume the model saw the previous page of results in a session. Agents therefore tend to request larger, complete payloads, and those payloads go straight into the model's context window. Their size and format affect both cost and accuracy.

Caching is a design goal. The spec says servers SHOULD return tools in deterministic order because it "improves LLM prompt cache hit rates when tools are included in model context." The same logic applies to tool results: a deterministic encoder produces the same bytes for the same data, so repeated context stays cacheable.

How should you format tool results under the new spec?

Keep machine-readable JSON in structuredContent, validated against your outputSchema. For the text block the model reads, the spec says a tool returning structured content SHOULD also return the serialized JSON there for backwards compatibility. For large uniform arrays, some servers deliberately put a compact encoding such as TOON in the text block instead. Here is a three-row inventory result as TOON, from @toon-format/toon 4.1.1:

items[3]{sku,name,stock,warehouse}:
  KB-201,Mechanical keyboard,42,AMS
  MS-110,Wireless mouse,0,AMS
  HD-330,"USB-C hub, 7-port",15,BER

On the current official benchmark (spec v4.1, 5,856 calls), TOON used 42.6% fewer tokens than JSON with 72.2% accuracy versus 71.4%. Agent research is more mixed. arXiv 2605.29676 tested formats on MCP benchmarks with open-weight models and found TOON was "not safe as a default in multi-turn agentic systems." Read our summary in TOON in multi-turn agents before you change every tool. For a working server, see building an MCP server with TOON.

What should MCP server authors do now?

  • Upgrade to a Tier 1 SDK release that supports 2026-07-28 and stop depending on initialize or session IDs.
  • Move per-user state into explicit handles that you validate on every call.
  • Return tools in a stable order and set sensible ttlMs values on list results.
  • Replace server-initiated Sampling and Roots calls with Multi Round-Trip Requests before the deprecation window ends.
  • Add outputSchema to data tools, and decide on purpose what goes into the text block. Our TOON for function calling guide covers the trade-offs.
  • Measure token usage on real tool outputs. Paste a response into the TOON token calculator to compare sizes, or call json_to_toon on our stateless TOON MCP server.

Frequently Asked Questions

What is the biggest change in MCP 2026-07-28?

The protocol core is now stateless. The initialize and initialized handshake and the Mcp-Session-Id header are gone. Every request carries its protocol version, client identity and client capabilities in _meta, so any server instance can answer any request.

Do my existing MCP servers stop working?

Not immediately. Roots, Sampling, Logging and the legacy HTTP+SSE transport are deprecated but keep working for at least twelve months under the new deprecation policy. Dynamic Client Registration still works for backward compatibility but will be removed in a future version.

What are ttlMs and cacheScope?

They are fields on list results such as tools/list. ttlMs tells the client how long it may reuse the result, and cacheScope says who may share the cached copy. Together they cut repeated list calls, and stable tool lists also help LLM prompt cache hit rates.

Does the new MCP spec change tool result formats?

The content types are the same: text, image, audio, resource links, embedded resources and structuredContent. Because there is no session, results must be self-contained, so compact formatting of large data in the text block matters more than before.

Recommended Reading

MCPModel Context ProtocolAI AgentsSpecificationTool CallingTOON