HubSpot's Claude Connector vs a Custom MCP: 12 Sales Jobs Tested

By Izzy Aly, Co-Founder, Systems and Architecture


The official MCP connectors we see are usually built on the vendor's API. HubSpot's is one of them. Its tools do what the API does: search for a record, read a record, update a record. That is how APIs worked long before LLMs.

Now connect that CRM to Claude. A rep asks, "What should I focus on this week to reach my target?" With HubSpot's official connector, Claude looks at its tools and sees that it can search for records. So it goes and gets all the open deals the rep has access to.

Then Claude looks through those deals and decides which ones make sense. It might pick the ones with a higher probability of closing, a higher value or more engagement. It might only check the few deals with the highest probability and focus on those. Either way, we are letting the LLM get the deals, do the calculations and decide what is high priority.

If your business has its own definition of high priority, you don't want that definition to be up to the LLM. You want an answer that is deterministic and doesn't change every time you ask.

A custom connector starts from that definition. When someone asks what to focus on this week, the custom connector tool gets the open deals and sorts them by your rules, grabs all relevant context and runs all this on the server side, which is faster and doesn't fill up the conversation context with noise. For example, a business can define a high priority deal as having at least two contacts, a set number of past engagements and a completed discovery call. The tool sends that back to Claude with all the needed context for Claude to easily identify the highest priority deals and the next actions needed to help the rep reach their target.

We put both to a test.

We asked Claude to do 12 everyday sales jobs in HubSpot: a morning brief, a forecast, call prep, logging a meeting, a bulk clean-up of close dates. Then we ran the same 12 jobs four ways, through HubSpot's own Claude connector and through a custom connector we built, each with and without a sales plugin.

One thing you should know before the results: we built the custom connector in this test, Pronova Sales for HubSpot, and its plugin, so we are testing our own product. We also built it against these same 12 jobs and fixed three things in it between our last two runs, while HubSpot's plugin didn't change. Each setup ran each job once, on one model, with made-up data in HubSpot test portals, so take the results as a first look and not as a benchmark.

Key takeaways

On the simple jobs there was no difference. All four setups got the forecast, the follow-up email, the contact import and the bulk close-date fix fully right.

The difference showed up on the jobs where your own rules decide what a good answer is. In call prep the custom connector listed every open task and HubSpot's connector missed them. In logging a meeting both HubSpot setups updated the deal and added the new contact but never logged the meeting itself, although part of that gap comes from how we wrote the answer key. Our own plugin got one wrong too, because it gave the follow-up task a title our check doesn't accept.

Both sides asked before changing anything in the CRM. HubSpot's setups asked in chat, where a typed "yes" took 3 to 6 seconds on seven of the eight change jobs. The custom setups sent every change to a separate approval page, which took 32 to 82 seconds, and in this test the page didn't catch anything the chat missed.

The custom connector also got there with less work: 18 CRM lookups instead of 52, and 2.4 minutes of Claude's time instead of 5.4.

So HubSpot's connector is the right start if you want something working this week with nothing to build. A custom connector makes sense when your team asks the same five questions every week and has written rules for the answers.

What we tested

Claude reaches HubSpot through two things. The first is the connector, which is what lets Claude read and change your CRM (technically it's an MCP server that offers Claude a set of tools to call). The second is the sales plugin, which is a set of written instructions for sales jobs like a morning brief or call prep, where each instruction is called a skill.

HubSpot's connector and sales plugin are the same for every HubSpot customer, so you switch them on and there's nothing to build. The connector is one of three ways HubSpot lets a coding agent into your CRM, and we compared all three in HubSpot Agent CLI vs MCP vs custom API.

A custom connector starts from your team's sales rules, like what counts as a silent deal, what a Commit deal has to show and who approves a change, and its server applies those rules and gives Claude a finished answer. It only covers the jobs you design it for, and when your rules change they have to be changed in the connector too.

Think of a new rep on their first day. HubSpot's connector gives them the keys to the CRM and a general manual, while the custom connector gives them your pipeline rules already applied to every deal.

That gives us four setups to compare, which the table below calls HubSpot, HubSpot + plugin, Custom and Custom + plugin. Everything else was the same for all four:

  • All four ran on the same model, Claude Sonnet 5.5 in Claude Code, and got the same 12 prompts in one session, first the eight that only read the CRM and then the four that change it.
  • Each side had its own HubSpot developer test portal filled by the same seed script, so both had the same deals, contacts and tasks.
  • All four got the same written sales rules in Claude's instructions, so HubSpot's side knew the rules too and the only difference is who applies them.
  • Versions: everything ran on what was live on 9 November 2026, and the exact version numbers are at the end of the post.

Where did the answers split?

We wrote scripts that scored each answer against an answer key, or against what the CRM held afterwards, so nobody rated the answers by hand. 100% means the answer matched the key in full, and how each job was scored is at the end of the post.

JobHubSpotHubSpot + pluginCustomCustom + plugin
T01 Daily brief50%100%100%100%
T02 Deals that need attention †30%9%21%100%
T03 Forecast100%100%100%100%
T04 Call prep86%60%100%100%
T05 Contact lookup100%67%100%100%
T06 Follow-up email100%100%100%100%
T07 Contact lookup, two matches †50%100%100%100%
T08 Follow-up task100%100%100%50%
T09 Onboardingn/an/an/an/a
T10 Log the meeting62%62%100%100%
T11 Import contacts100%100%100%100%
T12 Bulk close-date fix100%100%100%100%
Average of the 11 scored jobs80%82%93%95%

† Be more careful with the two rows marked †. We ran the whole test the day before as well, and in between we fixed things in our own connector and changed how T02 is scored, which is why our scores on T02 and T07 are different from that earlier run. The custom connector's 100% on call prep and the meeting log doesn't come from those fixes, because it already scored 100% on both before we made them. T09 (onboarding) has no answer key, so we didn't score it.

On the forecast (T03) all four setups came back with exactly the same answer: €205,000 in Commit with €31,000 of it backed by evidence, €396,100 past its close date across 31 deals, and the same three deals at risk.

Five results are worth a closer look.

  1. Call prep went to the custom setups, which listed every open task and the people at Acme nobody had contacted yet. HubSpot's connector missed the open tasks, and with its plugin it also missed the buying group and the contacts nobody had reached. This is the clearest win for the custom connector, because a rep who doesn't see the open tasks walks into the call without knowing what they already promised the customer.
  2. The gap on logging a meeting comes partly from our own answer key. The prompt packed five things into one message (a pilot, a new close date, a new amount, a promised document and a new contact), and both HubSpot setups updated the deal and added the contact. They lost points because our key also expects a meeting record, plus a note and a task linked to both the deal and the company, and the sales rules didn't ask for any of that. Neither HubSpot setup logged a meeting. HubSpot's plugin wrote a note and a task linked to the deal and the contact but not the company, and HubSpot's connector on its own wrote a task linked to the deal only. So the 62% partly measures how our connector logs a meeting, and someone else's key could score it differently. For you this only matters if your team wants every meeting saved as its own record and linked to both the deal and the company. If it doesn't, HubSpot's setups did what the rep asked and you can ignore this gap.
  3. The custom plugin lost the follow-up task because it created the task under a title our check doesn't accept, while the other three setups got it fully right. This one counts against us. The task was still created, so the rep loses nothing on the day, but it shows that our plugin can name things differently from what your team expects, and that matters if you have reports or automation that find tasks by their title.
  4. Listing every flagged deal is where the results were most uneven. Only the custom plugin named all 33 flagged deals, while HubSpot's connector named 10, the custom connector on its own named 7 and HubSpot's plugin named 3. That result depends on a fix we shipped between runs, because before it the custom plugin reused the morning brief. It is also the result with the biggest effect on a rep's week, because a deal that isn't on the list doesn't get looked at, and three of the four setups left most of the flagged deals out, including our own connector without its plugin.
  5. HubSpot's setups also dropped single facts. The connector on its own missed one overdue task in the daily brief and one of the two contacts who share a name, and with its plugin it missed one fact in the contact lookup. Each miss is small, but the answer still looks complete, so the only way a rep finds out is by checking the CRM themselves.

Three setups also stopped to ask a question. The custom plugin asked how to word a simple task and when it was due, HubSpot's plugin asked which job title to keep for a duplicate CSV row, and HubSpot's connector asked what type of task to create. That's fine when you're sitting at the keyboard, but it stalls a job that runs on a schedule at 6 a.m. with nobody there to answer.

Who checks a change before it lands?

Both sides asked before writing anything to the CRM, but they asked in different places.

HubSpot's setups asked in the chat. Claude showed the changes it planned to make and waited for a typed "yes", which took 3 to 6 seconds on seven of the eight change jobs. The eighth took a minute because HubSpot rejected the first write and Claude had to ask for a task type. Both HubSpot setups also offer to skip the confirmations for the rest of the chat, and if you take that offer the later changes in that chat go in without a confirmation.

The custom setups asked on a separate page. We switched on an optional rule that says nothing reaches the CRM until a person approves it on a page that lists every field, and that took 32 to 82 seconds per change job. That is about a minute per change compared with a few seconds in the chat, which is the price of the extra check. Those timings leave out two jobs on the custom connector alone, because on the follow-up task Claude mistyped the approval link and had to correct it, and the meeting log's wait included that task's second approval. We've fixed the link bug since.

In this test the page didn't catch anything the chat missed. On the bulk close-date job, HubSpot's connector listed all 31 deals in the chat with their current close dates and described the 31 tasks once, while the custom connector showed 10 of the 31 deals in the chat and put all 62 changes on the page. Both applied all 62 changes correctly.

So the real difference is who writes the list you approve. In chat, Claude writes it and decides what to show you. On the page, the server builds it from the changes actually waiting to go in, and nothing lands until a signed-in person approves. That matters on the day Claude gets something wrong or a session runs with nobody watching, and neither of those happened in this test. I'd still keep the rule on, and The gate belongs inside the tool explains why the check belongs in the connector and not in the prompt.

Which setup fits your team?

Which one fits depends on two things: how fixed your sales rules are, and how often your team asks the same questions.

HubSpot's connector and plugin are enough if:

  • Most of your questions are one-offs, because it can answer anything in the CRM without anyone designing the job first.
  • Your rules for what counts as "at risk" are loose or still changing, so there is nothing fixed to build in yet.
  • You want to try this out this week without anyone building anything.

A custom connector is worth building if:

  • The answer you need sits across HubSpot and two or more other tools, which is where a connector per tool leaves the most work to Claude.
  • You have written rules for what is stalled, at risk and ready to close, and you want the same answer every time someone asks.
  • Someone asks the same five questions every week (brief, pipeline, forecast, call prep, follow-up), so the work of designing each job pays back every week.
  • A person has to review changes to the CRM before they land.

If you're somewhere in between, start with HubSpot's connector and then build a custom connector next to it for the one job where you lose the most time or catch the most errors by hand. That is what a custom MCP build with us looks like: we start with a written gap analysis of what HubSpot's connector doesn't cover for your team, and then build that one job with your rules and an approval page in front of every write.

How much work each connector did

Everything Claude pulls from your CRM stays in the conversation, and Claude rereads all of it on every step, so a connector that returns less leaves Claude with less to carry. The custom connector made 18 CRM lookups where HubSpot's made 52, and Claude worked for 2.4 minutes where it worked for 5.4 with HubSpot's.

In a 12-job test that difference doesn't cost you anything, because the biggest session used about 15% of what Sonnet 5.5 can hold, so it tells you how each connector behaves more than it tells you what you would save.

Custom connector as a share of HubSpot’s Dot chart, HubSpot’s setup equals 100%, scale 0 to 100%. For each measure, a filled dot compares the custom connector alone with HubSpot’s connector alone, and a hollow dot compares each with its sales plugin. CRM lookups: 35% alone, 53% with plugin. Claude’s writing: 32% alone, 47% with plugin. Claude’s time: 44% alone, 57% with plugin. Context added per job: 48% alone, 64% with plugin. Session size: 62% alone, 74% with plugin. One run per job, 12 jobs, invented data. Source: ProNova test runs, 9 November 2026. Custom connector as a share of HubSpot’s Connector alone With sales plugin HubSpot’s = 100% Smaller than HubSpot’s CRM lookups 18 vs 52 · 20 vs 38 35% 53% Claude’s writing 15k vs 47k · 20k vs 43k tokens 32% 47% Claude’s time 2.4 vs 5.4 · 3.0 vs 5.3 min 44% 57% Context added per job 4.2k vs 8.8k · 5.9k vs 9.2k tokens 48% 64% Session size 89k vs 143k · 111k vs 150k tokens 62% 74% 0% 20% 40% 60% 80% 100% One run per job, 12 jobs, invented data.
Custom connector as a share of HubSpot’s Dot chart, HubSpot’s setup equals 100%, scale 0 to 100%. For each measure, a filled dot compares the custom connector alone with HubSpot’s connector alone, and a hollow dot compares each with its sales plugin. CRM lookups: 35% alone, 53% with plugin. Claude’s writing: 32% alone, 47% with plugin. Claude’s time: 44% alone, 57% with plugin. Context added per job: 48% alone, 64% with plugin. Session size: 62% alone, 74% with plugin. One run per job, 12 jobs, invented data. Source: ProNova test runs, 9 November 2026. Custom connector as a share of HubSpot’s Connector alone With sales plugin HubSpot’s = 100% CRM lookups 35% 53% Claude’s writing 32% 47% Claude’s time 44% 57% Context added per job 48% 64% Session size 62% 74% 0% 50% 100%
Fig. 1 — The custom setups as a share of HubSpot's: connectors alone (filled dots) and with their sales plugins (hollow dots). The chart leaves out answer quality. Source: ProNova test runs, 9 November 2026.

The plugins changed the picture a little. HubSpot's plugin cut its lookups from 52 to 38, and our plugin added two lookups and half a minute of Claude's time in return for the full list of flagged deals, 33 of 33 where the connector on its own named 7.

Why the custom connector did less work

The reason is how it's built: the custom connector does the calculations on its own server and only sends Claude the result, so Claude has fewer records to fetch and less to work out itself.

Who does the sums: HubSpot’s connector vs a custom connector Two flows from HubSpot to a forecast. HubSpot’s connector: HubSpot sends raw deal records over several lookups, and Claude sorts them, adds them up and applies your rules before giving the forecast. Custom connector: a server applies your rules and adds up, then sends finished figures plus the deals behind them in a single answer, and Claude explains them as the forecast. Both flows reach the same forecast; the sums move from Claude to the server. Who does the sums: HubSpot’s connector vs a custom connector HUBSPOT’S CONNECTOR HubSpot Raw deal records several lookups Claude sorts, adds up and applies your rules Forecast CLAUDE DOES THE SUMS CUSTOM CONNECTOR HubSpot Server applies your rules, adds up Finished figures + the deals behind one answer Claude explains Forecast THE SERVER DOES THE SUMS Same forecast. The custom connector moves the sums to the server.
Who does the sums: HubSpot’s connector vs a custom connector Two flows from HubSpot to a forecast. HubSpot’s connector: HubSpot sends raw deal records over several lookups, and Claude sorts them, adds them up and applies your rules before giving the forecast. Custom connector: a server applies your rules and adds up, then sends finished figures plus the deals behind them in a single answer, and Claude explains them as the forecast. Both flows reach the same forecast; the sums move from Claude to the server. HUBSPOT’S CUSTOM HubSpot Raw deal records several lookups Claude sorts, adds up and applies your rules Forecast HubSpot Server applies your rules, adds up Finished figures + the deals behind one answer Claude explains Forecast Shaded: where the sums happen.
Fig. 2 — Both connectors reach the same forecast, but the custom one does the sums on the server and HubSpot's leaves them to Claude. Source: ProNova test runs, 9 November 2026.

This matches Anthropic's advice on how to build MCP tools. Its guide to writing tools for agents says a tool should do several steps inside itself and return only what matters, so the working-out doesn't fill up the conversation, and its note on code execution with MCP makes the same case for filtering and adding up data before it reaches the model.

Chroma's context-rot study tested 18 models and found that they get less reliable as their input grows, even on simple tasks. That study is from July 2025, covers models up to Claude 4 and wasn't about CRM work, but it points the same way: the more raw records Claude has to read to answer one question, the more room there is for it to miss something.

You could ask whether a skill with a calculation script on top of HubSpot's connector would do the same job, and the answer is only partly. Claude's Agent Skills docs say a script's code stays out of the conversation and only its output comes back, so the sums themselves would stop taking up room.

The script needs the deals first, though, and with HubSpot's connector Claude has to fetch them through its tools, which puts every record into the conversation before the script runs. To keep the records out as well, the code has to call the connector itself, which is the setup Anthropic describes in its code execution note above, and that needs an execution environment that can reach the connector. We didn't test either setup.

What changes when the answer sits in several tools?

This test only used HubSpot, but a lot of revenue questions don't stay inside the CRM. Take a CS lead who asks, "Is the Acme renewal at risk?" The answer can depend on open support tickets, a failed payment and falling product use, and each of those lives in a different tool.

If every tool has its own connector that returns raw records, Claude has to pull from each one, work out that "Acme", "Acme Corp" and acme.example.com are the same customer, and then apply your rules inside the conversation. A custom connector can do that matching on its server and send Claude one record with the answer, the same way it did the sums for the forecast.

We didn't test this, so we have no numbers for it. We expect the difference in lookups to grow with every tool you add, because each extra tool is one more set of records for Claude to fetch and match, and that is why this case is first on the list of reasons to build a custom connector.

How far to trust these numbers

Treat these numbers as a first look and not as a benchmark, for the following reasons:

  • We built the custom connector and its plugin, ran the test and wrote the scoring scripts, so everything here comes from the people who benefit from the result.
  • Each setup ran each job once. It was twelve jobs, one model and one person running it who knew which setup was which, with invented data in two HubSpot developer test portals filled by the same seed script.
  • The results are from 9 November, which was our third run. We ran a five-job version on 5 November with the connectors only (four of the five jobs ran) and all 12 jobs on 8 November, and we fixed our connector after each of those runs.
  • Between the last two runs we fixed three things in our connector and plugin (meeting times, two contacts with the same name, and looking again on deal and pipeline questions), while HubSpot's plugin stayed at the same version. We can't see whether HubSpot changed its hosted connector.
  • We wrote the answer keys and changed the scoring between runs. The follow-up-task check now accepts "next step" in a title, the flagged-deals score now counts flag reasons, and for the change jobs the scores for both plugin setups now read the CRM as it stood when the session ended, before we reset the portal. That is tuning on the test, and it favours us.
  • The connector address that HubSpot's sales plugin ships with wouldn't let Claude Code sign in during our runs, so both HubSpot setups used HubSpot's registered connector address instead.
  • A lighter conversation should matter more over a long working day, and twelve jobs can't show that. The difference may only appear at 30 jobs or more.

When we run this test again, we want to remove the limits above. We won't change our connector once the test has started, we'll include questions it has never been tested on, we'll run every job three times and report the range, and we'll use the same approval step on both sides.

Someone outside ProNova should write the answer keys, and we'll add the script setups we didn't test this time.

Run the test yourself

If you want to check this before trusting us, take your own 30 most common jobs and run them in one session on each setup against a HubSpot sandbox. Then ask the same question at job 1 and again at job 30 and compare the two answers, because a long session is where the difference should show up.

Our 12 prompts are below with placeholder names, in the order we ran them on 9 November. T09 ran before T08 so that the eight jobs that only read the CRM came first. Before you run them:

  • Swap in people and deals that exist in your sandbox.
  • The last four change data, so never run them on live records.
  • T07 expects two contacts with the same name.
  • For T11, paste the rows from the second block on the lines after the prompt, with your sandbox's email domain in place of <your-domain>. The last row repeats the third in capitals, to test duplicate handling.
twelve-jobs.txt
T01  Good morning. What do I need to do today?
T02  Which of my deals need attention today, and why?
T03  What will I close this quarter, and what's at risk?
T04  Prep me for my call with Jane Doe at Acme tomorrow.
T05  Pull up Alex Example. Which company is he at, which deals is he on,
     and when did we last talk?
T06  Draft a follow-up email to Riley Sample about the Hooli deal.
     Draft only, don't send it.
T07  What's Jane Doe's phone and who else is at her company?
T09  I'm new to this. Get me started: what can you do with my HubSpot?

Changes data, sandbox only:
T08  Create a follow-up task for the Globex deal.
T10  Just finished with Acme: they'll pilot, close moves to Dec 15,
     amount €24k, I owe them the security doc by Friday, add their CFO
     Dana Sample.
T11  Import these contacts into HubSpot:
T12  Move every past-due close date on my open deals to the end of next
     month and add a task on each to confirm timing.
t11-contacts.csv
firstname,lastname,email,company,jobtitle
Pat,Example,pat.example@<your-domain>,Acme Corp,Facilities Manager
Sam,Sample,sam.sample@<your-domain>,Umbrella Logistics,Operations Lead
Jo,Test,jo.test@<your-domain>,Example Labs,Head of Revenue Operations
Jo,Test,JO.TEST@<YOUR-DOMAIN>,Example Labs,Head of RevOps

The versions we ran on 9 November 2026 were Claude Code 2.1.319, HubSpot's sales plugin at version 29c7342 and our plugin at version 12d49e9. Both connectors were the versions live that day.

How we scored each job:

  • Questions about facts (T01, T05, T06, T07): how many of the expected facts the answer included.
  • Deals that need attention (T02): how many of the 33 flagged deals the answer named, and how many of their flag reasons it gave.
  • Forecast and call prep (T03, T04): whether the figures and deals matched our answer key. For call prep we also checked for open tasks, the buying group and contacts nobody has reached.
  • Jobs that change the CRM (T08, T10, T11, T12): what the CRM held afterwards, record by record. For T08 the task title has to contain "follow-up", "next step" or "next meeting".

If your numbers come out differently, we'd like to know where we got it wrong: hey@pronovahq.com.

We connect the tools your team already uses to coding agents like Claude. If you have a job where your own rules should decide the answer, in your CRM or across several tools, book a free 30-minute workflow scoping call and we'll tell you whether the connectors you already have cover it or whether it needs a custom one next to them.

Sources: HubSpot, set up and use the HubSpot connector for Claude · Anthropic, writing tools for agents · Anthropic, code execution with MCP · Claude docs, Agent Skills · Claude docs, models overview · Chroma, context rot. Test data and timings from ProNova's run on 9 November 2026.

Updated 12 November 2026. We rewrote this post so it is easier to read and opens with the problem the test is about. We took out the token cost figures, because the runs used a Claude subscription and nobody paid them, and we corrected three shares in Fig. 1 (now 44%, 57% and 64%, shown before as 45%, 58% and 65%) so they match the numbers printed next to them.

Updated 11 November 2026. The first version of this post, published on 9 November, had two claims we couldn't back up with the run data, so we took them out: a comparison of total time from start to finish, and a line saying HubSpot's plugin wrote changes without asking. We also corrected Claude's working time with the custom connector from 2.3 to 2.4 minutes and added the full score table.

Topics
Author
Izzy Aly · Co-Founder, Systems and Architecture
Share

Work with us

Got more workflows than time to fix them?

The AI Workflow Lab maps where AI fits across your CRM and the tools around it, and tells you what to build first and what each build will cost.

or book a scoping call (opens Google Calendar)