@inferlane/agent-hooks

Drop-in lifecycle hooks that auto-log tokens, cost, active runtime, and web-search charges from Claude Agent SDK sessions (and Anthropic Managed Agents, Goose, SwarmClaw, or any runtime that exposes PreToolUse / PostToolUse / SessionStart / SessionEnd events) to InferLane.

Zero tool calls from the agent. No log_request spam in the model's output. The hooks run transparently alongside your agent loop and post events to InferLane's fleet API in the background.

Install

npm install @inferlane/agent-hooks

Quick start — Claude Agent SDK

import { query } from '@anthropic-ai/claude-agent-sdk';
import { claudeAgentHooks } from '@inferlane/agent-hooks';

const hooks = claudeAgentHooks({
  apiKey: process.env.INFERLANE_API_KEY!,
  runtime: 'CLAUDE_AGENT_SDK',
  agentName: 'my-coding-assistant',
  fleetId: 'fleet_abc123', // optional — attach to a named fleet for rollups + budget alerts
});

for await (const msg of query({
  prompt: 'Refactor the auth module',
  options: { hooks },
})) {
  // handle agent messages
}

That's it. Every tool call, web search, and session completion is logged to InferLane. The dashboard shows token cost + runtime cost ($0.08/hr for Managed Agents) + web-search cost ($10/1000 for Managed Agents) in real time.

What gets tracked

| Hook | Recorded data | |---|---| | SessionStart | Opens a fleet session (POST /api/fleet/sessions) | | PreToolUse | Starts the active-runtime timer; posts TOOL_USE event | | PostToolUse | Stops the timer; posts TOOL_RESULT event with token usage and active-runtime delta. Detects web_search / WebSearch calls and bumps the web-search counter. | | SessionEnd | Marks the session COMPLETED |

Runtime cost is only charged for ANTHROPIC_MANAGED sessions (Anthropic bills $0.08 per active session-hour on Managed Agents; local Agent SDK runtime is free). Web-search cost is only charged when the runtime is ANTHROPIC_MANAGED.

Advanced: granular control

If you'd rather manage hook registration yourself:

import { createInferLaneHooks } from '@inferlane/agent-hooks';

const il = createInferLaneHooks({
  apiKey: process.env.INFERLANE_API_KEY!,
  runtime: 'CLAUDE_AGENT_SDK',
  debug: true,             // log hook activity to console
  swallowErrors: true,     // default — don't break the agent if telemetry fails
});

const options = {
  hooks: {
    SessionStart: [il.sessionStart],
    PreToolUse:   [il.preToolUse],
    PostToolUse:  [il.postToolUse],
    SessionEnd:   [il.sessionEnd],
  },
};

// On graceful shutdown:
process.on('SIGTERM', async () => {
  await il.flush(); // closes the active session cleanly
});

Runtime field values

| Value | Use for | |---|---| | ANTHROPIC_MANAGED | Anthropic's cloud-hosted managed agent runtime | | CLAUDE_AGENT_SDK | Self-hosted Claude Agent SDK | | CLAUDE_CODE | Claude Code CLI / VS Code extension | | GOOSE | Block's Goose agent | | SWARMCLAW | SwarmClaw orchestrator | | CUSTOM | Anything else — self-reported |

Environment variables

| Variable | Purpose | |---|---| | INFERLANE_API_KEY | Bearer token for InferLane fleet API. Can also be passed via apiKey option. | | INFERLANE_BASE_URL | Override base URL (default https://inferlane.dev). Useful for self-hosted InferLane. |

Privacy

Hook options include debug: true which logs tool names and session IDs to console.log. Do not enable debug in production — tool names and payloads may contain user data.

The hooks never ship the contents of user prompts or tool inputs to InferLane by default. Only token counts, tool names, active-runtime durations, and web-search counts are recorded. To include tool payloads for debugging purposes, see the advanced configuration.

License

MIT