Connect LinkedIn to ChatGPT or Claude for prospecting

The protocol, and which clients actually showed up

The connection runs over the Model Context Protocol. One door, one key, and any client that speaks the protocol can hold a conversation with it. The key is a single string scoped to one workspace, and it carries its own ceiling: rows per answer, calls per hour, how many sending profiles the workspace may hold.

Since 8 September the door logged 27 registration attempts and completed 20 of them. The clients that arrived named themselves: Claude eleven times, a coding agent from OpenAI nine times, ChatGPT once, a third party agent framework twice, our own coding client twice, and one was our probe. So the honest answer to "ChatGPT or Claude" is that both arrive through the same door, and the busiest traffic so far comes from agent clients that run in a terminal.

What a key gives the model is a vocabulary of tools over your own data: who is in the audience, what went out, who replied, what is queued for tomorrow, what the budget looks like. We described the read half of that surface in read access to your own pipeline, and the ten minute setup path in setup in 10 minutes.

Two clocks, and knowing which one you are on

Of the keys that ever made a call, 37 have a measurable gap between the key existing and the first question asked through it. The median gap is 0.3 minutes. 26 of the 37 made that first call inside ten minutes, 28 inside an hour, 30 inside a day. The slowest waited 13.2 days, which tracks a person's calendar.

The sending side runs on the other clock. Across 41 LinkedIn profiles connected since 1 August, 20 have sent something, and the median gap from connecting the profile to the first invitation leaving it is 2.95 days. The longest was 18.9 days. That gap is deliberate: a profile that starts at volume on day one gets throttled by the platform, so the first days go to a ramp. The mechanics of that ramp are in warming a new sending profile.

Practical consequence for your first afternoon: plan on asking questions and shaping the audience on day one, and plan on the first invitation leaving later in the week.

What a prospecting session looks like in the log

Counting a session as calls with gaps under thirty minutes, the door saw 88 sessions. The median session is 4 calls and runs about a minute. 27 sessions were a single call, which is someone checking one number between meetings. 25 sessions ran 10 calls or more, and those average 96 minutes. The longest single session made 223 calls over 31 hours, which is an agent left running against its own to-do list.

The opening question is almost always orientation. 28 of the 38 keys with any traffic opened with one of the two tools whose whole job is to answer "where am I and what happens next": 19 with the next step tool, 9 with the guided setup step. Three opened with the budget, two asked which workspaces they could see, two went at the counts, and one asked for a message straight away.

After orientation the working tools take over. Audience search ran 67 times across 11 keys, handing back an average of 26 people per answer. Source search ran 46 times. The upcoming queue was pulled 23 times at an average of 64 rows. Recent outreach was pulled 20 times at an average of 113 rows. Exclusions ran 33 times, and each landed call took around 105 people out of reach in one move.

That last number is the one worth copying. The most useful work a connected model does in week one is narrowing: pulling current customers, open opportunities, partners and competitors out of the audience before a single message is written. One call, a hundred rows, done.

Speed: 186 ms for a read, seconds for anything that thinks

Latency splits cleanly along one line, and it changes how you ask.

Plain data reads, 453 of them, came back at a median of 186 ms, with a 90th percentile of 459 ms and a slowest case of 2.0 seconds. These are counts, lists, queue contents, reply history. Ask for them as often as you like; the cost of a second look is nothing.

Tools that run a model behind the door are slower by an order of magnitude. The two orientation tools sit at a median of 3.4 and 4.8 seconds. Writing a message ran 17 times at a median of 7.5 seconds. Enriching people against outside sources ran 5 times at a median of 23.7 seconds, because that one waits on other people's APIs.

So the shape that works is: read often and cheaply, batch the thinking. A model that asks for counts twenty times in a session pays almost nothing for it. A model that generates text person by person in a long loop spends its afternoon waiting, and the way out is to hand it the whole list at once.

What the door refuses, and why that is the point

The door turned down 56 calls out of 924, which is 6.1%. Here is what they were.

Several refusals were about grounding. The door declined to produce first lines for a set of people because it could find no checkable fact to stand on, and it said so plainly, asking for a fact to work from. That rule exists because the opening line is the whole bet; we wrote it up in one checkable fact per message. One draft came back at 900 characters against a 700 character cap and was refused on length. A handful were argument errors, the ordinary kind: a lead list or a company was required and neither arrived. One call would have struck out 0 of 1 people and pulled 0 rows, so it was stopped as a no-op.

The ceilings live in the key itself. Rows per answer run 50, 100 or 200 depending on the scope, calls run 120 per hour, and the entry scope holds one sending profile. Across the whole log, 924 calls carried 11,383 rows, and 28 calls came back at the row cap. A model that asks for the whole base in one answer gets a page of it.

The hard boundary sits at sending. The model queues; a separate layer decides when anything leaves, spreads the day, and holds the daily ceilings. Over the last 30 days that layer put out 1,397 invitations and 1,519 messages from 21 profiles, each one paced across the day. The caps it respects come from platform and vendor documentation, and we keep the current reading on LinkedIn limits.

The rule we hold to on this tier is short: viewing and narrowing are open to the connected model, and widening is a decision a person makes. An agent can see the audience, cut it, read every reply and draft every message. Adding people and pressing send stay with a human.

A setup you can run this week

1. Get a key and point your client at the door. One string, one workspace. This is a text field and a browser tab, and the log says the median person is asking questions 0.3 minutes later.
2. Ask the orientation tool first. 28 of 38 people did, and it is the cheapest way to learn what the key can reach.
3. Narrow before you write. Load customers, open deals and partners into the exclusion list. One call removes about a hundred rows, and it protects relationships that outbound can cost you.
4. Connect the sending profile on day one even though it sends later. The median gap to the first invitation is 2.95 days, and that clock starts when the profile connects.
5. Batch the generation. Reads cost 186 ms, writing costs 7.5 seconds per call. Hand the model the list, collect the drafts, review them in one sitting.

If you want the comparison between running this yourself and having it operated for you, the trade-offs sit on our AI SDR page, and the entry tier where you bring your own model is on the MCP page from $199 a month by the number of connected LinkedIn accounts.

Questions buyers ask