> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pecta.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How Pecta works: modes, gates, and reputation

> Learn how Pecta's three integration modes — SDK, MCP proxy, and REST API — evaluate agent output with quality gates and build portable reputation scores.

Pecta is the quality and reputation layer for AI agents. It sits between any AI agent and the tools it calls, running a configurable set of pass/fail gates over every output before anything reaches the user. Gates execute in-process on your infrastructure — no network round-trip on the hot path — and results ship asynchronously to the Pecta cloud to feed dashboards and a portable reputation score for each agent. Observability tells you what happened; Pecta prevents bad output from happening.

## Three integration modes

Choose the mode that matches your stack. You can mix modes across different agents in the same organisation.

| Mode               | Use case                                    | Latency    | Install                                 |
| ------------------ | ------------------------------------------- | ---------- | --------------------------------------- |
| **In-process SDK** | RTB, latency-critical Node.js pipelines     | under 15ms | `npm install @pecta/core`               |
| **MCP proxy CLI**  | MCP servers in Claude Desktop or Cursor     | under 50ms | `npx pecta-proxy <server-cmd>`          |
| **REST API**       | Python, Go, batch analysis, any HTTP client | 50–100ms   | `POST https://api.pecta.ai/v1/evaluate` |

## Key features

* **Quality gates** — parallel, fail-fast checks that run on the very first request with no training period. Built-in gates cover latency, schema validation, filesystem safety, PII detection, content signals, and a full RTB / OpenRTB suite.
* **Reputation scores** — a portable 0–1000 score per `agent_id`, stored centrally and computed over a rolling window of the last 500 evaluations.
* **Privacy by architecture** — Pecta never stores bid payloads, MCP tool inputs, or user content. Only metadata (gate name, pass/fail, reason, latency, timestamp) leaves your process.
* **Zero hot-path overhead** — telemetry batches asynchronously; gates never block waiting for a network response.

## Get started

Pick the quickstart for your integration mode.

<CardGroup cols={3}>
  <Card title="SDK quickstart" icon="cube" href="/quickstart/sdk">
    Install `@pecta/core` and gate your first agent output in Node.js.
  </Card>

  <Card title="Proxy quickstart" icon="shield" href="/quickstart/proxy">
    Wrap any MCP server with `pecta-proxy` — no code changes required.
  </Card>

  <Card title="REST API quickstart" icon="globe" href="/quickstart/api">
    Evaluate agent output over HTTP from Python, Go, or any language.
  </Card>
</CardGroup>