Gemini Notebook (formerly NotebookLM) Workflows for Agents

“Create notebook → generate podcast → send to phone.” A strategy consultant on the Kortex MCP waitlist wrote that as one wish, to be dictated to Claude or Codex. Add the step before it, finding the papers, and it’s four jobs. Gemini Notebook (formerly NotebookLM) makes you do each one by hand, then all four again next week. This guide shows how to do it once with your AI assistant, save it as a named workflow, and rerun it in one sentence or on a clock.

Key Takeaways

• 62 of 776 Kortex MCP waitlist respondents asked for multi-step agent workflows, and 23 of those 62 also wanted them on a schedule or trigger (Kortex MCP waitlist survey, Sep-Oct 2026).
• create_workflow saves the tool calls that actually ran, with inputs for what changes per run. run_workflow replays them.
• Every step is checked against your key’s profile first.
• A schedule rule can run a saved workflow, so any job you can say becomes a job on a clock.

What is a Gemini Notebook AI workflow?

A workflow is a named, saved sequence of Kortex MCP tool calls that your AI assistant can rerun by name. You don’t write code or draw a flowchart. Instead, you describe the job once, your assistant does it, and then you keep it. The Workflows & Schedules docs put it plainly: the saved thing is the sequence of tool calls that actually ran, not your sentence. As a result, “run my episode workflow” repeats exactly what worked last time, with no improvising. A workflow holds up to 20 steps and up to 10 inputs, and it can call any of the 150 tools Kortex MCP exposes, passing results from one step to the next. Is that the same as an automation rule? No. The rules in our Kortex automation workflows guide are trigger-and-action rules, such as “when a notebook gets this tag, make a Briefing Doc.”

From the waitlist: We built Kortex MCP, and we analyzed the 776 answers to its waitlist survey (Kortex MCP waitlist survey, 776 respondents, Sep-Oct 2026). We found that 62 people asked for multi-step agent workflows, about 8% of everyone who signed up. They also wanted more than most: 3.4 different jobs each on average, against 2.0 for everyone else who named a clear job. In addition, 23 of those 62 wanted the chain to run on a schedule or trigger, and 27 of them (44%) already copy and paste between tools by hand. For example, one wants to “have my chatGPT orchestrator control Notebook LM to make automate my workflow end-to -end.” A workflow is how a pile of jobs becomes one request, and the pipeline in this guide comes straight from one of those 62 requests.

Here’s where workflows rank among the asks:

Why can’t Gemini Notebook act like an agent on its own?

Gemini Notebook has no macros or saved sequences, and Google publishes no public consumer API for it. Therefore, nothing outside the app can drive it on your behalf. Is it an agent at all? Partly: Google says Deep Research makes a research plan and browses websites for you. However, an agent in the research sense does more. The ReAct paper, for instance, describes models that interleave reasoning with actions against outside tools, such as a Wikipedia API (Yao et al., arXiv, 2022). Gemini Notebook won’t chain its own steps that way. Meanwhile, an Audio Overview alone can take 15 minutes or more (Kortex MCP tool docs, 2026). So the four-step wish becomes a babysitting job. You search, import, wait, click Generate, wait again, download, then get the file onto your phone. And next week? You do the whole dance again.

The waitlist shows what that manual work costs. Specifically, 244 of the 776 respondents (31%) described a manual copy-paste workaround they run today (Kortex MCP waitlist survey, Sep-Oct 2026). One AI strategy consultant, for instance, researches in Perplexity, pastes sources into one notebook per client, then copies the outputs back into Claude by hand. Kortex MCP is the bridge that fills that gap from the outside. Your AI client calls https://mcp.kortex-notebooklm.com/mcp, and the call is relayed to your own Kortex Chrome extension, which talks to Gemini Notebook with your signed-in Google session. Kortex never stores your Google password or cookies. A workflow is that same path, saved and replayed. In other words, carrying things between tabs stops being your job.

What do you need before you start?

You need Kortex Pro, a Kortex MCP invite, an AI client and Chrome. Pro costs $9.99 a month or $4.92 a month billed yearly (Kortex pricing, 2026). Once your invite arrives, however, setup takes a few minutes.

1. Install Kortex from the Chrome Web Store and sign in.
2. Connect your AI client. Paste the server URL and approve the consent page, or use an API key. In Claude Code, for instance: claude mcp add --transport http kortex https://mcp.kortex-notebooklm.com/mcp. Our Claude setup guide covers the rest.
3. Pick a key profile that can write. New API keys default to Read-only, so a workflow that adds sources or makes audio needs a write-capable profile.
4. Keep Chrome open with the extension signed in.

How do you build a Gemini Notebook research workflow?

Build it by asking, not by configuring. In practice, you do the job once in conversation, check the result, then save the part you’ll repeat. Google’s Fast Research “rapidly scans for information, letting you immediately review and import sources” (Google, 2025), and Kortex MCP can drive it for you.

Step 1: Find the literature

research_sources runs Gemini Notebook’s own source research in three calls. start returns a task id, status returns what it found, and import adds all of it or the URLs you pick. Thus research can outlive a single MCP response. Fast mode is the working path today, and it can search the web or your Google Drive.

Create a notebook called "Sleep and memory 2026". Run fast web
research for recent papers on sleep and memory consolidation, show
me what it found, and import only the ones from journals or preprint
servers. Tell me when every source has finished processing.

Already have a reading list? Alternatively, import_pipeline takes a list of links, CSV text, an RSS or Atom feed, or a small site crawl. It handles up to 50 links per run and spends no Kortex import credits.

Here are 14 paper links. Preview them first without importing,
then add the ones that resolve to my "Sleep and memory 2026"
notebook. Tell me which ones failed.

This is where the bridge matters: Kortex MCP carries whatever your AI client can reach into a notebook. Similarly, if your client can read your email, see our Gmail to Gemini Notebook guide.

Step 2: Make the episode and send it to your phone

generate_audio_overview starts an Audio Overview in one of four formats: deep dive, brief, critique or debate. Meanwhile, it returns an artifact_id immediately while Gemini Notebook does the work. Once it’s done, add_podcast_episode adds it to a Kortex podcast. A Kortex podcast is a private RSS feed of your Audio Overviews, and get_podcast_feed_url gives you its address.

Generate a brief Audio Overview of "Sleep and memory 2026". When it
finishes, add it to my "Papers" podcast (create the podcast if I
don't have one yet) and give me the feed URL.

In Apple Podcasts, go to Library, tap More and choose “Follow a Show by URL” (Apple Podcasts for Creators). Likewise, most podcast apps take a feed URL. You do this once; after that, every episode just shows up.

One catch: the audio file for a new episode is prepared the next time you open the Kortex dashboard. In addition, treat the feed URL as private, since anyone who has it can listen. For format choices, see making Studio outputs from chat.

Step 3: Save it with create_workflow

create_workflow saves the steps that worked. Your assistant hard-codes ids that never change, like your Papers podcast, and anything that varies becomes an input. Nothing runs at save time. Notably, saving is rejected if a step names an unknown tool or a placeholder that can’t resolve.

Save the audio part of that as a workflow called "Episode to my
phone". Make the notebook an input, and the format an optional
input that defaults to brief. Hard-code my Papers podcast.

That’s the whole build. Consequently, “run Episode to my phone on my Transformers notebook” is now one sentence, and a new episode joins your Papers podcast once it’s ready.

What does a saved workflow actually contain?

A saved workflow is plain data: a name, an optional description, a list of inputs and a list of steps. Specifically, each step is {tool, args}, the tool’s normal arguments with placeholders where values vary. The limit is 20 steps and 10 inputs (Kortex MCP tool docs, 2026). Here’s roughly what the request above stores:

{
  "name": "Episode to my phone",
  "inputs": [
    { "name": "notebook", "label": "Which notebook?", "type": "notebook" },
    { "name": "format", "required": false, "default": "brief" }
  ],
  "steps": [
    { "tool": "generate_audio_overview",
      "args": { "notebook_id": "{{notebook}}", "format": "{{format}}" } },
    { "tool": "add_podcast_episode",
      "args": { "podcast_id": "PAPERS_PODCAST_ID",
                "notebook_id": "{{notebook}}",
                "artifact_id": "{{steps.0.artifact_id}}" } }
  ]
}

Three kinds of placeholder exist. {{notebook}} is an input you supply at run time. {{steps.0.artifact_id}} reads a field from an earlier step’s result, counting from zero. Finally, {{date}}, {{time}} and {{datetime}} stamp the moment the run starts, which is handy for notebook titles. What about the waiting? That’s handled for you. When generate_audio_overview or generate_artifact has steps after it, the run parks itself until the artifact is ready, then finishes the rest on its own. That’s why step 2 can safely use the artifact id. In contrast, a step saved with continueOnError may fail without stopping the run. You never have to read this JSON, though you can: the Kortex dashboard lists every saved workflow under Workflows, with its inputs, steps and runs.

How do you run, check and change a saved workflow?

Say its name to run it, ask how it went, and describe changes in plain words. Three tools do that work: run_workflow, get_workflow_run and update_workflow.

Run it by name

run_workflow runs the steps in order and waits up to 40 seconds by default (60 at most). Short runs come back finished. In contrast, longer ones return a run_id with status “running”, or “waiting” while an Audio Overview that can take 15 minutes or more is cooking (Kortex MCP tool docs, 2026).

Run "Episode to my phone" on my "Transformers reading group"
notebook. Check on it until the episode is in my Papers podcast,
then tell me how each step went.

Your assistant then polls get_workflow_run, which reports the overall status and each step’s result or error. Don’t ask it to call run_workflow again while one is going: that starts a second run, not a status check. Forgot what you saved? “Which workflows do I have?” calls list_workflows.

Which steps will your key allow?

Every step is checked against your key’s profile first. So “a key that cannot delete cannot run a workflow that deletes” (Kortex MCP tool docs, 2026). If any step is out of bounds, the whole run is refused before step one. Accordingly, a profile problem never leaves you with half a job done.

Key profile	What a workflow on this key can do
Read-only (default)	Read and ask only. Any step that writes blocks the run.
Minimal	Reads plus additive writes like adding sources and creating notebooks.
Standard	Everything except destructive tools. The usual choice.
Full	Every tool, including deletes.

Why does that matter? The tool your client calls is just run_workflow, which sounds harmless. In fact, without per-step checks, a read-only key could run a saved workflow that deletes notebooks. With them, the profile you chose for a client stays true no matter what that client saves.

Change it with update_workflow

Ask for the change in plain words. update_workflow edits a saved workflow’s name, description, inputs or steps, and it’s checked like create_workflow. Fields you leave out stay as they were. However, passing steps or inputs replaces that whole list, so your assistant must send the full list back with your edit included.

Update "Episode to my phone": make debate the default format, and
add a last step that tags the notebook "listened" once the episode
is in the podcast.

A good habit, therefore: compare the step list before and after.

How do you put a saved workflow on a schedule?

Use create_automation with a schedule trigger and a run_workflow action. Kortex schedules run hourly, daily or weekly on your computer’s clock while Chrome is open. The extension checks about every 15 minutes (Kortex MCP docs, 2026). The run_workflow action is the catch-all: anything you can do by hand becomes schedulable, including adding sources, which no other schedule action does.

Here’s the whole wish on a clock, literature included. A new notebook each week keeps you under the free plan’s 50-source cap (Gemini Notebook Help, 2026), and arXiv category feeds update daily (arXiv):

Save a workflow called "Monday papers": create a notebook named
"cs.CL papers" plus today's date, import the 15 newest items from
https://rss.arxiv.org/rss/cs.CL into it, generate a brief Audio
Overview, and add it to my Papers podcast. Run it once now. If it
works, schedule it for every Monday at 8.

Test it with a manual run first: run_automation fires the rule on demand, and list_automation_runs is the log.

Give new sources a head start. A run waits for an Audio Overview, but not for freshly imported sources to finish processing. If an episode skips papers, split the job in two. For example, one schedule imports into a standing notebook at 7, and a second runs “Episode to my phone” at 8. That said, a standing notebook keeps growing, so prune it now and then.

Two rules keep a schedule safe. First, a scheduled run has nobody to ask, so every required input needs a value in the rule or a default in the workflow. Second, a scheduled run isn’t made by an MCP key, so its steps are not narrowed by a key profile. As a result, it runs whatever the workflow was saved to run. The docs’ advice is right: keep anything irreversible as a workflow you run by hand. For daily digests and morning summaries, see the daily briefing guide.

What can’t a workflow do?

A workflow can only call Kortex tools, so it can’t read Gmail, Notion or other outside apps. Your AI client can, during a conversation, with its own connectors. Schedules have no client attached; hence they can’t read outside apps either (Kortex MCP docs, 2026). That split matters. 81% of waitlisters said their assistant is connected to Gmail, and 81% to Google Drive (Kortex MCP waitlist survey, Sep-Oct 2026). Use those connections in conversation: your assistant fetches the email or file, and Kortex MCP files it in a notebook. From the notebook onward, a saved workflow can take over.

A few other limits are worth knowing before you build:

• Google’s limits still apply. Since September 2, 2026, Gemini Notebook usage refreshes every five hours until a weekly cap, and a daily audio workflow spends it like you would. Our usage limits explainer has the numbers.
• Chrome has to be running. Minimised is fine, but a run missed with the laptop closed waits for the next one.
• Imports cap at 50 links per import_pipeline call, and research uses fast mode.

For a wider tour of what Kortex MCP covers, start with our Gemini Notebook MCP guide.

Get your first workflow running

The quickest win is the job you already repeat. Do it once with your assistant, check it, then say “save that as a workflow.” A good first one is “Episode to my phone”: a job you used to babysit becomes one sentence, and with a Monday rule, not even that.

Frequently asked questions

Is NotebookLM an AI agent?

Partly. Google says its Deep Research mode creates a research plan and browses hundreds of websites on your behalf. But Gemini Notebook (formerly NotebookLM) has no macros, saved sequences or public consumer API, so it won’t chain research, audio and a podcast feed on its own. Claude or Codex can, through Kortex MCP, a private beta for invited Kortex Pro members.

How is a Kortex workflow different from an automation rule?

A workflow is a saved list of Kortex MCP tool calls that you build by talking to your AI and rerun by name. An automation rule is a trigger plus actions, such as a tag added or a weekly clock. The two meet when a schedule rule’s action is run_workflow, which puts a saved workflow on a clock.

Can a saved workflow read my Gmail or Notion?

No. Every step is a Kortex tool, and Kortex doesn’t read outside apps. Your AI client can, with its own connectors, during a conversation. It fetches the material and Kortex MCP carries it into a notebook. A scheduled run has no AI client attached, so it only reaches what Kortex tools reach, like your notebooks and public links and feeds.

Do I need Kortex Pro to use workflows?

Yes. Workflows are part of Kortex MCP, which is a Kortex Pro feature at $9.99 a month or $4.92 a month billed yearly. Kortex MCP is also a private beta right now: only invited members can connect, and invites go out from the Kortex MCP waitlist.

What happens when one step of a workflow fails?

The run stops at that step, unless the step was saved with continueOnError, which lets it fail and carry on. Ask your assistant to check the run with get_workflow_run. It reports the overall status, every step’s tool, whether each step succeeded and the error for any step that didn’t.

Will a scheduled workflow run while my laptop is closed?

No. Schedules fire from the Kortex extension inside Chrome, which checks for due rules about every 15 minutes. Chrome can be minimised with no notebook tab open, but a closed laptop runs nothing. A daily or weekly rule only fires during its chosen hour, so a run missed with Chrome closed waits for the next one.

---

Editorial note: fact-checked against the official help pages and docs linked above on October 2, 2026. Spotted something out of date? Contact us and we will fix it, or read more about Kortex.