Using Claude with Attio via MCP: A practical guide
Connecting Claude to Attio via MCP (Model Context Protocol) gives an AI assistant secure, live access to your CRM: your records, deals, notes, emails and call recordings. Once connected, you can ask questions about your pipeline, update records in plain English, draft emails with full customer context, score deals, generate proposals and orchestrate workflows across your other business systems. Setup takes about two minutes.
We're an Attio consulting partner, and this connection has quietly become the highest-leverage thing we set up for clients. This guide covers how to connect Claude to Attio, what it's actually useful for day to day, real examples from client work, and just as importantly, when MCP is the wrong tool for the job.
What is MCP and why does it matter for your CRM?
MCP is an open standard that lets AI assistants like Claude connect to external tools and data sources. Instead of copying and pasting CRM data into a chat window, Claude can read and write to your Attio workspace directly, with your permission, using your access levels.
This matters more for Attio than for most CRMs because of what Attio captures automatically. If you've connected your email and calendar, and you're using call recording, Attio holds the full context of every customer relationship: every conversation, every meeting, every email thread. That's not just a database of names and deal values. It's a near-complete record of how your business actually talks to its customers.
Give an AI assistant access to that, and you're no longer asking it generic questions. You're asking it questions about your business, answered from your own data.
How do I connect Claude to Attio?
In Claude, go to Customise, then Connectors
Find Attio in the connector directory and click Connect
Authorise via OAuth and choose the Attio workspace you want to connect
Start a new chat. Claude can now search records, read call transcripts and emails, create notes and tasks, and update records in that workspace
Two things worth knowing before you start. First, the connection uses OAuth and inherits your permissions, so Claude can only see what you can see. Second, if you're an admin or consultant working across multiple Attio workspaces, check which workspace is connected before you run anything. Claude will happily answer questions from the wrong workspace with complete confidence (ask us how we know).
The everyday use cases
These are unglamorous and they're where most of the value sits.
Deal Q&A. "What's the latest on the Acme deal?" Claude reads the record, the recent emails and the last call transcript, and gives you the actual state of play rather than whatever the deal notes said three weeks ago.
Updating records conversationally. "Move Acme to Negotiation, set the value to $18k and add a note that they want to start in September." Done in one message, no clicking through the UI.
Drafting emails with real context. Because Claude can read the whole relationship history, a follow-up email references what was actually discussed on the last call, not a generic template. The difference in quality is immediately obvious.
Tasks and reminders. "Create follow-up tasks for every deal I haven't touched in two weeks" turns pipeline neglect into a to-do list in seconds.
CRM hygiene. This one is underrated. Your calls and emails contain accurate, current information that your structured CRM fields don't: job title changes, new stakeholders, revised budgets, shifted timelines. Ask Claude to review recent conversations against a set of records and update the fields to match. Your CRM starts reflecting reality instead of lagging behind it.
Bulk updates. Give Claude a list of records and a rule ("tag everyone who attended the June event", "set the industry field based on the company description") and it works through them. What used to be a CSV export, a spreadsheet session and an import becomes a single instruction.
Monitoring your whole business without pestering your team
Here's the shift that surprises most founders and sales leaders: because Attio captures calls and emails automatically, connecting Claude effectively gives you a research analyst with a complete view of your customer-facing activity.
Instead of asking a team member for a status update, you ask Claude. "Summarise what happened across the pipeline this week." "What did we commit to on calls with prospects this month?" "Which deals have gone quiet?" Nobody has to stop selling to write a report, and the answers come from what was actually said rather than what someone remembered to log.
This unlocks proper strategy conversations too. Where are we winning and why? What objections keep coming up? Which lead sources produce deals that actually close? These used to be quarterly-offsite questions answered by gut feel. With full conversation context, they're questions you can ask on a Tuesday and answer with evidence.
Deal scoring and forecasting
Forecasting is one of the most hated jobs in sales because it usually means chasing people for updates and then applying scepticism to what they tell you. With Claude connected to Attio, you can do it from the ground truth instead.
We recently built this for a client, a membership community whose key salesperson was leaving. Leadership needed fast, consistent visibility into a pipeline that lived largely in one person's head. We built a Claude skill that reviews each deal record along with every associated email, call and note, then writes back two fields: a structured, handover-ready deal summary and a likelihood score from 0 to 99.
The interesting part is the scoring philosophy: score on signals, not stage. A pipeline stage label like "Contract Sent" tells you almost nothing on its own, especially in sales processes where contracts go out early as a nudge. The signals that actually predict a close are recency of engagement, two-way dialogue, whether the real decision-maker is involved and warm, and explicit buying behaviour like asking about payment or booking the next step. A reply yesterday beats a contract sent a month ago. Claude can weigh all of that because it can read all of it.
The result was a scored, summarised pipeline the leadership team could triage in minutes, with every deal ready to be picked up by whoever inherited it.
Proposal generation
This was our own first serious use case. Claude reads the discovery call transcripts and email threads for a deal in Attio, synthesises what the prospect actually needs, and produces a structured proposal draft. In our pipeline, that output feeds a Google Slides template, so a document that used to take a couple of hours of re-reading notes and writing now takes minutes, and it's more accurate because nothing said on the call gets forgotten.
The general pattern applies to any sales document: the information you need already exists in your CRM as conversation. The AI's job is to turn conversation into artefact.
Orchestrating workflows across systems
MCP connections aren't limited to Attio. Claude can hold connections to several systems at once, which means it can run workflows that span them. This is where things get genuinely new.
One of our clients, an accounting firm, runs their client onboarding this way. A single Claude skill orchestrates the whole flow: it pulls the client's context from Attio, verifies the registered entity details (looking up the ABN online if it's missing, then writing it back to the CRM), creates the client in their proposal and billing platform with a duplicate check, selects the right proposal template based on the entity structure from a library of around two dozen, drafts the proposal, and pauses for human review before anything gets sent. Once the proposal is signed, the status change in Attio triggers the compliance workflow.
The same firm connected their client questionnaire tool and pre-fills onboarding questionnaires by combining what clients have already said in conversations with the financial data already in their systems. The founder's rationale stuck with us: you talk with clients constantly, there are gold nuggets in those conversations, and the last thing you want is to ask the same questions again. One recent questionnaire went out pre-filled and came back completed within a day.
That's a CRM, a billing platform and an onboarding tool coordinated by plain-language instructions in a skill file, built by an accountant, not a developer.
Does this work with ChatGPT?
Yes. ChatGPT supports MCP connectors too, and Attio's MCP server works with it, so the use cases in this guide are not Claude-exclusive. Setup is similar: add Attio as a connector, authorise via OAuth, and start asking questions.
We use Claude for this work and this guide reflects that experience. In our view Claude's skills feature, which lets you save a documented, repeatable procedure like the deal scoring and onboarding workflows above, is the strongest reason to prefer it for CRM workflows. But if your team is standardised on ChatGPT, the core capability travels: connect the MCP, get your CRM context into the conversation.
When MCP is the wrong tool
This is the section most write-ups skip, and it's where implementations go wrong.
Claude only acts when prompted. Everything above starts with a human typing an instruction. If you need something to happen on its own, on a schedule, when a field changes, or when a webhook fires, that's not a job for an MCP chat session. Use Attio's native workflows for in-app automation, and a tool like n8n for cross-system automation that runs unattended. Our rule of thumb: interactive and judgement-heavy work goes to Claude via MCP; hands-off, high-volume, deterministic work goes to workflows and n8n via the REST API.
The MCP and the API have different capabilities. A detail that catches people out: in our experience, email content is accessible through Attio's MCP but not through the public REST API. So a Claude session can read a customer's email history, while an n8n automation using the API cannot. If email context matters to your workflow, that constraint decides your architecture for you.
Unguided AI burns tokens. The MCP layer exists to help the AI work efficiently. A well-written skill with specific instructions is fast and cheap to run. Pointing Claude at a workspace with a vague request and letting it explore, run broad queries and figure things out from scratch consumes credits quickly. Specificity is an economy measure, not just a quality measure.
Three lessons from real implementations
Your data model documentation is now an AI prompt. Attio lets you write descriptions on most attributes, and Claude reads them to understand what your fields mean. One client hit a wall because relationship attributes carried no descriptions, so the AI couldn't reliably interpret how his entities connected. Treat every attribute description as an instruction to your future AI assistant, because that's what it is.
Build for the 80%. It's tempting to make a skill handle every edge case. Don't. A scenario that comes up once a year is cheaper to handle manually than to build, test and maintain automation for. Skills carry a maintenance cost that compounds with complexity.
Keep a human checkpoint on anything that leaves the building. Our client's proposal skill drafts everything but sends nothing without explicit sign-off. They learned this after a proposal once went out without review (to a forgiving recipient, fortunately). Draft with AI, review with humans, then send.
FAQs
Is it safe to connect Claude to my CRM?
The connection uses OAuth and respects your existing Attio permissions. Claude sees what you can see, nothing more. Apply the same judgement you would to any tool with CRM access, and keep human review on outbound actions.
Do I need to be technical to set this up?
No. The connection itself is a two-minute OAuth flow. The accounting firm workflow above was built by the firm's founder using plain-language skill files, not code.
What's the difference between the Attio MCP and the Attio API?
The MCP is designed for interactive AI sessions and includes access to emails and call recordings. The REST API is designed for programmatic automation via tools like n8n and, in our experience, does not expose email content. Most mature setups use both: MCP for interactive work, API for unattended automation.
