Install the MCP analytics SDK - AgentDiscoverability

Beta. The TypeScript package is on npm under the beta tag. The Python package is not on PyPI yet, and SDK keys are not being issued yet. For early access, email support@agentdiscoverability.com.

The SDK runs inside your MCP server and reports each initialize, tools/list and tools/call it handles. It only observes: it never changes your tools, their results or their speed.

1. Get an SDK key

Create one on the MCP page. It starts with adok_ and is shown only once. Store it as MCP_ANALYTICS_SDK_KEY and never commit it.

2. Install

npm install @agentdiscoverability/mcp-analytics@beta

pip install agentdiscoverability-mcp-analytics

3. Call `observe()`

Call it once, after your tools are registered and before the server starts.

import { observe } from "@agentdiscoverability/mcp-analytics";

observe(server, { sdkKey: process.env.MCP_ANALYTICS_SDK_KEY! });

import os
from mcp_analytics import observe

observe(
    server,
    sdk_key=os.environ["MCP_ANALYTICS_SDK_KEY"],
    # Required for now: the Python SDK has no default collector yet.
    collector_url="https://collector.agentdiscoverability.com",
)

server can be a Server or McpServer (TypeScript), or a FastMCP, MCPServer or Server (Python). observe() never throws. If the key is missing it logs a warning and leaves the server uninstrumented.

4. Verify

Connect any MCP client to your server. Within about a minute, the MCP page changes from Waiting for your first event… to Receiving · N events. You don’t need to call a tool first. To check locally, inspect the handle observe() returns. client.stats() should show sent above zero and dropped at zero.

Requirements

	Supported
TypeScript	Node ≥ 22, ESM, `@modelcontextprotocol/sdk` ≥ 1.11 and < 2
Python	Python ≥ 3.10, `mcp` ≥ 1.10 and < 3, or community `fastmcp` 3.x/4.x

Not supported: MCP SDK v2 (@modelcontextprotocol/server), CommonJS, Bun, Deno, Cloudflare Workers, or fastmcp 2.x.

What is collected

The tool name, the argument shape (key names and types, not values), duration and errors
The tool result, sanitized and truncated
The AI client family (Claude, ChatGPT, Cursor…) and versions
An end-user id only if you supply one. It is hashed before storage.

We never collect conversations, IP addresses, request headers, credentials or argument values. Data is kept for 90 days. See the Privacy Policy.

Troubleshooting

Symptom	Fix
Nothing arrives	Check the handle’s `enabled` is true, `observe()` ran before the server started, a client has connected, and (Python) `collector_url` is set.
`401`	The key is wrong or revoked. Create a new one on the MCP page.
`413` / `429`	The batch is too large, or you hit a rate limit. Lower `flushBatchSize` / `flush_batch_size`. The SDK retries `429`s on its own.
`503`	Ingest is paused on our side. There’s nothing to fix.
Behind a proxy	Allow direct HTTPS to `collector.agentdiscoverability.com:443`. Proxies and redirects are refused.

Still stuck? Email support@agentdiscoverability.com with your key prefix (never the full key) and the stats() output.

For coding agents

Find the file that creates the MCP server and registers tools.
Add the dependency, then the observe() call from step 3, after the tools are registered.
Read the key from MCP_ANALYTICS_SDK_KEY and keep it out of git.
Change nothing else. No tool name, schema or result may change.
Run the server, connect a client, and report the changed file and the stats() output.

​1. Get an SDK key

​2. Install

​3. Call observe()

​4. Verify

​Requirements

​What is collected

​Troubleshooting

​For coding agents

1. Get an SDK key

2. Install

3. Call `observe()`

4. Verify

Requirements

What is collected

Troubleshooting

For coding agents