MCP server

The Model Context Protocol is the open standard Anthropic published for letting AI agents call tools on your services. Elido ships an MCP server so any compatible client — Claude Desktop, Cursor, ChatGPT, Continue, Zed — can shorten URLs, read click analytics, manage QR codes, and run bulk imports without copy-pasting between tabs.

There are two ways to wire it up:

Hosted (recommended): point your client at https://mcp.elido.app/v1/ with a personal access token. No install, always up-to-date.
Self-hosted: run the binary or container locally; useful for air-gapped environments and offline work.

1. Get a personal access token

In the dashboard, go to Settings → API keys → New key. Pick scope mcp:full (or mcp:readonly if you only want analytics, not mutations). Copy the token — it’s shown once.


ELIDO_MCP_TOKEN="elt_…"

The token is workspace-scoped. Switch workspace, switch token.

2. Hosted setup — Claude Desktop

Open ~/Library/Application Support/Claude/claude_desktop_config.json on macOS (Windows: %APPDATA%\Claude\claude_desktop_config.json) and add:


{
  "mcpServers": {
    "elido": {
      "type": "http",
      "url": "https://mcp.elido.app/v1/",
      "headers": {
        "Authorization": "Bearer elt_…"
      }
    }
  }
}

Restart Claude. The Elido tool surface should appear in the tool picker.

3. Hosted setup — Cursor

Cursor reads from the same ~/.cursor/mcp.json:


{
  "mcpServers": {
    "elido": {
      "url": "https://mcp.elido.app/v1/",
      "headers": { "Authorization": "Bearer elt_…" }
    }
  }
}

Reload the Cursor window. @elido invocation works in chat panels.

4. Hosted setup — ChatGPT

ChatGPT reads MCP servers from Settings → Connectors → Add custom connector. Paste https://mcp.elido.app/v1/ and set the authorization header to Bearer elt_…. ChatGPT will surface the tools in any new conversation.

5. Self-host

The MCP binary ships in the same Docker image as the rest of the self-host bundle:


docker run --rm -p 8765:8765 \
  -e ELIDO_API_BASE=https://api.elido.app \
  -e ELIDO_MCP_TOKEN=elt_… \
  ghcr.io/elidoapp/mcp-server:latest

Or with the npm wrapper if you want stdio transport (e.g. for an agent that prefers stdio):


npx -y @elido/mcp-server@latest

Configure the client to launch the local process:


{
  "mcpServers": {
    "elido-local": {
      "command": "npx",
      "args": ["-y", "@elido/mcp-server@latest"],
      "env": { "ELIDO_MCP_TOKEN": "elt_…" }
    }
  }
}

6. Tool surface

The server exposes 18 tools across 5 namespaces. The most-used:

Tool	Scope	What it does
`links.create`	`mcp:full`	Shorten a URL with optional UTM template + expiry
`links.list`	`mcp:readonly`	Paginated link list with click counts
`links.update`	`mcp:full`	Patch destination, slug, rules, tags
`analytics.summary`	`mcp:readonly`	Top countries / referrers / devices for a link or campaign
`analytics.timeseries`	`mcp:readonly`	Hour / day buckets, returns chart-ready data
`qrs.create`	`mcp:full`	Designer QR with logo + colors
`bulk.import`	`mcp:full`	Upload a CSV; returns per-row status
`webhooks.replay`	`mcp:full`	Re-fire a delivery from the audit log

Full schema is at the API reference — every MCP tool maps 1:1 to a public REST endpoint, so the docs you read for the API are the docs for the agent.

7. Safety

Read-only by default: pick mcp:readonly if the agent is exploring; bump to mcp:full only when you need writes.
Per-tool allow-list: in the dashboard you can restrict a token to a subset of tools, e.g. allow links.* but deny bulk.import (which can create thousands of links in one call).
Audit log: every MCP call lands in the workspace audit log with the agent identifier, the tool, and the arguments — same surface as a human dashboard action.
Rate limit: hosted server is capped at 60 calls/minute per token; self-hosted is uncapped.

8. Edge cases

Long-running calls — bulk.import for >5000 rows runs async; the tool returns a job id, and the agent polls bulk.status until done. No client-side timeouts.
Workspace switching — one token = one workspace. Issue a separate token per workspace; the MCP client’s connector list surfaces each as a distinct server.
Token rotation — issue a new token, update the config, then revoke the old one. Don’t reuse token slots; the audit log relies on stable token ids.