Live meeting notes agent that wakes on a spoken name

Live meeting notes agent that wakes on a spoken name

An n8n workflow in two parts. The first sends a Recall.ai bot into a meeting and saves a row for it in Supabase. The second
receives each spoken line from Recall, appends it to that row, and watches for a name. When someone says the name, an
OpenAI agent reads the latest lines and saves a note if one was asked for.

What is inside

• workflow.json: the workflow as n8n exports it, 11 nodes and 7 sticky notes that label the two parts and mark what to
  replace. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The step that asked an OpenAI assistant is rebuilt as an AI Agent with a chat model, a memory and the note tool. The
  original used the Assistants API, which OpenAI shut down on 26 August 2026. The step that created the assistant's thread
  is gone, and the memory is kept per Recall bot instead.
• The assistant's instructions lived in the author's OpenAI account, not in the file, so the agent has a short system
  message written by Mindket. Read it and adjust it.
• The bot request uses the current Recall.ai fields (recording_config with a webhook under realtime_endpoints), and the
  nodes read the current webhook payload. The original used older field names.
• Two expressions pointed at nodes that do not exist, and the query returned only a thread id while the next step read the
  dialog from it. Both are repaired.
• The SQL no longer has text from the meeting pasted into it. Each value is passed as a bound value, so a speaker's display
  name cannot change the query.
• The first part had no trigger, so a manual trigger starts it.
• The meeting link, the webhook address of the author's own n8n and the webhook path became placeholders or a fresh path.
  The author's talk, video link and set-up essay are left out.

Before you start

• An n8n instance that Recall.ai can reach from the internet, such as n8n Cloud or a server with a public https address.
• A Recall.ai account and its API key.
• A Supabase project, with its URL, its service role key and the connection details of its database.
• An OpenAI API key.

Set it up

1. In Supabase, open the SQL editor and create the table the workflow uses:

   create table public.data (
     id uuid primary key default gen_random_uuid(),
     date_created timestamptz not null default now(),
     input jsonb,
     output jsonb
   );

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json, then save.
2. Add a Supabase credential with the project URL and the service role key. Select it in Create data record.
3. Add a Postgres credential with the connection details of the same Supabase database. Select it in
   Insert Transcription Part and Create Note.
4. Add an OpenAI credential and select it in OpenAI Chat Model.
5. Add a Header Auth credential. Name the header Authorization and give your Recall.ai API key as its value. Select it in
   Create Recall bot. The host in that node, us-west-2.recall.ai, must be the region of your Recall.ai account, so
   change it if your account is elsewhere.
6. Open Transcript webhook and copy its production URL. In Create Recall bot, replace <YOUR_N8N_WEBHOOK_URL> in the
   body with it. Then activate the workflow, because Recall can only reach a production webhook.
7. Start a meeting, and in Meeting URL replace <YOUR_MEETING_URL> with its link. Run the workflow from
   Run to send the bot. The bot joins, and the host may need to let it in.

How a run goes

1. Run to send the bot starts the first part. Meeting URL holds the link of the meeting.
2. Create Recall bot asks Recall.ai to send a bot. The body turns on English live transcription at low latency, sends each
   finished line to your webhook as a transcript.data event, and sets when the bot leaves on its own.
3. Create data record saves a row in public.data with the bot's id and the link in input, and an empty dialog list
   in output.
4. As people speak, Recall posts each line to Transcript webhook.
5. Insert Transcription Part finds the row by the bot's id and appends the line with its order, speaker, language and
   time. It returns the whole list.
6. If Jimmy is said checks whether the line contains the name. Lines without it stop here.
7. AI Agent receives the latest ten lines, oldest first, as speaker: words. It calls Create Note once if it finds a
   request to note something or a decision, a task or a key point. Create Note appends the text to the notes list of the
   same row.
8. Window Buffer Memory keeps the last ten turns of the agent for each bot, so it knows what it saved earlier.

Change the look

• The name it listens for: the right value in If Jimmy is said, and the name written in the system message of AI Agent.
  The check is case sensitive.
• What it notes and how it answers: the system message in AI Agent.
• How many lines it reads: slice(-10) in the prompt of AI Agent.
• The model: OpenAI Chat Model, set to gpt-5.6-terra.
• The language: language_code in the body of Create Recall bot. The low latency mode there supports English only.
• When the bot leaves: the automatic_leave block in the same body.

Good to know

• The facts about Recall.ai and OpenAI come from pages read on 12 November 2026.
• A bot joins and transcribes the meeting. Tell the people in it, and check the rules for recording where you and they are.
• Recall's documentation sets live transcription through recording_config and says the low latency English mode delivers
  a line about one to three seconds after it is spoken.
• OpenAI's documentation lists function calling as supported for gpt-5.6-terra, which the agent needs for its note tool.
  If you prefer another model, choose one that supports it.
• The webhook has no authentication. Anyone who learns its address can add lines to your database. Recall's documentation
  asks for its signature to be checked before a payload is used, and this workflow does not do that.
• Each line is stored in full in Supabase, and the agent sees only the latest ten lines when the name is said.
• Window Buffer Memory is temporary: n8n's source says it lives inside the instance and does not work in queue mode.
• Billing: Recall.ai bills the bot under your plan, OpenAI bills each time the name is said, and Supabase hosts the table.
• Spoken words and display names go to Recall.ai, to your Supabase project and, for the latest ten lines, to OpenAI.