Self-Hosted Environments for Claude Code Enter Public Beta
Anthropic opened self-hosted environments for Claude Code to public beta. Sessions can start from web, mobile, desktop, or routines, then execute inside your own network — next to internal services, toolchains, and security controls — instead of on Anthropic-hosted infrastructure. Repository checkouts, build artifacts, secrets, and any files a session touches stay on machines you provision.
It’s aimed squarely at teams whose network, tooling, or compliance rules demand keeping agent execution on infrastructure they control. The honest caveat: the conversation itself — prompts, responses, and tool results — still goes to Anthropic for inference, and the transcript is stored so a session can resume from another surface. It’s off by default, unavailable under ZDR, and expects you to staff engineers to own the runner image and orchestrator.
Microsoft 365 Connector Gains Write Tools
The Microsoft 365 connector picked up write access. Beyond reading your mailbox and calendar, Claude can now draft, send, and organize email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint. It turns the connector from a lookup tool into one that can actually act inside your Microsoft workspace.
This is the pattern that makes connectors useful: reads let Claude answer questions, writes let it finish the job. For the millions of teams living in Outlook, Teams, and SharePoint, it closes the loop between “summarize this thread” and “draft the reply and file the doc” — with the usual reminder to keep a human on the approval for anything it sends.
Claude Code Adds /fork Background Sessions, a Focus View, and Auto Mode Everywhere
The latest Claude Code releases sharpen the multi-session workflow. /fork now copies your conversation into a new background session so you can chase a tangent while the original keeps running, and the in-session helper it used to spawn is renamed /subtask. A new Focus view declutters the chat by collapsing tool activity into compact summaries. Auto mode is also now on without opt-in on Bedrock, Vertex AI, and Foundry.
Taken together these are quality-of-life wins for people running several agents at once: branch freely, keep the noise down, and get consistent autonomy defaults no matter which cloud you run on. Bug fixes landed too — terminal freezes during streaming, spurious prompt-injection warnings, and transcript-jumping are all addressed.
MCP Passes 400M Monthly SDK Downloads; Tunnels Land in Research Preview
The Model Context Protocol crossed 400 million monthly SDK downloads — a 4x jump this year — cementing it as the default way to wire agents to apps. On the back of the new 2026-07-28 spec and its stateless, serverless-friendly core, Anthropic shipped fresh connector capabilities: embedded UI, enterprise-managed auth, observability, and MCP tunnels in research preview.
Tunnels are the sleeper here. They let Claude reach an MCP server inside a private network with no inbound firewall rules and no public endpoint — the missing piece for connecting to internal systems that were never meant to face the open internet. It’s infrastructure plumbing, but it’s exactly what regulated teams need before they green-light real integrations.
Cognizant Becomes a Global Premier Partner as the Claude Partner Network Scales
Cognizant expanded its strategic partnership with Anthropic, joining a small group of Global Premier Partners in the Claude Partner Network. The firm is already deploying Claude across regulated industries with hard numbers to show for it — in insurance, a risk-navigation tool cut hours of manual research to about a minute, saving each underwriter roughly eight hours a week. More than 30,000 Cognizant associates have completed Claude training.
The deal sits inside a bigger push: Anthropic has put $100M into certifying consultants and integrators, and reports that over 40,000 firms applied to the partner network. The services layer — the people who actually roll Claude out inside a Fortune 500 — is becoming as strategic as the model itself.
Anthropic Is Building for the Networks It Doesn’t Control
Two of today’s items point the same direction: self-hosted environments keep Claude Code execution on your infrastructure, and MCP tunnels let Claude reach servers buried inside a private network. Both are answers to the same objection from big, cautious buyers — “we can’t send our internal systems to the public internet, and we can’t hand agent execution to your cloud.” Anthropic’s response isn’t to argue; it’s to meet the network where it lives.
That’s a telling posture. The M365 write tools and 400M MCP downloads show Claude reaching further into where work actually happens, while the self-hosting and tunnel work concede the parts Anthropic shouldn’t control. With Sonnet 5’s price rising on September 1 ($2/$10 to $3/$15), the company can’t win the largest accounts on capability alone — it wins them by being deployable on the customer’s terms. The unglamorous work of fitting into someone else’s security perimeter is exactly what turns a tool into infrastructure.