# Nibble

Nibble is an AI-powered production incident investigation prototype built with the Model Context Protocol (MCP) and the OpenAI Agents SDK.

## What it does

Nibble exposes production observability data through an MCP server, including:

- Service health
- Service logs
- Recent deployments
- Deployment details

An OpenAI agent can use these tools to autonomously investigate production incidents and identify likely root causes.

## Example

Prompt:

> Customers are reporting that checkout is failing. Investigate the incident and determine the likely root cause. Then indicate which deployment(s) caused it and who is responsible for the deployment(s).

Agent run output:

![OpenAI agent incident investigation output](docs/agent-run-output.svg)

The agent can autonomously:

1. Check service health
2. Identify degraded services
3. Inspect relevant logs
4. Correlate failures with recent deployments
5. Inspect deployment changes
6. Determine a likely root cause

## Architecture

```text
OpenAI Agent
    |
    | MCP / stdio
    v
Nibble MCP Server
    |
    +-- get service health
    +-- get service logs
    +-- get deployments
    +-- get deployment details
    |
    v
Mock production data
```

## Running locally

Install dependencies:

```bash
npm install
```

Create a `.env` file:

```env
OPENAI_API_KEY=your_api_key
```

Run:

```bash
npx tsx src/agents.ts
```