Chat agent that researches YouTube channels and comments

Chat agent that researches YouTube channels and comments

An n8n chat agent that answers questions about YouTube. You type a request in the chat, and the agent finds a channel by its handle, lists its videos, reads a video's description and comments, fetches a transcript and judges a thumbnail. It keeps the chat history in Postgres. Use it to see what an audience asks for, how top videos are packaged and what to make next.

What is inside

• workflow.json: the workflow as n8n exports it, 28 nodes. 22 of them run and 6 are sticky notes. It holds no credential.
• README.md: these notes.

The file holds two scenarios. The first is the chat agent. The second is a set of tools in the same workflow, which the agent calls through seven sub-workflow tool nodes.

What we changed from the original copy

• AI Agent was on version 1.6 in OpenAI Functions mode. n8n's source, read on 12 November 2026, marks that mode as deprecated and removed in n8n v3. It is an AI Agent on version 2.2 now, with the same prompt, memory and seven tools.
• The seven tool nodes pointed at a workflow in the file author's n8n by its ID. The tools now call this workflow itself, by {{ $workflow.id }}, because the second scenario is in the file.
• OpenAI Chat Model named no model, so n8n used the node's default, gpt-5-mini. OpenAI's deprecation page, read on 12 November 2026, shuts it down on 11 December 2026 and names gpt-5.6-terra. The node names gpt-5.6-terra now.
• The OpenAI node that judges thumbnails used gpt-4o. OpenAI's page shuts that model down on 14 January 2027 and names gpt-6-luna, which takes image input. The node uses it and allows 1500 tokens, up from the default of 300, because that model reasons before it answers.
• Three sticky notes carried a banner picture, a profile link, a community link, a referral link and a video thumbnail. Two are rewritten in Mindket's words, and the one with the video thumbnail is gone.
• The credential links and the saved test data in Execute Workflow Trigger are gone, and the file's name is ours.
• Nothing else. The system message, the tool descriptions, the router and the requests are as the file had them.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google Cloud project with the YouTube Data API turned on, and an API key for it.
• An Apify account and its API token, for transcripts.
• An OpenAI API key.
• A Postgres database for the chat memory.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Create a Query Auth credential for YouTube. Name the query parameter key and give it your Google API key. Select it in Get Channel Details, Get Video Description, Get Comments, Run Query and Get Videos by Channel.
3. Create a second Query Auth credential for Apify. Name the query parameter token and give it your Apify token. Select it in Get Video Transcription.
4. Add an OpenAI credential and select it in OpenAI Chat Model and OpenAI.
5. Add a Postgres credential and select it in Postgres Chat Memory.
6. Save the workflow. The tools call the saved version of it by its ID, so save again after any change to the second scenario.
7. Open the chat and ask for something, such as "Find the channel with the handle example_handle and summarize the comments on its latest video."

What happens in a run

1. When chat message received takes the message, and AI Agent reads it as chatInput. Postgres Chat Memory uses the chat's sessionId, so each chat keeps its own history.
2. The system message tells the agent to plan the tools in a sensible order and to ask when a request looks like a mistake. It also tells the agent to check a video's length when it needs to leave out shorts.
3. The agent calls a tool by name: get_channel_details, get_video_description, get_list_of_videos, get_list_of_comments, search, analyze_thumbnail or video_transcription. Each tool sends a command and its input to the second scenario.
4. Execute Workflow Trigger takes the call, and Switch sends it to the matching request. The channel, video, search and comments requests go to the YouTube Data API. Edit Fields turns the comments into lines of author: comment. The transcript request runs an Apify actor and waits for its result. OpenAI describes a thumbnail from its address and a prompt.
5. Response returns the result to the agent, which answers in the chat.

What to change

• How the agent behaves: the system message in AI Agent.
• The model: OpenAI Chat Model for the agent and OpenAI for thumbnails.
• How many comments come back: maxResults in Get Comments, which is set to the most the API gives in one call.
• A new tool: add a Switch output with its own request, then a tool node that sets its command.
• The transcript source: the Apify request in Get Video Transcription.
• How much the chat remembers: Postgres Chat Memory.

Good to know

• A new YouTube Data API project allows 100 search calls a day by default, and 10,000 units for all other calls combined, according to Google's quota page. Run Query and Get Videos by Channel are search calls. A long session can reach the limit.
• Get Comments asks for one page of up to 100 comment threads and does not go further.
• Transcripts come from an Apify actor that someone else publishes. It was public and updated in July 2026 when we read it. A run can take a minute or more, and Apify charges for it. Check the actor's page before you rely on it.
• A transcript is long. It goes into the agent's context and into the chat history, so a few of them add up in cost.
• Comments carry viewers' public display names, and the agent passes them to OpenAI with the rest.
• Data goes to Google for the lookups, Apify for transcripts, OpenAI for the model and the thumbnails, and your Postgres for the history.
• The chat is private to n8n unless you make it public in the trigger. A public chat lets anyone spend your keys.