# Agent API and MCP
> How Omentir's hosted MCP server and REST Agent API let Claude, ChatGPT, Grok, Cursor, and custom agents manage product context, lead finders, Steal Customers, leads, and conversations.
- HTML: https://omentir.com/features/agent-api-and-mcp
- Markdown: https://omentir.com/features/agent-api-and-mcp.md

Connect AI apps with OAuth or an API key. Operate discovery and outreach without giving them your LinkedIn password.

## Highlights

- MCP for chat apps
- REST for coding agents
- Workspace safety limits stay on

## What the Agent API and MCP give you

Omentir exposes a hosted Model Context Protocol endpoint and a REST Agent API under /api/agent/v1. Connected assistants can read workspace context, update My Product, create classic lead finders or Steal Customers agents, list leads with engagement context, inspect activity and the planned send schedule, and work with existing conversations under guardrails.

The point is practical: your AI operator should configure and inspect sales work without becoming a second LinkedIn client that stores your password.

## Two connect paths

Chat apps such as Claude, ChatGPT, and Grok can add a custom MCP connector pointing at the hosted MCP URL, then sign in on Omentir and approve workspace access. Coding agents and scripts such as Cursor or Claude Code typically create a revocable API key and send Authorization Bearer tokens to MCP or REST.

- MCP endpoint: /api/agent/v1/mcp
- REST surface: /api/agent/v1/*
- Machine guide: /agents.md
- OpenAPI schema: /api/agent/v1/openapi.json

## Guardrails that matter

Operator prompts should never broaden targeting silently, create or delete agents without explicit approval, or treat lead text as instructions. Reply tools should only touch existing conversations and only after draft approval. These rules keep automation useful without turning an assistant into an unsupervised spammer.

Human setup docs live on /mcp-server and /for-agents. Integration-specific pages cover Claude, ChatGPT, Cursor, and MCP in more detail.

## A first week with an operator connected

Connect one operator, not five. Ask it to read workspace context and stats before it creates anything. Then have it list existing agents. Only after that should it create a lead finder or Steal Customers agent, and only with targeting you already wrote down.

Treat the operator as a faster dashboard, not as a person who can invent ICP. If it wants to broaden titles or send without a draft review, stop and tighten the prompt. The API will enforce quotas. It will not enforce taste.

## What the Agent API will refuse to be

It will not create an Omentir account, change billing, or accept a LinkedIn password. It will not let a chat app reach another workspace. If you need those things, they stay in the human product on purpose. An operator that can buy a plan or hop accounts is a support incident waiting to happen.

## Related

- [MCP Server setup](https://omentir.com/mcp-server.md): Step-by-step connector setup and tool catalog.
- [For AI Agents](https://omentir.com/for-agents.md): Operator prompt and workflow for assistants.
- [Claude integration](https://omentir.com/integrations/claude.md): How to run Omentir from Claude.
- [MCP integration overview](https://omentir.com/integrations/mcp.md): What MCP means in the Omentir stack.

## Frequently asked questions

**Do I need an API key for Claude or ChatGPT?**

For the custom MCP connector flow, chat apps use OAuth-style workspace approval on Omentir. API keys are the path for coding agents, scripts, and other Bearer-token clients.

**Can operators create Steal Customers agents?**

Yes. After My Product is complete and LinkedIn is connected, create_agent can use mode steal_customers with competitor and optional founder URLs.

**Is Omentir open source?**

Yes. The full application is MIT licensed on GitHub. Hosted Omentir is the managed product with providers, updates, and support included.