Code with Claude London: What Anthropic Will Show

Code with Claude London opens Monday May 19, 2026 at Park Plaza Westminster Bridge — Anthropic's first European developer summit, running the same two days as Google I/O 2026. Three product tracks are expected: self-hosted agent sandboxes entering public beta, MCP tunnels entering research preview, and an Anthropic Managed Agents API that would formalize the agent lifecycle into a single managed surface comparable to what Google is expected to announce at I/O.

The date overlap is not accidental. By counter-programming against Google I/O, Anthropic forces enterprise buyers on Claude to choose which keynote to track live. Teams that attend London are, by definition, self-selecting as Claude-primary organizations. That is the strategic bet, and it is worth understanding before Monday.

This preview covers what is confirmed, what is rumored for each track, how the event fits into the broader four-vendor managed-agent collision unfolding May 18–20, and which surface-area developments matter most for engineering teams already running Claude Agent SDK in production.

Code with Claude London is Anthropic's first European developer summit, confirmed on anthropic.com/events before May 17. The venue is Park Plaza Westminster Bridge on the South Bank — a deliberate choice that puts Anthropic next to central London's tech and financial district audience rather than a startup campus.

The format mirrors the two previous Code with Claude summits in San Francisco (May 2025) and New York: single-track keynotes in the morning, developer breakout sessions in the afternoon across three tracks — Claude Code, Agent SDK and MCP, and enterprise deployment. Dario Amodei is confirmed for the opening keynote on Monday morning. A closing fireside with CTO Tom Brown is reportedly planned for Tuesday, though not officially confirmed before May 17.

Attendance is invite-only with a reported cap under 500. The invite list is weighted toward Claude for Enterprise customers, AWS Bedrock and Vertex enterprise buyers, and Code with Claude alumni from the previous summits. Access through the Anthropic enterprise team — there is no public registration path.

The dates are not a scheduling coincidence. Google I/O 2026 runs May 19–20 at Shoreline Amphitheatre in Mountain View — the same two days. SAP Sapphire Madrid opens May 19 as well, completing a three-event collision that any enterprise technology buyer will need to navigate.

The strategic read — consistent with pre-event analysis from Stratechery and Interconnects — is that Anthropic benefits from the overlap rather than suffering it. By running London on exactly I/O week, Anthropic creates a natural fork in the enterprise buyer journey. Developer and enterprise teams that show up in London are expressing a preference. They are attending a Claude event when they could be watching Google's keynote. That self-selection is the point.

The secondary framing worth tracking: timing London before I/O means Anthropic can shape how its Managed Agents story lands in the press before Google's version takes the oxygen. The keynote on Monday May 19 runs in the London morning, hours before I/O's prime-time sessions on the US West Coast. If Anthropic ships first, the managed-agent narrative belongs to Anthropic for at least a news cycle.

Pre-event leaks at The Information and Semafor point to self-hosted agent sandboxes entering public beta at London. The expected architecture: enterprise customers run Claude agents inside their own VPC, with Anthropic supplying the sandboxing primitives rather than the execution environment. The distinction matters operationally — your infrastructure, your data perimeter, but Anthropic-managed isolation boundaries around what the agent can touch.

The benchmark every enterprise buyer will apply is Vercel Sandbox, which reached general availability in January 2026 using Firecracker microVMs. That primitive is public, well-documented, and already embedded in Vercel's AI agent workflows. Anthropic's version will need a clear differentiation story around enterprise compliance guarantees, VPC integration depth, and the interaction with Bedrock-EU data residency for the European audience in the room.

What Anthropic has not announced publicly before May 17 is any pricing model for the self-hosted tier. Expect the London keynote to frame the technical architecture first, with commercial details following in Bedrock and enterprise contract channels. Do not assume a public pricing page launches simultaneously.

Pre-event engineering discussion on Latent Space points to MCP tunnels entering research preview at London. The expected primitive: a mechanism for connecting on-premises MCP servers to remote Claude agents through an Anthropic-authenticated tunnel layer. The design pattern mirrors Cloudflare Tunnel and Tailscale Funnel — your MCP server stays behind your firewall, the tunnel authenticates the outbound connection, and remote Claude agents reach the server through the tunnel without exposing a public endpoint.

For enterprise buyers, the value proposition is direct: existing on-premises MCP servers built against the MCP spec 2025-11-25 (stdio + Streamable HTTP transport) can participate in remote agent workflows without a network architecture redesign. No public firewall rule, no reverse-proxy complexity for the security team to approve. The MCP ecosystem — which reportedly crossed 10,000 tracked servers by April 2026 — has significant on-premises deployment density that this primitive would unlock for cloud-hosted agents.

MCP tunnels are characterized here as a research preview because the pre-event signals consistently use that framing. Research preview in Anthropic's product cadence means the API is available and functional but may change before general availability. The MCP Inspector developer workflow remains the primary local-development tool while tunnel support matures.

The most anticipated announcement — surfaced via GitHub commit activity on the Anthropic engineering repositories before May 17 — is a formalized Managed Agents API in the Anthropic SDK. The expected model: a single API call that manages the full agent lifecycle, including tool calls, context management, retry logic, and termination, rather than requiring the caller to orchestrate each turn manually.

This would represent the productization of the patterns already documented in the Claude Agent SDK migration playbook. The @anthropic-ai/claude-agent-sdk npm package (renamed September 29, 2025) is the current entry point; a Managed Agents API would sit above it in the abstraction stack.

The competitive pressure here is clear. Google is expected to ship a Managed Agents API at I/O the same week. Cursor Composer 2.5 launches the day before on May 18 with xAI Colossus training. Vercel Sandbox is already GA. Anthropic's version needs to differentiate on the Claude reasoning quality per agent turn, the enterprise compliance guarantees, and the integration depth with the self-hosted sandbox track. That framing is likely to be the core of the Monday keynote.

The three product anchors on the confirmed London agenda correspond to Anthropic's three active developer surfaces. Understanding where each surface is before May 19 is the prerequisite for interpreting what the keynote actually moves.

Claude Code 1.3 is the current version, documented at code.claude.com/docs. The London breakout track on Claude Code is expected to cover the Skills Marketplace update and enterprise skill governance — how organizations manage which Claude Code skills are available to their developers, and how skill provenance is audited. No version number change has been confirmed for May 19.

Claude Agent SDK (@anthropic-ai/claude-agent-sdk, renamed September 29, 2025) is the current orchestration entry point. The London breakout is expected to preview the Managed Agents API as a higher-level abstraction above the SDK, not a replacement. The SDK migration patterns documented in our Claude Agent SDK production patterns guide remain current for teams on the SDK today.

MCP spec 2025-11-25(stdio + Streamable HTTP transport — SSE transport was deprecated March 26, 2025) is the spec version at London. The breakout session is expected to deliver an enterprise-focused OAuth 2.1 + PKCE implementation guide, bridging the gap between the spec's auth requirements and what enterprise security teams will sign off on. For teams building MCP servers today, the TypeScript MCP server tutorial covers the current transport and auth patterns.

The managed-agent SDK landscape will be materially different by end-of-day May 20 than it was at start-of-day May 18. Four vendors are landing — or have recently landed — managed-agent primitives within a 72-hour window. The table below maps the collision:

The Cursor dimension deserves context. In April 2026, SpaceX reportedly exercised a $60B acquisition option for Cursor, with a $10B breakup fee triggered approximately 30 days after the planned SpaceX IPO (June 12, 2026, at a reported $1.75T valuation). Cursor Composer 2.5, training on xAI Colossus with approximately one million H100-equivalent chips, is the product delivery that lands into this transaction. The day-before timing relative to the Anthropic keynote means Cursor is directly competing for developer attention in the same news cycle.

The actionable framing for enterprise teams: do not evaluate any of these four surfaces in isolation. The differentiation question after May 20 will be sandbox model (hosted vs. self-hosted vs. hybrid), auth surface depth, and integration with your existing cloud vendor — not which vendor announced first. Our AI transformation practice will publish a post-event four-way comparison once the London and I/O announcements are on the record.

Most coverage of Code with Claude London treats the venue as a logistics choice. The more useful read is as a platform-investment signal. The EU AI Act Chapter V obligations for General-Purpose AI models came into force August 2, 2025. Any model provider with significant European enterprise revenue — and Anthropic has both direct Claude Enterprise customers and indirect exposure through Bedrock-EU and Vertex-EU deployments — has compliance obligations around transparency, model documentation, and data handling that a London-based engineering office and European enterprise events directly support.

The three expected London tracks map cleanly to European enterprise requirements. Self-hosted sandboxes address data-residency concerns directly — agents running in the customer's VPC never send data to Anthropic-operated compute. MCP tunnels allow on-premises tool servers to participate in agent workflows without leaving the customer network perimeter. The enterprise OAuth 2.1 + PKCE implementation guide from the MCP session addresses what European security teams actually need to see before approving production deployments.

AWS Bedrock EU regions and Vertex AI EU regions are the deployment paths through which most European enterprises will operationalize Claude commercially. London as the venue for the Managed Agents announcement is Anthropic signaling that the European versions of those deployment paths are first-class, not afterthoughts. See the AWS Bedrock regional documentation for current EU availability status.

Whether you are watching London remotely or attending, the five things most worth tracking before Monday morning are below. Each maps to a specific technical decision your team will likely need to make within weeks of the event.

The broader context worth framing: Code with Claude London is the first event where Anthropic will position its agentic infrastructure story directly against a live Google I/O keynote. The two narratives will interact in real-time across enterprise buyer conversations. Teams currently evaluating Google I/O 2026 announcements alongside Anthropic's London tracks should bookmark both event channels and compare the managed-agent architecture announcements side-by-side before the week is out.

For organizations building MCP server infrastructure, the MCP Apps and interactive Claude UI guide covers the current surface area. The London tunnel and enterprise auth announcements will extend — not replace — those patterns. Watch the MCP spec changelog for any 2025-11-25 revision that ships alongside the tunnel announcement.