Connecting Capsule CRM to Claude: The MCP Door Reads, the API Door Writes

I partner on Capsule, I am a customer on a trial clock, and in October 2026 I wanted Claude to have a seat in my own account. Two jobs for it: run the affiliate-programme CRM, which is 239 programmes as organisations and projects on a board, and run the booking layer from the free scheduling recipe. Capsule gives an AI two ways in, and I set up both on 12 October. They are not the same door.

Two doors

The finding of the afternoon is the Writes row. Capsule's own tool reference is honest about it: the three write tools "can't rename a record, reassign its owner, or change any other core field". So an assistant connected through the front door can see your whole pipeline and cannot add a note, create a task, move a deal or make a pipeline. For the management work I wanted done, the door is the API, and the API is where I spent the rest of the day.

What the front door took

The developer portal wants a GitHub login. Register an application: a name, a description, a redirect URL, which must be HTTPS, no wildcards, one per app, and an application type, where "Web" is the one that gets a client secret. The secret sits in clear text on the app page next to a reset button.

Claude's side is a two-step dialog. You give it a name and the server address, it probes the server, and it offers three ways to introduce itself, which is a short education in how AI clients and servers are learning to meet:

1. Use Claude's published identity. The server fetches Claude's client details from an address Anthropic hosts. Nothing to register. Capsule does not support it.
2. Register automatically. Claude registers a client with the server on the fly. Capsule does not support that either.
3. Paste your own client id and secret. The only one Capsule accepts.

Capsule chose the most manual of the three, and I can see why: it controls abuse, it gates by plan at the account level, and a small team ships the minimum. The cost lands on every user, who has to go and find the callback address Claude never shows in its own dialog. Mine came from a vendor's help page, and a one-line hint in the dialog would save every Capsule customer the same search.

Capsule announced the MCP server on 26 June 2026 as an early-access beta for Growth and above, and the post's verb is "ask": questions about contacts, pipeline, tasks and projects, with answers "drawn straight from your live CRM data". The three write tools came after the post, and they are the smallest writes a vendor could add while still saying the assistant can change something.

What the back door took

A token from the preferences page. The token is the user: everything it writes is attributed to that person, which is the right design and worth knowing before you hand one to a script. There is a read-only scope, and a read token on a write call gets a 403 that says so.

Then the API itself, which is well made and has one fossil. Projects are still kases in every path and root key. Cases were renamed Projects in the interface in June 2022 and the API kept the old word so nobody's integration broke. Every new developer meets the fossil on day one, and I like what it says about the company: backward compatibility as a visible value, paid for in small confusion forever.

Writes wrap the object in a root key, updates are partial so a field you leave out is untouched, sub-items are added without an id and changed with one, and the filter endpoint is the same engine as the saved lists in the app, twelve operators including "is within last N days". Four thousand requests an hour per user, with the remaining count in every response header.

What the API does not have is history. There is no endpoint that tells you what a field said last week. I wrote about that on the won-deal page as the filing cabinet test, and here it has a practical consequence: every change an automation makes should also write a note, or it is simply gone.

What I built, and what the first job taught me

A small MCP server of my own, eleven generic tools over the REST API, running on my machine and registered with Claude Code. Rather than sixty typed tools, one per operation, it has list, get, search, filter, create, update, delete and a cheat sheet of the body shapes, so the model does not have to re-read Capsule's documentation to add a note. The hosted server went the other way, one tool per operation, and I think both are right for their audience: the generic shape is easier to keep complete when a model is driving, and the typed shape is easier for a vendor to sandbox.

Every write it makes is appended to a log file: time, method, path, body. Five lines of code, and it is the audit trail the product itself does not keep.

The first real job was pushing 239 affiliate programmes into Capsule as organisations and projects on a new board, with six stages, eight custom fields and a note per project. Dry run first, which printed every action it would take. Then ten rows. Then the rest. Roughly seven seconds per row for three writes with a polite quarter-second gap, so the whole board took the better part of twenty-five minutes and never came near the hourly limit.

Three things from that run that I would tell anyone wiring a model to a CRM:

The model needs a projection, not the record. Two hundred and forty full project records with their embedded fields came back as 180 kilobytes of JSON and blew the assistant's context in one call. The fix was a compact mode on my server that returns one slim line per record. Capsule's hosted tools page at a hundred records for the same reason.

A CRM synced from files is only as current as the files. One programme had been approved on its network two days earlier, and my tracker still said nothing, so the board put it one stage behind the world. That is an argument for the vendor's own hooks, which Capsule has, nineteen event types for registered OAuth apps, and an argument for letting the CRM own status rather than a spreadsheet.

Vendor-seeded records are a trap for importers. My account already held an organisation called Capsule, with a logo and five social links, before I created anything, and the sample data Capsule seeds for new accounts sits beside the real data until you remove it. Match on exact names and expect company.

When the front door is the right one

For a person who is not technical and wants to ask the CRM questions in a chat window, the hosted server is exactly right and twenty minutes well spent. "Which opportunities on the Partnerships board have not moved in a month" is a reading question, and reading is what the door does.

For anything that changes the record, use the API, give the key the narrowest scope that does the job, and log what it writes. Vendors ship read-mostly first because a wrong edit by an assistant is their support ticket, and in October 2026 Capsule's ChatGPT connector gained the ability to log activities, set tasks and update custom fields, so the write side is widening. Until the product keeps a history of its own, the five-line log is the thing that lets you undo an AI's confident mistake, and the nightly export is the thing that makes the whole cabinet replaceable.

What each one pays me

Status as of 12 October 2026.

One tool on this page pays me, and it is the one I am mildly critical of. If the commission had chosen the verdict, the front door would have been fine.

Sources: Capsule's developer documentation on custom MCP connections, authentication, reading, writing and filters; Capsule's support article "Capsule MCP integration tool reference"; Capsule's blog, "What if you could just ask your CRM? Introducing Capsule MCP", 26 June 2026; Capsule's developer documentation for the Project operations, which records the 2022 rename; Anthropic's documentation on custom connectors and on MCP in Claude Code. All read 12 October 2026. The timings and counts are from my own account on the same day.