Three product announcements this week all point in the same direction: enterprise AI is moving from individual developer tools to team-scale infrastructure. The Claude apps gateway for AWS gives organizations centralized control over Anthropic's Claude Code and Desktop clients. JetBrains AI for Teams and Organizations promises a vendor-agnostic control plane for agentic development. Slack connected Slackbot to Salesforce's entire platform through MCP, turning a chat client into an orchestration layer.
All three are solving a problem that has become unavoidable: when every developer has their own AI credentials, their own model settings, and their own cost baseline, the organization has no leverage. Spend leaks. Policy is aspirational. Onboarding and offboarding are manual. The industry is converging on a model where AI tools are governed by a central gateway, configured through identity, and instrumented with telemetry.
The Claude apps gateway: what it actually does
Anthropic and AWS shipped a self-hosted control plane that sits between developers and the inference endpoints. The gateway handles identity through any OIDC provider, routes requests to Bedrock or Claude Platform on AWS, enforces policies and spend caps, and sends telemetry over OpenTelemetry Protocol. It runs as a stateless container on ECS, EKS, or EC2, backed by a PostgreSQL database for session state.
The key design choice: the gateway lives inside the customer's private network, behind an internal ALB with a TLS certificate from ACM. Developers reach it through the private network, not over the public internet. The gateway holds the upstream credential (an IAM task role) and calls Bedrock or Claude Platform on behalf of each developer after validating their short-lived token.
Policies are defined once on the server. At sign-in, the CLI receives managed settings and the gateway enforces them on every request. The model picker shows only allowed models. Tool permissions -- file writes, web access -- can be restricted. Spend caps are per organization, group, or user, with daily, weekly, and monthly windows.
The concrete impact: onboarding is "add the developer to your IdP." Offboarding is "remove them from the IdP" and the session expires within the configured token lifetime (one hour by default). No long-lived secrets on developer machines.
This matters because the alternative is what most teams are doing today: each developer provisions their own API key, manages their own settings, and the CTO finds out about costs at the end of the month. The gateway replaces that with a single control point that connects to existing identity workflows.
JetBrains AI for Teams: the vendor-agnostic bet
JetBrains announced a broader initiative. JetBrains AI for Teams and Organizations connects external AI tools through Model Context Protocol (MCP) and external agents through Agent Client Protocol (ACP). The vision is a unified system for agentic software development where teams can run cloud agents, share context, and get centralized governance through something called JetBrains Central.
The product is not out yet. JetBrains plans to introduce capabilities gradually over the coming weeks. The announced features include team automations triggered by repository events, a context system that gives agents cross-repository intelligence, and a central CLI that brings tools like Claude Code, Codex, and Gemini CLI into the same organizational environment.
The interesting part is the vendor-agnostic posture. JetBrains is not building its own AI model. Instead, it wants to be the layer that routes between models and tools, enforcing governance and cost controls regardless of which underlying provider a developer picks. That is the right bet for anyone whose team uses multiple AI tools.
The practical challenge: JetBrains Central only works if teams actually adopt JetBrains tools. Organizations that are already on VS Code or Neovim may not switch. The company is betting that the governance and cost attribution features are compelling enough to pull teams onto their platform. I am not sure that works. Developers are stubborn about their editors.
Slackbot becomes a work operating system
The third announcement is the most ambitious because it is the most visible. Slack connected Slackbot to Salesforce's entire platform through dedicated MCP servers. A user can ask Slackbot for a customer's deal history, get a live Tableau visualization, update a CRM record, and trigger a DocuSign approval all within the chat window.
MCP is the technical backbone. Salesforce exposes CRM records, Tableau visualizations, Data 360 customer profiles, and Agentforce agents as MCP servers. Slackbot acts as an MCP client, discovering tools and routing queries to the appropriate backend. The bot respects each user's Salesforce permissions, so field-level security carries over automatically.
The operational model is multiplayer. Actions happen in shared channels. When Slackbot pulls a customer profile or flags a deal risk, the whole team sees it. A colleague can redirect the agent's work in real time. That is the structural difference from single-player AI tools. As Slack CEO Denise Dresser said, "work is a team sport" and "for AI to really take hold in the enterprise, it has to be multiplayer."
The MCP approach has tradeoffs. Tool discovery on every connection consumes context tokens. One technical analysis estimated that a server with 300 tools costs 5,000 to 10,000 tokens per session before the model does any useful work. Salesforce will need to filter and segment MCP servers carefully at enterprise scale.
There is also the partner tension. Anthropic, whose protocol MCP is, reportedly plans to expand Claude Tag to Microsoft Teams and email. The same partners Slack invites into its ecosystem are building their own surfaces for enterprise work. Slack's moat is the workspace context and the permissions layer that comes from being the communication hub. Whether that is enough depends on how quickly competitors build equivalent orchestration layers.
Practical takeaways
The three announcements share a pattern. Each is a gateway that sits between users and AI infrastructure, controlling access, enforcing policy, and tracking cost. Each connects to existing identity providers rather than creating new credential systems. Each uses telemetry to give organizations visibility into usage.
For teams deploying AI tools today, the immediate question is which gateway model fits. The Claude apps gateway is specific to Anthropic's tools. JetBrains Central is tool-agnostic but platform-specific. Slackbot is chat-agnostic but ecosystem-specific. None of them solve the problem of managing multiple gateways simultaneously.
The deeper takeaway is that AI governance is following the same trajectory as cloud governance. In 2015, teams provisioned individual cloud accounts and managed access manually. By 2020, everyone used identity-aware proxies, service meshes, and cost management platforms. AI tools are going through the same cycle at higher speed. The Claude gateway, JetBrains Central, and Slack's MCP integration are the first generation of that infrastructure.
The hard part is not the technology. The hard part is that governance only works if developers actually use the tools through the gateway. If a developer can bypass the gateway by pasting an API key into a terminal, the controls are useless. The organizations that get this right will be the ones that make the gateway path frictionless and the bypass path difficult. The vendors who understand that will win the next phase of enterprise AI.