Key Signals

• SpaceX and Cursor announce a partnership worth $10 billion to $60 billion, reshaping the AI coding landscape. SpaceX will pay at least $10 billion for a collaboration to "create the world's best coding and knowledge work AI," with an option to acquire Cursor outright for $60 billion later this year. Cursor has scaled to over $2 billion in annualized revenue and may be worth $50 billion privately, placing it alongside OpenAI and Anthropic as one of the three dominant forces in AI-assisted development. The deal pairs Cursor's frontier coding models and massive developer usage data with xAI's compute infrastructure, potentially breaking the compute bottleneck that Cursor says has constrained its model training. [1]

• Roo Code shuts down its VS Code extension and pivots entirely to a cloud-based autonomous agent, signaling a growing industry shift away from IDE-centric AI coding. The open-source tool, which reached approximately 3 million installs as a fork of Cline, will sunset its extension, Cloud, and Router services on May 15. CEO Matt Rubens stated that "IDEs aren't the future of coding" and introduced Roomote, a prompt-driven cloud agent that completes tasks end-to-end across Slack, GitHub, and Linear. Kilo Code, another Roo Code fork, is positioning itself as an alternative for users who want to stay in the IDE, arguing that "the IDE is not over — far from it." [2]

• Anthropic tested removing Claude Code from its $20/month Pro plan, revealing the strain that agentic coding workloads place on subscription economics. The change, applied to approximately 2% of new prosumer signups, triggered immediate community backlash before the pricing page was reverted. Head of Growth Amol Avasare explained that when Max launched a year ago, Claude Code and long-running async agents did not exist — subscription plans "weren't built for this" level of engagement. OpenAI seized the opening, publicly pledging that Codex will remain available on both free and Plus ($20) plans, claiming sufficient compute and efficient models to sustain it. [3][4]

• GitHub Copilot CLI ships two pre-releases and launches C++ code intelligence in public preview, expanding language-server-grade tooling to the terminal. The Microsoft C++ Language Server, powered by the same IntelliSense engine used in Visual Studio and VS Code, gives Copilot CLI precise semantic data — symbol definitions, references, call hierarchies, and type information — for a language that is notoriously difficult for text-based search alone. The v1.0.35-4 pre-release adds named sessions with --name/--resume and moves MCP OAuth into a shared runtime flow, while v1.0.35-3 improves rendering performance for large outputs, self-correcting tool calls in the agentic loop, and instant plugin activation without restart. [5][6][7]

• Claude Code v2.1.117 introduces forked subagents, replaces Glob/Grep with native search tools, and fixes a critical Opus 4.7 context window miscalculation. The release enables forked subagents via CLAUDE_CODE_FORK_SUBAGENT=1, embeds bfs and ugrep directly into native macOS/Linux builds for faster searches without a separate tool round-trip, and corrects a bug where Opus 4.7 sessions were computed against a 200K context window instead of the native 1M — causing inflated /context percentages and premature autocompaction. Default effort for Pro/Max subscribers on Opus 4.6 and Sonnet 4.6 is now "high," and concurrent MCP server connection is the new default for faster startup. [8]

• Cloudflare reaches general availability on Sandboxes and publishes an enterprise MCP reference architecture, professionalizing the infrastructure layer for AI coding agents. Sandboxes now offer secure credential injection via zero-trust egress proxies, PTY terminal support, persistent code interpreters, filesystem watching, and snapshot-based session recovery at $0.00002 per vCPU-second — with Figma already running production agent workloads. Separately, Cloudflare's MCP architecture introduces "Code Mode," which collapses tool interfaces into dynamic entry points and reduces token usage by up to 99.9%, alongside an AI Gateway for cross-provider model routing and per-user cost controls. [9][10]

AI Coding News

• SpaceX and Cursor are partnering on a deal worth between $10 billion and $60 billion to combine Cursor's AI coding models with xAI's compute infrastructure. SpaceX will pay at least $10 billion for the collaboration, with an option to buy Cursor entirely for $60 billion. Cursor, which has reached more than $2 billion in annualized revenue, acknowledged that compute has been its bottleneck, while xAI's grok-code-fast-1 model has not gained meaningful traction in the coding market. The deal gives Cursor access to xAI's massive data center footprint and gives xAI access to Cursor's developer usage data across both its own models and competitors' products deployed in production. [1]

• Roo Code will shut down its VS Code extension, Cloud, and Router services on May 15, 2026, and is launching Roomote, a cloud-based autonomous coding agent. The open-source project, which emerged as a fork of Cline in late 2024 and reached approximately 3 million installs, is abandoning the IDE model entirely. Roomote takes a prompt and carries out tasks end-to-end, integrating with Slack, GitHub, and Linear to generate pull requests, run tests, and verify results before handing work off for review. Cline founder Saoud Rizwan acknowledged that Roo Code "contributed to Cline more than anyone else" among forks, while Kilo Code is positioning itself as an IDE-based alternative for departing Roo Code users. [2]

• Anthropic tested removing Claude Code access from its $20/month Pro plan for approximately 2% of new signups, exposing the pressure that agentic workloads place on AI subscription pricing. The change was spotted on the pricing page and prompted rapid community pushback on Reddit and X, after which Anthropic reverted the page. Head of Growth Amol Avasare explained that Max was originally designed for heavy chat usage and did not include Claude Code or Cowork — long-running async agents are now everyday workflows, and "engagement per subscriber is way up." OpenAI responded publicly, with a team member stating that "Codex will continue to be available both in the FREE and PLUS ($20) plans." [3][4]

• At the MCP Summit in New York City, AWS revealed how Amazon Bedrock is shaping the Model Context Protocol through contributions driven by real customer use cases. Luca Chang, an MCP Specification Maintainer on the Bedrock team, explained that Amazon's major contributions — the Tasks and Elicitations features — arose from mapping AWS cloud products to the protocol and finding gaps. The MCP maintainer cohort is now broad enough that a recent pre-conference meeting could not cover all proposed topics, reflecting the protocol's rapid evolution as an industry standard for connecting AI agents to tools and data. [17]

• Cloudflare announced the general availability of Sandboxes, providing persistent isolated Linux environments purpose-built for AI agent workloads. Launched in beta last June, the GA release adds secure credential injection through zero-trust egress proxies where the agent never sees the token, PTY terminal sessions proxied over WebSocket, persistent code interpreters that maintain state across calls, and snapshot-based session recovery that reduces startup from 30 seconds to 2 seconds. Active CPU pricing charges only for cycles actually used at $0.00002 per vCPU-second, and Figma is already running production agent workloads on the platform. [9]

• Cloudflare published a reference architecture for scaling MCP deployments in the enterprise, with centralized governance and a "Code Mode" that reduces token usage by up to 99.9%. The architecture positions remote MCP server deployment, SSO/MFA authentication through Cloudflare Access, and an AI Gateway for cross-provider routing and per-user cost monitoring as essential for production agent systems. Code Mode collapses verbose tool interfaces into a small set of dynamic entry points that models discover on demand, addressing context window limitations. Forrester notes that MCP is fundamentally a transport mechanism, not a governance layer — governance, observability, and policy enforcement are emerging as a separate "control plane" concern. [10]

Feature Update

• GitHub Copilot CLI v1.0.35-4 adds named sessions and moves MCP OAuth into a shared runtime flow. Users can now create sessions with --name and resume them by name with --resume=, making it easier to maintain context across terminal sessions. LSP server entries in lsp.json now support configurable spawn, initialization, and warmup timeouts. MCP OAuth state is automatically cleaned up when removing an MCP server, and the Copilot agent now has shell access on Windows. [6]

• GitHub Copilot CLI v1.0.35-3 introduces a usage contribution graph, improved rendering, and agentic loop fixes. The /usage command now displays a GitHub-style contribution graph that adapts to terminal color mode. Sync task calls under MULTI_TURN_AGENTS now block until completion instead of auto-promoting to background after 60 seconds. Plugins take effect immediately after install without a restart. Fixes address self-correcting custom tool calls in the agentic loop, emoji and multi-codepoint character handling, session token auto-refresh on expiry mid-turn, and prevention of loading ~/.claude/ configs as Copilot project config. [7]

• The Microsoft C++ Language Server is now in public preview for GitHub Copilot CLI, providing semantic code intelligence for C++ projects. Powered by the IntelliSense engine from Visual Studio and VS Code, it delivers symbol definitions, references, call hierarchies, and type information to complement grep-style search. The language server is distributed as an npm package and requires a compile_commands.json file — a bundled skill handles this automatically for CMake projects. MSBuild support is planned for a future release. [5]

• GitHub Copilot now supports bring-your-own-key for language models in VS Code for Business and Enterprise users. Teams can reuse API keys from Anthropic, Gemini, OpenAI, OpenRouter, Azure, Ollama, and Foundry Local. BYOK models are available across VS Code Chat, the built-in plan agent, and custom agents — usage is billed by the chosen provider and does not count against GitHub Copilot request quotas. The policy is enabled by default, and administrators can disable it through the Copilot policy settings. [11]

• The Copilot usage metrics API now reports aggregated active and passive user counts for Copilot code review. Six new fields are available in both 1-day and 28-day reports at enterprise and organization levels, covering daily, weekly (trailing 7-day), and monthly (trailing 28-day) windows. Active users are those who manually requested a review or applied a suggestion; passive users are those whose reviews were auto-triggered by policy. When both signals are present, active always trumps passive — enabling organizations to distinguish intentional adoption from policy-driven usage. [12]

• Claude Code v2.1.117 ships forked subagents, native search tooling, and an Opus 4.7 context window fix. Forked subagents can be enabled with CLAUDE_CODE_FORK_SUBAGENT=1, and agent frontmatter mcpServers now load for main-thread agent sessions via --agent. On macOS and Linux native builds, the Glob and Grep tools are replaced by embedded bfs and ugrep available through the Bash tool for faster searches without a separate tool round-trip. The release fixes Opus 4.7 sessions computing against a 200K context window instead of the native 1M, which caused inflated /context percentages and premature autocompaction. Plugin marketplace enforcement, concurrent MCP startup, and OAuth token auto-refresh on 401 are among the other improvements. [8]

• OpenAI Codex shipped three Rust-based alpha releases (0.123.0-alpha.8 through alpha.10) on the same day, reflecting rapid iteration on the agent platform. The releases — published at 12:02 UTC, 21:56 UTC, and 23:09 UTC — carry minimal changelog entries, indicating continuous integration builds during active development of the Codex CLI. This cadence aligns with OpenAI's broader push to accelerate Codex's agentic capabilities. [13]

• Gemini CLI v0.39.0-preview.2 patches a subagent initialization bug in the broader v0.39.0 preview series. The cherry-picked fix addresses duplicate initialize calls on refreshed agents. The v0.39.0 line includes substantial features: a unified invoke_subagent tool replacing legacy wrappers, skill extraction that now requires recurrence evidence, MCP resource listing and reading tools, prompt-driven memory editing across four tiers, JSONL session logging, bundled ripgrep for offline support, and a security fix preventing IDE stdio override via workspace .env files. [14]

• OpenAI introduced Codex-powered workspace agents in ChatGPT, automating complex workflows that run in the cloud. Workspace agents help teams scale work across tools securely, extending Codex's agentic capabilities beyond the CLI into ChatGPT's interface. Separately, OpenAI published a technical deep dive showing how WebSockets and connection-scoped caching in the Responses API reduced overhead and improved model latency in the Codex agent loop, providing architectural insights into how agentic workflows are being optimized at scale. [15][16]