control-tower-sdk (Node.js)

Report your agent's status, tokens, cost and tool calls to the Agent Control Tower.

Install

# from a published registry (once published)
npm install @intelli-1113/control-tower-sdk
# or directly from git / a local tarball
npm install git+https://your.git/control-tower-api.git#path:sdk/control-tower-js
npm install ./control-tower-sdk-1.0.0.tgz     # after `npm pack`

Use (zero-config via env)

export CONTROL_TOWER_URL=http://your-tower:3090
export CONTROL_TOWER_AGENT_ID=weather-agent

import { ControlTower } from '@intelli-1113/control-tower-sdk';

const ct = new ControlTower({ name: 'Weather Agent', framework: 'LangChain', model: 'gpt-4o' });
ct.start({ tools });                       // begins heartbeats → agent auto-registers in the tower

// capture an LLM call's tokens
const res = await ct.track(client.messages.create({ /* ... */ }));

// task lifecycle + tool calls
ct.taskStart('Weather in Tokyo');
ct.toolCall('get_weather', { city: 'Tokyo' });
ct.taskComplete('Sunny, 24°C');

// optional LangChain integration (needs @langchain/core)
const cb = await ct.asLangChainCallback();

serverUrl/agentId come from CONTROL_TOWER_URL/CONTROL_TOWER_AGENT_ID if not passed. The admin can then enrich the agent's metadata and map it to a Business Unit from the tower UI.

Core SDK has no required dependencies (@langchain/core, llama3-tokenizer-js are optional).

OpenTelemetry mode (zero-code telemetry + heartbeat)

Prefer auto-instrumentation over wrapping calls? Launch your agent with the bundled OTel preload — no code changes, works with any framework:

# 1. install OTel + an instrumentation for your framework
npm i @opentelemetry/sdk-node @opentelemetry/exporter-trace-otlp-http
npm i @arizeai/openinference-instrumentation-langchain   # LangChain JS
# (or one line:  npm i @traceloop/node-server-sdk  — auto-covers OpenAI/Anthropic/LangChain)

# 2. run with the preload (must load BEFORE your app)
node --import @intelli-1113/control-tower-sdk/otel your-app.js

Reads the same env (CONTROL_TOWER_URL, CONTROL_TOWER_AGENT_ID, CONTROL_TOWER_NAME, CONTROL_TOWER_DEPARTMENT). It:

• exports GenAI traces (tokens, cost, latency, tool calls, lineage) to <CONTROL_TOWER_URL>/v1/traces, and
• starts a heartbeat so the agent shows online/offline even when idle.

In OTel mode, don't also call ct.track()/asLangChainCallback() — OTel captures the tokens, and doing both double-counts. Keep explicit SDK calls only for escalations/triggers. If the OTel packages aren't installed, telemetry is skipped but the heartbeat still runs.

MCP proxy (live MCP usage metering)

This package also ships control-tower-mcp-proxy — a transparent MCP stdio proxy. Put it between an MCP host (e.g. Claude Desktop) and the real server to meter every tool call (tokens, latency, errors) live:

"weather": {
  "command": "npx",
  "args": ["-y","-p","@intelli-1113/control-tower-sdk","control-tower-mcp-proxy","--","npx","-y","@scope/weather-mcp@latest"],
  "env": { "CONTROL_TOWER_URL":"http://localhost:3090","CONTROL_TOWER_AGENT_ID":"weather-mcp" }
}

Python SDK

The Python package (control_tower) lives under python/ — pip install ./python.

This repo contains the client SDKs for the Agent Control Tower (the server lives in a separate repo). Publish: npm publish (JS) · cd python && python -m build && twine upload dist/* (PyPI).