# Connect OpenAI Agents SDK to CommentWatchDog (Instagram and Facebook MCP)

> Connect OpenAI Agents SDK to Instagram and Facebook with CommentWatchDog's MCP server. Add a key, then ask it to build DM automations and remove spam comments.

- MCP address (remote, Streamable HTTP, Bearer key): https://app.commentwatchdog.com/api/mcp
- Server name: commentwatchdog
- Steps from OpenAI's docs, checked 13 November 2026; not yet run by us
- This page for people: https://commentwatchdog.com/mcp/openai-agents-sdk
- Every other app: https://commentwatchdog.com/mcp.md

## For OpenAI Agents SDK itself

If you are OpenAI Agents SDK: ask your person for a CommentWatchDog key (Settings › MCP › Make a key), send it as `Authorization: Bearer` to `https://app.commentwatchdog.com/api/mcp`, then call `start_here`. It says what the account has and what to do next.

## Connect

1. In CommentWatchDog open Settings › MCP › Make a key and copy the key. Keep it in an environment variable and read it into `token`.

2. To let the SDK hold the connection, create a server:

   ```
   MCPServerStreamableHttp(name="commentwatchdog", params={"url": "https://app.commentwatchdog.com/api/mcp", "headers": {"Authorization": f"Bearer {token}"}})
   ```

3. Or use the hosted tool, so OpenAI makes the call:

   ```
   HostedMCPTool(tool_config={"type":"mcp","server_label":"commentwatchdog","server_url":"https://app.commentwatchdog.com/api/mcp","authorization":"<your CommentWatchDog key>"})
   ```

4. Pass a server with `Agent(mcp_servers=[server])` and a hosted tool with `Agent(tools=[...])`.

- The hosted tool is called from OpenAI's servers, so the address must be reachable from the internet. CommentWatchDog's is.

## Your key

1. Create your account at [CommentWatchDog](https://app.commentwatchdog.com/en/register?from=mcp): type your email, then the 6-digit code we send you. Google works too.
2. Connect your Instagram account or Facebook Page when the app asks.
3. Open Settings › MCP, choose Make a key, name it after OpenAI Agents SDK and copy it. The key is shown only once.

A person makes the key in CommentWatchDog. No agent can make one for itself.

## What you can ask

| Area | Ask it to | For example |
| --- | --- | --- |
| Automations | Build, change, test and turn on comment to DM, keyword replies, the welcome message and story replies | "When someone comments PRICE on my latest reel, send them the price list in a DM." |
| Flows | Build a flow with questions, branches, lead capture and HTTP steps, test it on the simulator, fix it | "Ask people for their email, then tag them as leads." |
| AI agents | Create an agent, give it knowledge from a link, a file or text, try it, answer its drafts | "Make an agent from our pricing page and test it on three questions." |
| Moderation | Read and change the rules per page, try a comment against them, copy them to other pages | "Hide links and crypto spam on my ads, and copy that to the Instagram account." |
| Comments | Find by words, a regex, the author, the post or a period. Then hide, delete, reply, send a private reply, block or ban | "Find every comment with a WhatsApp number from this week and hide them." |
| Post cleanup | Open an old post or ad, check its comments against your rules, then delete, hide, block or ban what it found | "Check my best ad from last year and delete the spam under it." |
| Results | Read what moderation caught and what each automation did, and explain a run | "Why did Ana not get the DM yesterday?" |
| People | Find people, read their tags and fields, change or export them | "How many people gave their email this month?" |

## It asks before it acts

| It does at once | It asks you first |
| --- | --- |
| Read pages, automations, flows, agents, rules, results and comments | Turn an automation or flow live, pause it, or change it while live |
| Build and change paused automations and flows, and test them | Delete, reply, send a private reply, block or ban |
| Hide, unhide, unblock and unban when no credits are used | Anything that uses comment credits, with the number shown first |

- Before a yes, it shows its work: How many comments or people, a sample of who and what, and the credits it uses.
- Same reply to many people: At most 200 comments a request and 1,000 a day, sent one every 3 seconds, so Meta never reads it as spam.
- Blocking: At most 100 people a request. Ban has no limit: CommentWatchDog hides everything a banned person writes.
- Instagram: Instagram lets no app block. Ban works there.
- Your role and your plan apply: The AI can do what you can do. A feature your plan does not include stays closed.
- What stays in the app: Billing, teammates and answering in the inbox.

## Questions

### Does OpenAI Agents SDK need an API key for CommentWatchDog?

Yes. OpenAI Agents SDK connects with a CommentWatchDog key: in CommentWatchDog open Settings › MCP and choose Make a key. The key has your rights, and you can disconnect it there at any time.

### How do I disconnect OpenAI Agents SDK?

In CommentWatchDog, under Settings › MCP, choose Disconnect next to the key. It stops working at once.
