Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.equinix.dev/llms.txt

Use this file to discover all available pages before exploring further.

The Equinix Fabric MCP is a real, Equinix-operated MCP server at mcp.equinix.com/fabric. It exposes the Fabric API as agent-friendly tools — search/get for reads, create/update/delete for mutations, with the mutating tools blocked-by-default until human confirmation. equinix.dev’s CLI uses the same MCP under the hood. So does your AI runtime, once you’ve installed the agent skills and connected your client.

​
What the MCP gives you

  • Discovery — search_router, search_connection, search_service_profile, etc.
  • Inspection — get_router_package, get_connection, get_service_profile.
  • Validation — validate_route_filter, validate_stream_sink, validate_terraform_block.
  • Pricing — quote_router, quote_connection (gives you a JSON shape compatible with the pricing.md spec).
  • Mutations (blocked by default) — create_router, create_connection, update_router, delete_connection.

​
Connect your client

Add the Fabric MCP to your client config: 
{
  "mcpServers": {
    "equinix-fabric": {
      "url": "https://mcp.equinix.com/fabric/sse",
      "headers": { "Authorization": "Bearer ${EQUINIX_TOKEN}" }
    }
  }
}
The same configs are in the Quickstart — single source of truth.

​
How equinix-dev uses the MCP

When you run equinix-dev plan, the CLI executes this sequence on the MCP: 
1. search_router(metro_code, package)         → discover what's possible
2. search_service_profile(metro_code, name_like) → find partner endpoints
3. get_router_package(package)                → look up pricing + limits
4. validate_terraform_block(hcl)              → server-side syntax check
5. quote_router(package, metro_code)          → estimated monthly cost
                                              → write to pricing.md spec
                                              ───────────────────────
6. create_router(...) [BLOCKED]               → would mutate; refused
                                              → returns preflight gates
The mutating step is intercepted by the gateway. The agent surfaces the proposed plan + the preflight gate list back to the human, who runs terraform apply outside the equinix.dev shell after confirming.

​
Authoring Terraform from natural language

This is the canonical agent-driven flow: 
USER:  Design a private AI inference path in IAD to Lambda. Plan only.

AGENT: ⏵ search_router(metro_code="DC", package="BASIC")
       ⏵ search_service_profile(metro_code="DC", name_like="Lambda")
       ⏵ validate_terraform_block(<proposed HCL>)
       ⏵ quote_router(package="BASIC", metro_code="DC")
       ⏵ Drafting Terraform...

       I've drafted a Fabric Cloud Router (BASIC, $145/mo) plus a
       1Gbps connection to Lambda Labs Private AI in IAD ($310/mo).
       Total: $455/mo. The full HCL is below. Apply is blocked until
       these preflight gates pass:
         - equinix_account_authorized
         - ssh_key_uploaded
         - billing_owner_acknowledged

USER:  Add a Network Edge FortiGate in line.

AGENT: ⏵ search_module(kind="network_edge_device")
       ⏵ get_network_edge_device_package(type_code="FG", package_code="VM02")
       ⏵ Updating Terraform...

       Added a FortiGate VNF (VM02, 2 cores, $1,150/mo). Total: $1,605/mo.
No actual Equinix resources were created in the trace above. All four traces would be visible in the Workspace request log.

​
Mutation policy

The Fabric MCP exposes its mutation policy explicitly: 
{
  "tool": "create_router",
  "arguments": { ... },
  "result": {
    "status": "BLOCKED",
    "reason": "mutation_policy = blocked_by_default_requires_human_confirmation",
    "preflight_gates": [
      "equinix_account_authorized",
      "ssh_key_uploaded",
      "billing_owner_acknowledged"
    ]
  }
}
The result is a gateway-level refusal — the MCP server doesn’t even attempt the API call. This is enforced by the Equinix MCP infrastructure, not by equinix-dev or your agent. You can read the full policy at docs.equinix.com/equinix-api/mcp-servers/overview.

​
Skills bundle

The @equinix/agent-skills bundle teaches your AI runtime which MCP tool to call for which intent. Without the skills, an agent can still hit the MCP, but it’ll guess at tool selection. With the skills, the agent has structured guidance — “for ‘design a private AI path,’ call search_router first” — that lifts response quality immediately. Install: 
npx skills add ./equinix-agent-skills

​
What the MCP doesn’t do

  • It doesn’t run terraform plan or terraform apply for you. You still hold the Terraform tooling.
  • It doesn’t store secrets. EQUINIX_TOKEN is per-request; the MCP never persists it.
  • It doesn’t talk to non-Equinix providers. AWS / Azure / GCP MCPs are separate products with their own endpoints.

​
Next

Recipes

Three full Use Cases — each comes with the MCP trace an agent walks through.

Safety model

The six controls that keep the MCP plan-only by default.