The setup splits into two halves with very different clocks. Pointing your model at an outbound stack and getting a live answer back takes minutes, and our logs put the median at 24 seconds. Getting a LinkedIn profile into a state where it can send takes about two days. Knowing which half you are in on any given afternoon is most of what this page is for.
We run an MCP door over our own pipeline and hand keys to the people whose outreach we operate. 103 keys issued since 17 August 2026, 55 of them live, 31 with at least one tool call, 882 calls through 7 October. Small numbers, counted carefully, and they describe the setup path as it actually runs.
The four things you connect
In order, because the order is what people get wrong:
- A key. One string, scoped to one workspace. It arrives from the onboarding page or from the bot, and it carries its own caps: rows per answer, calls per period, how many sending profiles the workspace may hold.
- Your model client. Claude, ChatGPT, a coding agent, anything that speaks the Model Context Protocol. The door registers the client itself through dynamic client registration: 12 registration attempts over the window, 10 of them completed. The clients that showed up named themselves Codex four times, Claude twice, ChatGPT once, a third party agent framework twice, and one was our own probe.
- A LinkedIn profile. The model calls one tool, the door hands back a one time link, you open that link and sign in. This is the half with the slow clock.
- A mailbox, when you want the second channel. Separate tool, separate link, same shape.
Steps one and two are a text field and a browser tab. Step three decides when the first invitation leaves.
The first ten minutes, as the log sees them
Across 31 keys that ever made a call, the median gap between the key existing and its first tool call was 0.4 minutes. 22 of the 31 made that first call inside ten minutes, 24 inside an hour, 26 inside a day. The slowest key waited 7.2 days, which tracks a person's calendar.
Inside each key's first ten minutes the door saw 237 calls across 22 distinct tools, a median of 7 calls per key, with a range of 1 to 19. So the shape of a successful setup is roughly: connect, then ask seven questions in a row.
What the very first question is, almost always, is orientation. 20 of the 31 keys opened with one of the two tools whose whole job is to answer "where am I and what happens next": 12 with the state summary, 8 with the guided setup step. Three opened with the budget, two asked which workspaces they could see, and the rest went straight at the audience or the counts.
26 of those 237 first ten minute calls came back refused, an 11% failure rate against 5.8% over the whole window (51 of 882). The reasons are the setup itself talking: a marker that reads through a connected LinkedIn account before one is connected, a required argument the model guessed at, a source with an empty credit balance, a text generation that came back in the wrong shape and asked to be called again. Three of those four go away if step three is done before the model starts exploring, which is the one scheduling lesson in the data.
The step with the slow clock
Connecting the LinkedIn profile is the part worth planning around.
15 connect links issued, 8 opened, 4 carried through to a connected account. On those 4, the median time from the link being created to the account being live was 16.7 hours. The fastest was 42 seconds, by someone who asked for the link while already looking at the screen. Measured from the moment the link was opened, the median was 7.2 hours: people open these links, look at them, and come back later.
Then the profile has to become a sender. Across 25 profiles that eventually sent something, the median gap between connection and first invitation was 1.86 days, with one profile sending the same day and the slowest taking 18.9 days. That gap is deliberate: a brand new sender climbs a day by day ramp, and the pacing rules that keep an account healthy spread the first week thin on purpose.
The first seven days of a profile's life, measured on the same 25 profiles, averaged 17.9 invitations, with a median of 15 and a range of 3 to 80. So a realistic expectation for the first week after a ten minute setup is around fifteen invitations, and a realistic expectation for a first reply is the week after that.
For scale: 1 689 invitations went out across all sending profiles in September, 438 of them were accepted, which is 25.9%, and 80 of those people replied for the first time, which is 4.7% of everyone invited and 18.3% of everyone who accepted. The benchmarks page keeps the current version of those rates.
What the model may do once it is in
A client key sees 28 tools. 16 of them are marked read only, 12 change something, and 3 of the 12 carry the destructive hint, which is how a well behaved app knows to ask you before it runs them: queueing a touch to a live person, excluding companies, and pulling a queued touch back.
The call mix follows that shape. 654 of the 882 calls were reads, 228 changed something, and 152 of the 228 were the guided setup step itself, which writes to the session as you answer it. The reads carried 11 091 rows out, and the 200 row ceiling was reached 27 times.
Two properties of the door matter more than any single tool.
One key sees one workspace. 246 calls carried at least one workspace id in their result, and every one of them carried a single id. A key issued for one workspace answers questions about that workspace and declines the rest, and the door holds that line in code.
A model may narrow and stop, and widening stays with people. The model can exclude companies, pull a queued touch, hold a draft. New people and new text going to live humans pass the same checks a human operator's work does. Five of the six drafts submitted through the door came back refused with the reason named: an em dash, a 900 character text against a 700 cap, wording the outbound rules keep out of cold messages, an empty body. The door refusing your model's first draft is the system working as designed, the same way it refuses ours.
That is also why the empty answer is normal early. 89 of 163 conversation reads returned zero rows, and 50 of 63 audience reads returned zero while the other 13 carried 1 739 rows between them. A pipeline connected an hour ago is empty, and a door that says so plainly beats a door that pads the answer.
Latency: a fifth of a second against twenty seconds
Reads that touch only our own tables land in a fifth of a second: 177 ms median on conversations, 182 on the counts, 194 on the brief behind one message, 198 on the upcoming queue, 208 on recent sends. Your model can ask six of those in a turn and the turn still feels live.
Anything with a model or an outside provider behind it runs on a different clock. The guided setup step has a 6.8 second median and a 27.4 second ninetieth percentile. Writing a first text runs 8.0 seconds at the median and 21.2 at p90. Enrichment, which goes out to providers, sits at 23.7 seconds median and 58 at p90.
The practical consequence: have your model do its reading first and its writing last. A session that interleaves them spends most of its wall clock waiting, and that is the only part of this setup that genuinely feels slow.
A checklist you can run tomorrow
Three things, in this order, and the whole thing fits in an afternoon:
- Request the key and connect the LinkedIn profile first, before you open your model at all. This removes the most common first ten minute failure and starts the 1.86 day clock immediately.
- Open your model, add the connector, and ask the orientation tool one question: what is the next step. 20 of 31 keys did exactly this, and it is the cheapest possible first call.
- Read before you write. Audience, queue, recent sends, counts. Four calls, under a second total, and you will know whether the pipeline has anything in it to reason about.
Then expect fifteen invitations in week one and the first replies in week two. The read access article covers what the twenty reads actually answer once there is data behind them, and the AI SDR page covers where this step sits next to the operated versions.
The step itself is $199 a month for one connected LinkedIn account. The setup is ten minutes. The profile is two days. Plan for the two days.
Questions buyers ask
How long does it take to connect LinkedIn to Claude for outreach?
The key and the connector take minutes: across 31 keys the median gap between the key existing and its first working tool call was 24 seconds, and 22 of the 31 made that call inside ten minutes. The LinkedIn profile itself is slower, with a 16.7 hour median from link to connected account on the four that completed, and a 1.86 day median from connected profile to first invitation across 25 profiles.
What can a model actually do through an outbound MCP server?
A client key sees 28 tools, 16 of them read only. Reads cover the audience, the queue, recent sends, replies, the counts and the budget, and they land in about 200 ms. The 12 writing tools cover exclusions, pulling a queued touch, submitting a draft and requesting a human decision, and three of them are marked destructive so your app asks before running them.
Is pipeline data from one workspace visible to another key?
Each key answers for exactly one workspace, and the door holds that line in code. Over 882 calls, 246 carried at least one workspace id in their result, and every one of those 246 carried a single id. A key issued for one workspace answers for that workspace and declines questions about the others.
How many invitations should I expect in the first week?
Across 25 profiles, the first seven days after connection averaged 17.9 invitations, with a median of 15 and a range of 3 to 80. The low end is deliberate: a new sender climbs a ramp over its first days. For context, September across all sending profiles came to 1 689 invitations, 438 acceptances at 25.9%, and 80 first replies at 4.7% of those invited.