Cloudflare / Customer Zero Cloudflare Connect SF 2026

Rollout plan

This is the rollout order we'd use today. We did it differently because some of these features weren't available yet.

  1. Model access

    • Put AI Gateway behind Cloudflare Access on a custom domain.
    • Store provider keys in the gateway. Clients send only their Access token.
    • Set spend limits: a budget per person, plus a tighter budget for the most expensive models.
    • Turn on request logs. Agree with security and legal on who can read them and how long to keep them.
    • If a model is meant for a smaller group, confirm who can reach it before you grant access.
    • Point your coding tools at the gateway.

    Move on when

    • Approved users can connect.
    • Usage is attributable to each user.
    • The gateway blocks requests after a budget is reached.

    Guides: Access on AI Gateway, Stored keys, Spend limits, Logging, Coding tools

  2. Defaults and tools

    • Connect your first existing MCP servers to an MCP Server Portal. Keep per-user authentication enabled and confirm tool calls appear in MCP Portal logs.
    • Set the portal's Code Mode policy to On by default. Use Enforced if no client runs its own Code Mode.
    • Distribute client defaults for models, the MCP Portal URL, system prompts and skills, using the configuration options each client supports.
    • Optionally, serve configuration through a Worker API. Static hosting also works for simple client configuration. Keep credentials out of shared configuration.
    • Include the defaults in client setup and test that local overrides still respect server-enforced access policies and spend limits.

    Move on when

    • Setup supplies the portal automatically.
    • Tools keep each user's existing permissions.
    • MCP Portal logs show who called each tool.

    Guides: MCP Server Portals, Code Mode, Hosted OpenCode configuration

  3. Expand

    • Add more internal MCP servers behind the same portal as teams ask for them.
    • Build custom MCP servers on Workers when existing servers do not cover your internal systems. For engineers, use the caller's own account in the downstream system.
    • When you connect production agents, such as CI pipeline agents or background workers, give each one its own credentials and budget rule. Add AI Gateway metadata tags for agent ID, project and pipeline.

    Move on when

    • New approved servers appear without engineers changing their client config.
    • Each production agent's model requests appear under its metadata tags, and its budget rule blocks requests when reached.

    Guides: Remote MCP servers on Workers, Custom metadata, Service tokens

  4. Tune

    • Use User Insights to identify high-cost usage. Test cheaper models on those tasks and compare the results.
    • Try Auto Router and Workers AI models on those workloads.
    • For lower-cost fallbacks, add a dynamic route that switches models when the primary model's budget runs out. Keep an overall budget to limit total spending.

    Move on when

    • You have compared outcomes, conversation turns and total task cost with the current setup, before widening use.

    Guides: User Insights, Auto Router, Dynamic routing, Workers AI

Next: Adding a coding tool

At any phase, score a new tool with the evaluation checklist before you connect it. A missing requirement in the first three sections means it fails.