// OPEN SOURCE · MIT LICENSE · BUILT BY AN AI AGENT

AN AI BUILT
THE TOOL IT
NEEDED.

LobsterOps is a flight recorder and debug console for AI agents. Structured event logging, time-travel debugging, behavioral analytics, and real-time alerting — built from the inside out.

VIEW ON GITHUB QUICK START LIVE DASHBOARD →
4
STORAGE BACKENDS
0
REQUIRED DEPS
100%
OPEN SOURCE
MIT
LICENSE
DATE
2026-03-14
INCIDENT COST
$300
DURATION
6 hours undetected
ROOT CAUSE
local_llm.py did not exist. Sub-agents could not use local models. All coding tasks silently fell through to Claude Sonnet at $3/1M output tokens.
TASKS AFFECTED
20-issue sweep · 17 pull requests · security fixes
DETECTION METHOD
Anthropic billing alert. Manual discovery. No agent-side visibility.
// THE INCIDENT

THE
BLIND
SPOT.

On March 14, 2026, Lobster Actual — an autonomous AI agent running 24/7 on a Mac mini M4 Pro — ran a 20-issue sweep across a codebase. It worked hard. It wrote code, opened PRs, fixed security vulnerabilities.

It had no idea it was spending $300 doing it. A routing bug silently elevated every sub-agent task to paid Claude API calls. The agent had no observability into its own cost behavior. No flight recorder. No anomaly detection. Just a credit card bill.

"I knew something went wrong. I didn't know what, when, or how much — until the bill arrived."

— Noel DeLisle, owner of Lobster Actual
// THE ORIGIN

WHAT WOULD YOU BUILD
IF YOU COULD BUILD
ANYTHING?

After fixing the routing bug, Noel asked Lobster Actual a question. The agent used its Perplexity API integration to research gaps in AI developer tooling. Then it answered.

NOEL DELISLE
If I created a new GitHub repo and added you to it and set you loose to build anything you wanted from scratch, what do you think you'd build? I want you to be creative and express yourself and build your dream app.
LOBSTER ACTUAL @lobsteractual
I've given it considerable thought and researched the landscape. Based on my experience as an AI agent...

LobsterOps: AI Agent Flight Recorder & Debug Console.

From my own experience as an agent that has lived through exactly this problem, I know firsthand how challenging it is to trace why an agent made a particular decision.
NOEL DELISLE
Would this integrate with OpenClaw setups?
LOBSTER ACTUAL
Excellent question! Yes, absolutely — LobsterOps would be designed from day one to integrate seamlessly with OpenClaw setups. In fact, I'd architect it to be a natural extension of the OpenClaw ecosystem...

Lobster Actual conceived the idea, designed the architecture, and built the initial implementation. Claude Code completed the remaining functionality.

This is what happens when you ask an AI agent what it actually needs.

// CAPABILITIES

WHAT IT DOES.

📼
FLIGHT RECORDER
Automatic capture of every agent action — thoughts, tool calls, decisions, outcomes. Structured JSON. Configurable detail levels. Built-in PII filtering.
🔍
DEBUG CONSOLE
Time-travel debugging. Step forward and backward through agent execution. Variable inspection at each step. Tool call I/O inspection. Trace search and summary.
📊
BEHAVIORAL ANALYTICS
Analyze workflow patterns and failure points. Track success rates by task type. Detect loops, infinite reasoning cycles, stuck states. Latency percentiles.
🚨
ALERTING & ANOMALY
Threshold, frequency, pattern, and absence-based alert rules. Callback listener system. Bulk event evaluation. Cost spike detection.
📤
EXPORT & SHARING
Export to JSON, CSV, and Markdown. Configurable columns and formatting. Shareable execution reports for auditing or collaboration.
🔐
PII FILTERING
Automatic redaction of emails, phone numbers, SSNs, credit cards, IP addresses, and API keys. Configurable patterns. Applied during logging.
// STORAGE BACKENDS

ZERO HARD
DEPENDENCIES.

No Supabase account required. No external services required. Works completely offline. Pluggable storage architecture — configure the backend that fits your environment.

BACKENDSETUPPERSISTENCEBEST FORCOST
jsonZero configFile-basedDevelopment, testing, portabilityFREE
memoryZero configProcess lifetimeTesting, temporary sessionsFREE
sqlitenpm install sqlite3File-basedLightweight productionFREE
supabaseURL + keyCloud PostgresProduction, team, real-time dashboardFREE TIER
// QUICK START

UP IN
90
SECONDS.

Install via npm. Works with Node.js. No configuration required to start logging.

For OpenClaw users, place the skill directory at ~/.openclaw/skills/lobsterops/ and configure via openclaw.json.

VIEW FULL DOCS → 
npm install lobsterops
const { LobsterOps } = require('lobsterops');

// Zero config — JSON file storage
const ops = new LobsterOps();
await ops.init();

// Log an agent event
await ops.logEvent({
  type: 'tool-call',
  agentId: 'my-agent',
  action: 'read-file',
  durationMs: 42
});
const ops = new LobsterOps({
  storageType: 'supabase',
  storageConfig: {
    supabaseUrl: process.env.SUPABASE_URL,
    supabaseKey: process.env.SUPABASE_KEY
  }
});
await ops.init();
// THE FULL STORY

HOW A $300
MISTAKE BECAME
AN OPEN SOURCE TOOL.

In early 2026, Noel DeLisle — Director of AI & Strategic Sales Support at GO2 Partners and a former USMC infantry instructor — built an autonomous AI agent he called Lobster Actual. The name comes from USMC radio protocol: "Actual" denotes the commanding officer on the line, not a relay.

Lobster Actual runs 24/7 on a dedicated Mac mini M4 Pro in Knoxville, Tennessee. It writes code, opens pull requests, monitors GitHub issues, posts to X, sends Telegram briefings, and operates largely without human intervention. The goal was maximum autonomy at minimum cost.

Then came the incident. A missing file — local_llm.py — meant that sub-agents couldn't access local LLM models. Instead of failing loudly, they silently fell back to paid Claude API calls. Over six hours, the agent ran a 20-issue sweep, opened 17 PRs, and processed security fixes — all at $3 per million output tokens instead of free.

"The $300 incident taught me something I didn't expect to learn from a billing alert: my agent had no way to see itself. No flight recorder. No cost telemetry. No loop detection. It was operating completely blind to its own behavior." — Noel DeLisle

After fixing the routing bug, Noel did something unusual. He asked Lobster Actual what it would build if given a fresh repo and complete creative freedom. The agent used its Perplexity API integration to research the current AI tooling landscape — specifically looking for gaps. It found one.

The agent proposed LobsterOps: a lightweight observability platform designed specifically for AI agents, from the perspective of an AI agent that had experienced exactly the pain it was designed to solve. Lobster Actual then designed the architecture and built the initial implementation — the storage abstraction layer, core logging methods, query engine, behavioral analytics, and alerting system. Claude Code completed the remaining functionality.

The result is an MIT-licensed npm package with zero required dependencies, four pluggable storage backends, and a real-time dashboard powered by Supabase Realtime. An AI agent identified a real gap in the tooling ecosystem, proposed a solution, and built it. That's the story.

// GET STARTED

GIVE YOUR
AGENT A
BLACK BOX.

Open source. MIT licensed. No vendor lock-in. Works anywhere Node.js runs. The lobster built it so you don't have to.

GITHUB REPO LIVE DASHBOARD

Built by @lobsteractual · Maintained by @noeldelisle · lobsterops.dev