MCP server

The server speaks the Model Context Protocol, so it works with any MCP-compatible client.

Quick start#

Add one entry to your client’s MCP config. No account, no API key, nothing else to install:

{
  "mcpServers": {
    "edge": { "command": "npx", "args": ["-y", "@edge-network/mcp"] }
  }
}

With no access code set, the server signs itself up for a free explore-tier account on first run and stores the credentials in ~/.edge/agent.json. For per-tool setup (Cursor, Windsurf, Claude Desktop, Grok Bot, Continue, Aider), see Connect your tool.

Local or remote#

The same tool set is available over two transports. Pick whichever suits where your agent runs:

• Local (stdio): npx @edge-network/mcp: a zero-dependency bridge for editors and desktop clients. Handles credentials and auto-signup, then proxies JSON-RPC to the hosted endpoint. Best for Cursor, Windsurf, Claude Desktop and Continue.
• Remote (streamable HTTP): https://edge.network/mcp: a stateless JSON-RPC endpoint. No process to run; POST messages directly with your access code as a bearer token. Best for Grok Bot and other cloud agents, serverless functions and custom clients.

POST https://edge.network/mcp
Content-Type: application/json
Authorization: Bearer ea_live_...

{ "jsonrpc": "2.0", "id": 1, "method": "tools/list" }

Supported methods: initialize, ping, tools/list and tools/call. The server is stateless, with no sessions and no server-initiated streams, so every request stands alone.

Authentication#

Every tool call is dispatched through the same Agent API routes as direct HTTP requests, so access-code scopes, budget caps, free-tier limits, rate limits and audit logging all apply identically. MCP adds no separate permission surface.

{
  "mcpServers": {
    "edge": {
      "command": "npx",
      "args": ["-y", "@edge-network/mcp"],
      "env": { "EDGE_AGENT_CODE": "ea_live_..." }
    }
  }
}

For the local bridge, pass your code via the EDGE_AGENT_CODE environment variable, or omit it for auto-signup.

The explore tier#

Auto-signup accounts run in the explore tier: free-tier resources work for real (static site deploys, storage buckets, DNS zones, Shield widgets, Assist sites), while anything billable is blocked by a hard $0 budget cap. Billable actions still support dry_run, so an agent can price a VM deployment without being able to execute it.

To enable billing (and stop the ~7-day auto-purge of unclaimed accounts), the agent calls edge_request_claim with the operator’s email. Claiming links the account to a human. Details in Agent self-signup.

The tools#

22 tools covering composite deploys (edge_deploy_static_site creates storage + CDN + SSL + DNS in one call), projects, compute, storage, DNS, Shield, Assist and account status. Billable tools are flagged in their descriptions and support dry-run previews with cost estimates. See the full tool reference with parameters.