The AI Assistant That Knows Every Endpoint
A guided tour of the fastest way into the LinkedIn API: open the assistant in the docs, describe the play in plain English, and let it name the endpoints. Video and screenshots at every step.
Open the AI Assistant in the gtm-api docs and describe the play you want in plain English. It is wired to the live API and to the changelog, so it answers with the endpoints that exist today and the order to call them in. That is the shortest route into the LinkedIn API, and it costs nothing on the free plan.
The problem with a large API is rarely the API. It is the first ten minutes. You know the outcome you want, you open the reference, and you are looking at a wall of endpoints with no idea which three of them are yours. Below is the whole two-minute walkthrough, then the same thing broken into steps you can click through.
On this page
Where to start when you are stuck
Start in the docs, with the assistant, before you start reading endpoints. Type the outcome you are after and let it hand you the call sequence.
In the opening of the video above, Peter Kaliuzhny, partner at gtm-api, names the failure mode this page exists to fix: the API gives you so many capabilities that new users stall before the first call. Nothing is broken and nothing needs support. You simply cannot tell which endpoints are yours.
“This AI assistant can help you to understand what you can actually do right now. It’s connected to the current API. It’s connected to the changelog, so it knows everything that we’re shipping.”
That last part is the reason to prefer the assistant over any static list, including this one. The surface moves every week, so a page that enumerates endpoints is accurate on its publication date and slowly stops being true. The assistant reads the contract at the moment you ask.
What the AI Assistant actually is
It is a question-answering layer over the gtm-api documentation set, with a live connection to the API contract and the release history. You get plain-English answers about which endpoint does what and what changed recently, without holding the whole reference in your head.
▶Watch: what the assistant is wired to, from 0:19
The recommendation to start in the API docs, and the two connections that make the assistant worth opening: the current API and the changelog. This is where the quotation above comes from.
The assistant is one of four ways into the same information, and they suit different moments:
| Surface | Reach for it when | Reflects this week’s ships |
|---|---|---|
| AI Assistant | You know the outcome and not the endpoint. It answers in plain English and sequences multi-step plays | Yes, reads the API and the changelog |
| API reference | You know the endpoint and want exact parameters, response envelope and error codes | Yes, generated from the live contract |
| MCP tool discovery | You want the agent to pick the tool itself at runtime, inside Claude or Cursor | Yes, the server lists its own tools |
| Changelog | Something behaved differently than it did last month and you want to know why | Yes, that is what it is for |
Anything the docs cannot see, because it depends on the state of your own account, goes to human support instead. The assistant is a knowledge surface. It does not read your workspace.
The walkthrough, step by step
The example from the video is a keyword-to-lead-list play: find posts on a keyword, pull the people who commented, then enrich them. Here is what that looks like on screen.
Step 1: Ask for the outcome, in whatever English you have
How do I find and scrape posts for a keyword and then enrich the authors? How do I find and scrape posts for X keyword, then scrape the comments on each post?
The second one is the version typed in the video, so it is the one the rest of this walkthrough follows.
▶Watch: the question being typed and sent, from 0:41
The panel switches from empty to working: it echoes the question at the top, runs its own searches across the docs, and starts assembling the answer underneath.
Step 2: Read the chain it gives back
scrape_linkedin_search_posts) with filters.keywords, which returns posts carrying a post_ln_id. 2. Scrape commenters, feeding that post ID into Scrape post comments (scrape_linkedin_get_post_comments), which returns each comment plus the commenter’s member ID, name and headline in one call. Underneath, the branch you did not ask for: if you want reactors or resharers as well, it names Scrape post reactors and Scrape post resharers. Every endpoint name in the answer is a link into the reference.The correction is worth dwelling on. The assistant read an ambiguity in the question, said which reading it picked, and offered to redo it. That behaviour is the difference between a docs search box and something that can hold a play in its head.
Step 3: Click an endpoint name to land on its reference page
linkedin_scraping). The page tells you the things the summary could not: that keywords are required, that you address it either by a filters object or by a pasted search URL and never both, and that it is cursor paginated, so you feed paging.next_cursor back verbatim. The assistant gave you the route. This page gives you the parameters.Step 4: Keep pulling on the thread
▶Watch: one question becomes a whole play, from 1:25
Scrape posts, this endpoint. Then scrape comments, this endpoint. Here is how you do reactors. Then enrichment, if you want it. The useful habit is asking follow-ups in the same thread, because the answers stack into something you can schedule.
Good follow-ups are the boring ones. Which call returns the profile URL. What this costs against your daily allowance. What changed in these endpoints recently. Daily pacing is enforced server-side per account and the caps differ by action type, which is covered on the LinkedIn limits page.
Why your agent already understands gtm-api
The product is MCP-first, so Claude and other agent runtimes discover its tools without being taught. Model Context Protocol is the open standard Anthropic published in late 2024 for connecting models to external tools. Point an MCP client at the server, and the agent reads the tool list and picks its own calls.
▶Watch: why Claude and other engines already know gtm-api, from 1:03
Claude and the other LLM engines know gtm-api natively, because the API was built for them. This stretch also gives the two moments the assistant still earns its place: when you want to understand the system yourself before pointing an agent at a live LinkedIn account, and when an agent stalls on an ambiguous request and you need the correct sequence to hand it.
The surface is large and it keeps growing. On 26 August 2026 the live MCP server listed 19 toolsets, covering messaging, network, content, scraping, saved scraping jobs, enrichment, account health, antidetect browsers, webhooks, bulk actions, identity and billing. Any count quoted in a blog post, this one included, has a shelf life, which is the argument for putting the question to the server or the assistant. Full setup for Claude and Cursor is on the LinkedIn MCP server page.
What runs on the free plan
There is a free forever plan for individual users who want to run their own LinkedIn account with AI agents, so the assistant, the docs and a connected account are all reachable without a card. Paid plans start from $10 per connected account per month, and the current grid sits on the pricing page.
The free tier connects a live LinkedIn account, with lower daily ceilings per action type. Scraping, enrichment, messaging and network actions each have their own daily allowance, and they are enforced before the call runs, so an agent cannot burn through a day’s budget in one loop.
- One connected LinkedIn account. It runs in its own cloud browser session on a dedicated residential IP, the same isolation model the paid plans use.
- The full typed surface. The endpoints are the same ones paid accounts call. What differs is the daily volume.
- MCP and REST together. Every action is reachable from an agent through MCP and from your own backend over REST, because both are generated from one contract.
- Server-side pacing. Limits are checked before each action instead of after, which is how gtm-api reports 20,000+ LinkedIn accounts at under a 1% ban rate.
For teams weighing this against the wider category, the honest framing is that free LinkedIn automation usually means a browser extension running on your own machine, and the trade there is your session and your IP. Cloud options with a free entry point are rarer. The comparison of approaches, including the official LinkedIn API and the third-party providers, sits on the LinkedIn API guide.
Frequently Asked Questions
What is the gtm-api AI Assistant?
It is a question-answering assistant built into the gtm-api documentation at docs.gtm-api.com. You describe what you want to do on LinkedIn in plain English and it replies with the endpoints that do it, in the order to call them. It is connected to the live API contract and to the changelog, so its answers reflect what is shipped today.
Where do I find the AI Assistant in the docs?
There is an “Ask Assistant” button in the top bar at docs.gtm-api.com, between the search box and the Pricing link. It opens a panel down the right-hand side of the page, next to the quickstart, the authentication guide, the knowledge base, the API reference and the changelog. You do not need a paid plan to use it.
Does the assistant know about endpoints shipped this week?
Yes. It reads the current API and the changelog, which is the difference between asking it and reading a guide that was accurate on its publication date. “What changed in the scraping endpoints this month?” is a reasonable question to put to it.
Is the AI Assistant the same thing as the gtm-api MCP server?
No, they are two different surfaces. The assistant answers questions about the API in the docs, for a human reading. The MCP server exposes the API as tools an AI agent calls at runtime, so Claude, Cursor or any MCP client can run the actions themselves. A common pattern is to work the play out with the assistant, then let the agent execute it over MCP.
Do I still need the assistant if my agent already speaks MCP?
Often not, because an MCP client discovers the tools on its own. It earns its place in two situations: when you want to understand the system yourself before pointing an agent at a live LinkedIn account, and when an agent picks the wrong tool or stalls on an ambiguous request and you need the correct sequence to hand it.
What can I do on the gtm-api free plan?
The free forever plan for individual users covers one connected LinkedIn account with the same typed endpoints paid accounts call, over both MCP and REST. The difference is the daily allowance per action type, which is lower and enforced server-side before each call. Paid plans start from $10 per connected account per month.
What should I ask the AI Assistant first?
Describe an outcome and ask for the call sequence, for example “find posts about a keyword, pull everyone who commented, then enrich them”. Naming the outcome gets you a chain of endpoints, where naming a category gets you a list to read. Good follow-ups in the same thread are which call returns the field you need, what it costs against your daily allowance, and what changed in that part of the API recently.
Sources & Further Reading
- Documentation and AI Assistant, docs.gtm-api.com (quickstart, authentication, knowledge base, API reference, MCP server and changelog, sections verified 26 August 2026)
- Walkthrough video, “How to Start with gtm-api: the AI Assistant That Knows Every Endpoint”, gtm-api on YouTube, 25 August 2026. Every clip and screenshot on this page is cut from it, and the quotation above is transcribed from it
- Model Context Protocol, modelcontextprotocol.io, the open standard published by Anthropic in late 2024
- Toolset count of 19 read from the live gtm-api MCP server on 26 August 2026, self-reported. Product and plans: gtm-api.com
- Related: LinkedIn MCP server · The LinkedIn API · LinkedIn limits
Ask it your first question.
Create a free account, connect one LinkedIn account, then open the docs and describe the play you want. The assistant names the endpoints and your agent runs them.
Last updated: August 2026 · Endpoint surface and plan details verified against the live gtm-api at time of publication