Skip to content

MCP SDK v2 adoption survey

A survey of which MCP SDK v2 and MCP 2026-07-28 features AutoMobile used, which it did not, and the order in which the unused ones were worth adopting. The usage map and baseline describe the code as of the survey; later work (live observe shell, elicitation, opt-in modern-era serving, resource_link artifacts, Skills) changed several rows.

Baseline at survey time

  • Installed: @modelcontextprotocol/server and client 2.3.1, node 2.1.1. SDK 2.4.0 is published and adds the Skills extension, ServerEvent params, and an outputSchema $ref fix. Its upgrade notes do not touch us: no tool uses x-mcp-header or a z.object().meta({ id }) output schema.
  • The SDK negotiates the 2025-11-25 wire on initialize. The 2026-07-28 (“modern”) era is reached only through createMcpHandler / PerRequestHTTPServerTransport, which the daemon does not use. The daemon serves the 2025 era statelessly.
  • No confirmed host client sends a 2026-07-28 request yet. The community extension support matrix lists no client for Tasks at all, and lists Skills as partial in ChatGPT, fast-agent and MCP Inspector only.

What 2026-07-28 changed

  • Stateless protocol (no initialize, no Mcp-Session-Id; version and capabilities in every request’s _meta)
  • server/discover
  • subscriptions/listen replacing resources/subscribe and the GET stream
  • multi round-trip requests (input_required results instead of server-initiated elicitation, sampling and roots requests)
  • ttlMs / cacheScope on list and read results
  • Tasks moved to the io.modelcontextprotocol/tasks extension with polling and tasks/update
  • SSE resumability removed
  • ping, logging/setLevel and roots/list_changed removed
  • Roots, Sampling and Logging deprecated
  • trace-context _meta keys
  • routing headers

Usage map

Area SDK / spec surface State in AutoMobile
Tools: inputSchema, outputSchema, structuredContent core Used by every tool. structuredContent mirrors the compact text; extras are opt-in.
Tool annotations and title core Used (src/server/toolAnnotations.ts).
Progress and cancellation notifications/progress, notifications/cancelled Used. Progress rides the request’s own stream; socket disconnect cancels via the execution tracker.
Resources and templates, listChanged core Used broadly. tools/list_changed is an internal push with a hash-polling fallback in the proxy.
Resource subscriptions resources/subscribe, resources/updated Used (src/daemon/proxy/ResourceSubscriptions.ts). Legacy API; subscriptions/listen is unused.
Tasks extension extension, SDK ships types only Used, own SQLite-backed implementation (src/server/tasks/, src/daemon/tasks/). Eligible per call: startDevice, executePlan, installApp, videoRecording stop. Advertised on a 2025-era initialize to a declaring client.
MCP Apps ui:// resources Used for observe (static snapshot); live shell is #4682.
resource_link content blocks core Unused.
Multi round-trip requests, elicitation inputRequired, acceptedContent, legacy shim Unused. Tasks have requestInput plumbing no tool calls.
Server instructions core Unused.
Completions completable Unused.
server/discover, cache hints (ttlMs, cacheScope) modern era Unused.
subscriptions/listen modern era Unused.
Skills over MCP SDK 2.4.0 ext/skills Unused.
Trace context _meta traceparent, tracestate, baggage Unused.
Auth (bearer, protected-resource metadata, OAuth client) SDK server and client helpers Unused. The HTTP ingress is loopback with a per-daemon random token and a host allowlist.
Sampling, Roots deprecated Unused; do not adopt.
Logging deprecated sendLoggingMessage only inside the stdio proxy; do not extend.
SSE resumability / event store removed from the spec Unused; do not adopt.
Tunnels Anthropic product, not SDK or spec Not used; see Tunnels.

Ranked proposals

Rank is value to the agent (the primary consumer) first, then to a human supervisor, divided by effort and risk, and discounted by real client support.

# Proposal Benefit Effort Risk Depends on Issue
1 Return large non-image artifacts (video exports, session logs, exported plans) as resource_link, inline payload opt-in Fewer model-context bytes and client spills on big results; works on every client, since automobile:// resources already exist S-M Low-medium: a client that never fetches links loses the payload, so images and small results stay inline Per-tool byte measurement #13942
2 Byte-identical, deterministic tools/list Stable client list caches and LLM prompt-cache hits; the spec now says SHOULD S Low none #13943
3 Choose further Tasks-eligible tools from measured durations (snapshot create/restore, large putAppFile, cold-boot kill, uninstall with data wipe) Long device operations stop holding a request open; survives client reconnects S to measure, M per tool Low (opt-in per request) Device time for measurement; a Tasks-declaring client #13944
4 Publish a usage skill through the Skills extension Workflow guidance (session binding, observe before act, acquisition) pulled on demand instead of sitting in every context; the alternative to a permanent instructions block S-M Low-medium: few hosts consume it yet; content must stay current with the tools SDK 2.4.0 #15430
5 Elicitation through multi round-trip requests for an ambiguous deviceId and destructive confirmations A supervised host can ask instead of failing with a candidate list; the SDK’s legacy shim serves pre-2026 clients with the same handler M Medium: a result that asks a question stalls unattended agents, the majority of traffic Owner decision; must never block an unanswered call #13945
6 Measured instructions block (at most about 300 tokens) Unknown until benchmarked; adds tokens to every client S Low-medium Benchmark; overlaps proposal 4 #13946
7 Serve the modern era beside the 2025 era (server/discover, cache hints, subscriptions/listen, MRTR, spec-correct Tasks negotiation) High once a host speaks 2026-07-28; none today L Medium-high: our own tools/list and tools/call handlers and the session-key model need an adapter around the SDK handler A host client sending 2026-07-28 #13947
8 Finish the live observe App shell, then an element picker Human oversight; the picker also helps agents M Medium, host-gated A host that renders Apps with session data #4682
9 Remote device access through MCP Tunnels Cloud-hosted agents reaching an owner-run device bench M-L High: new non-loopback auth surface Product decision #4674 (park)

Opportunistic, not tracked: completions on resource-template URIs (few clients call completion), trace-context propagation (needs an exporter first).

Notes on the order

  • Proposals 1 to 3 need no new client capability, so their benefit is immediate and their downside is bounded to our own output shaping. Measure before changing any default: clients differ in whether they follow resource_link.
  • Proposals 4 and 6 are alternatives for the same job. Prefer the skill: it costs context only when a host loads it, and the same text works on both protocol eras.
  • Proposal 7 gates several others (cache hints, subscriptions/listen, spec-correct Tasks negotiation). In the modern era the Mcp-Session-Id header is gone and the spec’s answer to cross-call state is an explicit server-minted handle passed as an ordinary tool argument. AutoMobile’s header key (x-auto-mobile-mcp-session-id) is an extension the host must be able to set; the in-band __mcpSessionId argument is the spec-shaped path and must keep working.
  • Tasks remain opt-in per request. Nothing that returns in under about two seconds should become a task.
  • SSE resumption or an event store (removed from the spec; a broken stream means re-issuing the request, and tasks already give a durable handle)
  • sampling and roots (deprecated)
  • per-request log level (logging is deprecated)
  • server-side request batching
  • taskifying short interaction tools (tapOn, inputText, pressButton)

Tunnels

MCP Tunnels is an Anthropic research-preview product, not part of the SDK or the specification. An outbound-only cloudflared connector plus an Anthropic proxy run in the customer’s network; the proxy terminates inner TLS and routes each request to an upstream MCP server by hostname. The tunnel carries traffic but does not authenticate to the upstream server: each MCP server brings its own OAuth or bearer token. The consumers are Claude Managed Agents and the Messages API MCP connector; a tunnel made in the Console is not available as a claude.ai connector.

Fit against the daemon:

  • What it would add: a cloud-hosted agent reaching devices on a private bench or emulator farm without inbound ports. Local clients (Claude Code, Cursor, IDE plugins) already reach the daemon through the stdio proxy and need nothing.
  • What the daemon lacks for it: a non-loopback Streamable HTTP endpoint with real authorization. The ingress is loopback-only with one random token and a host allowlist. A tunnel needs OAuth (the SDK has bearer-verification and protected-resource-metadata helpers) and the tunnel hostname in the allowlist.
  • Semantics that do not carry over: a hosted MCP connector sends static headers, so the per-conversation session key header cannot be set; binding would rely on the in-band __mcpSessionId handle. The proxy-side liveness that keeps a device lease alive has no hosted equivalent; the lease keepalive Tasks already use is the closest mechanism.
  • Recommendation: park. Revisit only if supporting cloud-hosted agents against owner-run device benches becomes a product goal; the prerequisite work is the auth surface, not the tunnel.

Sources

MCP 2026-07-28 changelog, Tasks extension, Skills extension, extension support matrix, MCP tunnels, and the TypeScript SDK v2 release notes.