Skip to content

mcp-server.md

Latest commit

 

History

History
181 lines (135 loc) · 6.78 KB

mcp-server.md

File metadata and controls

181 lines (135 loc) · 6.78 KB
Raw

Cotlet MCP server

← Back to README · Full docs: cotlet.dev/docs/mcp (Markdown: cotlet.dev/docs/mcp.md)

One remote MCP server gives your agent every app on your account, a read tool and a write tool per app. Add it once; new apps show up as you enable them for agents.

Server

Setting Value
URL https://cotlet.dev/mcp/v1
Transport Streamable HTTP
Auth Authorization: Bearer COTLET_API_KEY
URL for OAuth sign-in https://cotlet.dev/mcp/v1/connect (Claude.ai, Claude Desktop, ChatGPT, Claude Code without a key)
Protocol versions 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05

Connect a client

Claude Code

claude mcp add --transport http cotlet https://cotlet.dev/mcp/v1 \
  --header "Authorization: Bearer $COTLET_API_KEY"

Cursor

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "cotlet": {
      "url": "https://cotlet.dev/mcp/v1",
      "headers": { "Authorization": "Bearer ${env:COTLET_API_KEY}" }
    }
  }
}

Codex

# ~/.codex/config.toml
[mcp_servers.cotlet]
url = "https://cotlet.dev/mcp/v1"
bearer_token_env_var = "COTLET_API_KEY"

Cline

// cline_mcp_settings.json
{
  "mcpServers": {
    "cotlet": {
      "type": "streamableHttp",
      "url": "https://cotlet.dev/mcp/v1",
      "headers": { "Authorization": "Bearer <your key>" }
    }
  }
}

Claude Desktop

Settings > Connectors > Add custom connector
Name: Cotlet
URL:  https://cotlet.dev/mcp/v1/connect

Then Connect and sign in to Cotlet.

ChatGPT

Settings > Apps & Connectors > Create
MCP server URL: https://cotlet.dev/mcp/v1/connect
Authentication: OAuth

n8n

Node: MCP Client (MCP Client Tool inside an AI Agent)
Server Transport: HTTP Streamable
MCP Endpoint URL: https://cotlet.dev/mcp/v1
Authentication: Bearer Auth
  Token: <your key>
Tool: skool_read
Input (JSON): {"operation": "list_groups", "arguments": {}}

Make

App: MCP Client > Call a tool
MCP Server: + New MCP Server
  URL: https://cotlet.dev/mcp/v1
  API Key / Access token: <your key>
Tool Name: skool_read
Arguments: {"operation": "list_groups", "arguments": {}}

Zapier

Action: MCP Client by Zapier > Run Tool
Server URL: https://cotlet.dev/mcp/v1
Transport: Streamable HTTP
OAuth: No
Bearer Token: <your key>
Tool: skool_read
  operation: list_groups
  arguments: {}

In n8n, Make and Zapier pick the app's read or write tool, such as skool_read. Put the operation in operation and its inputs in arguments. An API key is enough there; OAuth is not needed.

No key in Claude Code? Add https://cotlet.dev/mcp/v1/connect without the header, run /mcp and sign in to Cotlet in the browser. The client gets its own credential, listed on the API keys page, where you can disconnect it like a key.

Product names are trademarks of their owners. Cotlet is not affiliated with them.

Tools

Each app has two tools named like the app, with dashes as underscores: skool_read for the endpoints that read and skool_write for the ones that change data. An app that only reads has no write tool. So 70 endpoints cost your agent two tools, not 70.

  • {"operation": "list_members", "arguments": {...}} runs an operation. Its arguments are the same JSON Schema as the REST endpoint.
  • Each tool's input schema carries every operation's arguments, so your agent needs no extra call to learn them.
  • Signed in, tools/list returns the apps you enabled for agents, within the key's scope; a read-only account or key gets no write tool. Without a key it returns every app's two tools with their operation counts, no schemas.
  • Read tools carry readOnlyHint: true, so clients can run them without asking. Write tools carry destructiveHint: true, so clients ask before each call.
  • Each app is also a resource, cotlet://apps/{app}, with its description and endpoints.
  • Want fewer operations in front of your agent? Create a key for only the apps (or operations) you use.

Account tools, all free: cotlet_whoami (which account a connection uses), cotlet_apps (which apps work now and what the user still has to do), cotlet_request_apps, cotlet_balance, cotlet_get_call (outcome of a call waiting for approval), cotlet_report_issue (a bug or missing feature, straight to the Cotlet team) and cotlet_upload_file (a local file for tools that take one, up to 25 MB).

Calling a tool

Agents call tools for you; this is what goes over the wire:

curl -X POST https://cotlet.dev/mcp/v1 \
  -H "Authorization: Bearer $COTLET_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call",
       "params": {"name": "skool_read", "arguments": {"operation": "list_members", "arguments": {"group": "yourgroup"}}}}'

The result carries the output twice: as structuredContent for code and as JSON text for the model.

Errors

A failed tool call returns a result with isError: true and a message your agent can act on, with the same codes as the REST API:

{
  "content": [{ "type": "text", "text": "AUTH_EXPIRED: Skool signed the user out. Ask the user to sign in to the app again in their Cloud Browser at browser_url, then call again. browser_url: https://cotlet.dev/app/browser" }],
  "isError": true
}

Approval

Tools that reach other people (messages, posts, accepting members) wait for your approval by default. The call returns status: pending with a call_id instead of a result.

  • Approve or reject it in the console under Activity, Approvals. Without a decision it lapses after 24 hours.
  • Your agent checks the outcome with cotlet_get_call and the call_id.
  • For automations, set a tool to "Run without approval" in Activity, Approval rules.

Drive your Cloud Browser

When no Cotlet tool does the job, your agent can work in your Cloud Browser itself, billed per started minute of an open session.

  • Browser tools: cotlet_browser_open opens the app's browser, then cotlet_browser_navigate, _read, _screenshot, _scroll, _click, _type, _select, _press, _upload and _close. No extra software on your side.
  • Chrome DevTools: cotlet_browser_connect with {"app": "skool"} returns cdp_url, a DevTools WebSocket already signed in to the app. Connect Playwright (chromium.connectOverCDP(cdp_url)), Puppeteer or npx chrome-devtools-mcp@latest --wsEndpoint=<cdp_url>. The URL works for new connections for one hour.
  • cotlet_browser_sessions lists open sessions with what they cost so far and closes them.

MCP or REST?

Use MCP when an agent decides what to call. Use REST at https://cotlet.dev/api/v1 when your own code does: same key, same inputs, same outputs.