---
title: "Can Claude or ChatGPT Migrate Your CRM? MCP Limits Explained"
slug: can-claude-or-chatgpt-migrate-your-crm-mcp-limits-explained
date: 2026-11-17
author: Raajshekhar Rajan
categories: [CRM, Salesforce, HubSpot]
excerpt: "Can Claude or ChatGPT migrate your CRM? Yes for a few hundred records, no for full history, attachments or regulated data. Here are the vendor-cited MCP server limits."
tldr: "An AI agent can move a small CRM dataset through MCP. Salesforce's one-record writes, HubSpot's 10,000-result search cap and an ID map that has to live outside the chat make MCP the wrong path for a production migration."
canonical: https://clonepartner.com/blog/can-claude-or-chatgpt-migrate-your-crm-mcp-limits-explained
---

# Can Claude or ChatGPT Migrate Your CRM? MCP Limits Explained


# Can Claude or ChatGPT Migrate Your CRM? MCP Limits Explained

**Can Claude or ChatGPT migrate your CRM?** Yes, if you are moving a few hundred records, reviewing every output and treating it as a one-off. No, if you need full history, attachments, association integrity, audit-field preservation or anything touching regulated data.

The limit is not the model. It is what vendor MCP servers expose, the limits they enforce, and the controls a production migration needs around any tool that writes to your CRM. Switching from one model to another does not change what the vendor's MCP server offers.

This article walks through what the Salesforce, HubSpot and Pipedrive MCP servers can and cannot do for a migration, where the constraints start to damage data integrity, and where an [AI-assisted approach genuinely works](https://clonepartner.com/blog/blog/ai-agent-data-migration-via-mcp-the-2026-guide).

> [!NOTE]
> **Last verified: 17 November 2026.** The MCP facts in this article were checked against each vendor's own documentation on that date. MCP servers change often, so re-check the linked pages before a production cutover.

## What is a CRM MCP server?

**A CRM MCP server lets an AI client, such as Claude or ChatGPT, read and write CRM records through tool calls instead of hand-written API code.** Salesforce, HubSpot and Pipedrive each publish one. All three are designed for conversational work on a handful of records at a time, not bulk data movement.

An MCP server exposes specific tools, such as `search_crm_objects` or `createSobjectRecord`, that the agent can call on its own. That sounds like the right mechanism for a migration. You prompt the agent to "read all accounts in Salesforce and recreate them in HubSpot". The agent queries the Salesforce MCP server, holds the results in its context, and pushes them to the HubSpot MCP server. Past a few hundred records, the constraints below take over.

> [!WARNING]
> **Treat MCP as an assistant surface, not as your migration engine.** If the move includes years of activity history, attachments, deleted or archived records, inactive owners or sensitive free text, you need a deterministic loader behind the chat interface.

## MCP server capabilities: cross-platform comparison

This table summarises the migration-critical points for the three vendors. A dash means the point was not part of our 17 November 2026 check. The [MCP server limits reference](https://clonepartner.com/blog/blog/mcp-server-limits-that-break-data-migrations-2026) covers eight platforms in the same format.

| Capability | Salesforce | HubSpot | Pipedrive |
|---|---|---|---|
| **Write tools** | Create, update and delete on every standard and custom object. `createSobjectRecord` creates one record per call. | Create and update for CRM objects, activities, pipelines and custom properties. Quotes writable in beta. | `add` and `update` tools for deals, persons, organizations, leads, products, activities, notes and projects |
| **Read limits** | `soqlQuery`: 50,000 records per transaction. `find`: 2,000 records. | `get_crm_objects`: 100 IDs per request. Search: 200 per page, 10,000 total. | — |
| **Delete tool** | Yes, in `sobject-all` and a separate SObject Deletes server | No CRM delete tool | None listed |
| **Attachment or file tool** | — | None listed | None listed |
| **Upsert or deleted-record queries** | No upsert, QueryAll or undelete tool | — | — |
| **Original created date** | Only on insert, with the "Set Audit Fields upon Record Creation" setting and permission | Create date read-only, except on deals | — |
| **Bulk path outside MCP** | Bulk API 2.0 | Imports | Spreadsheet importer: XLS, XLSX or CSV, up to 200,000 rows, under 50 MB |

Sources: Salesforce ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/hosted-mcp-servers/guide/sobject-all.html)), HubSpot ([developers.hubspot.com](https://developers.hubspot.com/docs/apps/developer-platform/build-apps/integrate-with-the-remote-hubspot-mcp-server)), Pipedrive ([support.pipedrive.com](https://support.pipedrive.com/en/article/mcp-tools)).

The pattern is consistent. Every platform exposes a conversational read and write surface. The bulk, archival and idempotency features a production migration needs live on other paths.

## Salesforce hosted MCP server: where it runs out of road

The `platform/sobject-all` server exposes create, read, update, delete, query, search and relationship tools for every standard and custom object in the org ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/hosted-mcp-servers/guide/sobject-all.html)). That sounds like enough for a migration. It is not, for specific reasons:

- **One record per write.** `createSobjectRecord` creates one record per call. Migrating 50,000 leads means 50,000 tool calls. The tool fails if required fields are missing or values break validation rules, so its description tells agents to call `getObjectSchema` first.
- **SOQL reads are capped per transaction.** `soqlQuery` allows a maximum of 50,000 total records per transaction across all queries. An org with 200,000 contacts needs the agent to partition its reads.
- **SOSL search returns at most 2,000 records.** That is not enough to find and match records at migration scale.
- **No upsert tool.** Upsert by external ID is what makes a Salesforce load idempotent, and lets child records reference parents by external ID ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/api-rest/guide/dome-upsert.html)). Without it, a retry after a timeout can create a duplicate.
- **No QueryAll or undelete tool.** Salesforce's REST QueryAll resource, unlike Query, "returns records that are soft deleted due to a merge or delete", and also returns archived task and event records ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/api-rest/guide/resources-queryall.html)). The MCP tool list has no equivalent. Deleted records stay in the Recycle Bin for up to 15 days, and no undelete tool is available through MCP.
- **Cascading deletes.** Deleting a master record in a master-detail relationship also deletes all of its detail records.
- **No bulk tool.** Salesforce documents Bulk API 2.0 as the path for large loads ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/salesforce-app-limits-cheatsheet/guide/salesforce-app-limits-platform-bulkapi.html)). The MCP tools do not call it.

A capped query result is easy to mistake for a complete one. Compare record counts with the source before you trust any export.

### Preserving history and audit fields

A migration moves time and ownership, not just text.

Suppose an agent reads a Salesforce Account created in 2019 by an account executive who left in 2022, and writes it to a new org through a standard create call. The new record shows today's date as its creation date and the connected user as its creator.

Salesforce allows the audit fields `CreatedDate`, `CreatedById`, `LastModifiedDate` and `LastModifiedById` to be set only on insert. The org must enable **"Set Audit Fields upon Record Creation"**, and the loading user must have that permission. `SystemModstamp` can never be set. Keeping records owned by departed employees needs the **"Update Records with Inactive Owners"** permission ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/object-reference/guide/system-fields.html)).

These are org settings and user permissions that an administrator turns on. The hosted MCP server acts with the authenticated user's field-level security, object permissions and sharing rules, so the settings must be in place before any load. Without them, every migrated record carries today's date and the connected user, and SLA reports, time-in-stage metrics and historical quota attribution break.

This is why an agent-run Salesforce migration tends to move off hosted MCP and onto the regular APIs: upsert by external ID for safe retries, QueryAll for deleted and archived records, and Bulk API 2.0 for volume.

## HubSpot MCP server: read-only gaps and the 10,000-record wall

HubSpot's remote MCP server can write contacts, companies, deals, tickets, leads, line items, products, appointments, courses, listings, projects, services and custom objects. It can also write activities (calls, emails, meetings, notes and tasks), pipelines, custom properties and static segment membership. Quotes are writable in beta with an extra scope ([developers.hubspot.com](https://developers.hubspot.com/docs/apps/developer-platform/build-apps/integrate-with-the-remote-hubspot-mcp-server)).

The migration-specific gaps:

- **Carts, invoices, orders and subscriptions (beta), and conversations, are read-only.** If your migration includes historical billing records, the MCP server cannot write them.
- **No attachment or file tool** for CRM records. Deal contracts, signed proposals and ticket screenshots cannot travel through this path.
- **No CRM delete tool.** If a test run creates bad records, cleaning them up needs the standard API or manual deletion.
- **Sensitive Data blocks conversation data.** When the Sensitive Data setting is on, conversation data is blocked from the MCP server. This does not apply to the standard CRM APIs.

### The 10,000-record ceiling

The most dangerous constraint is the search limit. The HubSpot MCP search tool relies on the CRM search API. That search is capped at **10,000 total results** for any query and **200 objects per page**, and allows **five search requests per second** per account ([developers.hubspot.com](https://developers.hubspot.com/docs/api-reference/latest/crm/search-the-crm)).

If you ask an agent to "export every contact created since 2019", it uses the search tool. The export stops at 10,000 records, and no error is raised unless someone checks the total. Without a record-count parity check against the source, thousands of contacts can be left behind.

HubSpot also notes that newly created or updated objects can take a few moments to appear in search results. A migration that writes a record and immediately searches for it to verify the write can see a false gap.

### Deduplication, timestamps and associations

- **Create dates.** "The Create date property is read-only for all objects except deals" ([knowledge.hubspot.com](https://knowledge.hubspot.com/records/update-a-property-value-for-a-record)). You cannot backdate the creation of a contact or company, so historical records arrive looking new.
- **Deduplication.** Companies created through the API "will not be deduplicated by the Company domain name property", and import rows without an email or Record ID become new records ([knowledge.hubspot.com](https://knowledge.hubspot.com/records/deduplication-of-records)). If an agent processes the same company twice in a retry loop, you get a duplicate it may not notice.
- **Associations.** Associations need the IDs of both records, so parents must exist in HubSpot before children can be linked ([developers.hubspot.com](https://developers.hubspot.com/docs/api-reference/latest/crm/associations/associate-records/guide)). The migration must keep a source-to-destination ID map that survives retries, restarts and later delta runs.

For the full object and attachment details on this route, see [Salesforce to HubSpot migration: pipeline, tickets and attachments](https://clonepartner.com/blog/blog/salesforce-to-hubspot-migration-pipeline-tickets-attachments).

## Pipedrive MCP server: broad but record by record

Pipedrive's MCP tool list has `get`, `add` and `update` tools for deals, persons, organizations, leads, products, activities, notes and projects, plus search tools for most of them ([support.pipedrive.com](https://support.pipedrive.com/en/article/mcp-tools)). The projects tools need the Projects add-on, and the email tools need a connected email sync. The published list has no delete tool, and the page does not mention files or bulk operations.

For bulk work, Pipedrive's spreadsheet importer is the vendor-native path. It accepts XLS, XLSX or CSV files with no more than 200,000 rows and a file size under 50 MB ([support.pipedrive.com](https://support.pipedrive.com/en/article/importing-data-into-pipedrive-with-spreadsheets)). If your Pipedrive migration includes years of files, plan a separate path for them.

## Why the source-to-destination ID map belongs outside the agent's context

Every CRM migration needs a **source-to-destination ID map**: a persistent lookup from each old record ID to its new ID in the destination. Without it, associations (contact to company, deal to contact, note to deal) cannot be rebuilt.

An agent working through MCP tends to carry that map in its conversation. Two things make that fragile.

**Tool output limits.** Claude Code warns when MCP tool output exceeds 10,000 tokens and limits it to 25,000 tokens by default. Successful text results longer than 50,000 characters are saved to a file ([code.claude.com](https://code.claude.com/docs/en/mcp)). A migration's ID map grows with every record, so it soon outgrows what a conversation is designed to pass around.

**Copying and determinism.** Anthropic's engineering team writes that models "may be more likely to make mistakes when copying data between tool calls". Its recommended fix is to have the agent write code that processes data outside the model: "The agent sees five rows instead of 10,000" ([anthropic.com](https://www.anthropic.com/engineering/code-execution-with-mcp)). Anthropic's documentation also says results are not fully deterministic even at temperature 0 ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/glossary)). A field mapped one way in the test run can be mapped another way in the full run or the delta. A changed date format, picklist value or phone rule means production data no longer matches the sample the team reviewed.

That is sound advice for building a migration script. It also means migration-scale data belongs in code and storage outside the model's context. Here is our illustration of an ID ledger kept in a file:

```yaml
accounts:
  001xx000003DHP0AAO: 194533001
contacts:
  003xx000004TQ9LAAW: 90215519
deals:
  006xx00000ABCDeAAH: 118402991
```

The ledger lets the association pass use exact destination IDs instead of whatever the conversation still holds. It also makes reruns idempotent and keeps the final delta consistent with the sample your team signed off.

## Why the mapping must be frozen across the sample, full run and delta

A production CRM migration is never a single event. It has a sample run, a full historical load in the days before go-live, and a final delta just before the team starts work in the new system, to catch records users changed in the meantime.

All three runs need the same **field mapping**: the contract that says exactly how each source field lands in the destination. The mapping approved in the sample must be the mapping used for the full run and the delta.

Suppose you ask an agent to map a custom Salesforce field called `Client_Tier__c` to HubSpot for the full run, then ask again a few days later for the delta. It may map it to a different property, or format the picklist values differently. A delta with a different mapping corrupts the destination, creating stray properties or overwriting clean historical data with differently formatted values.

This is why production migration tooling **freezes the mapping** after sign-off and applies it unchanged to the sample, the full run and the delta. The mapping lives in a file or database, not in a model's context.

## GDPR and the extra processor

Sending CRM records through a model adds a party to the data flow.

- **Article 28.** A controller may use only processors "providing sufficient guarantees", under a binding contract ([gdpr-info.eu](https://gdpr-info.eu/art-28-gdpr/)).
- **Article 5.** Personal data must be limited to what is necessary, and the controller must be able to demonstrate compliance ([gdpr-info.eu](https://gdpr-info.eu/art-5-gdpr/)).

When an MCP-connected agent reads CRM records with EU personal data, such as names, email addresses and deal notes, the AI provider and any MCP server operator join the flow. A field can move from source to destination without a model reading it, so the extra processor is not needed for the job.

The plan matters too. On Anthropic's Free, Pro and Max plans, data retention extends to five years if the user allows training, and stays at 30 days otherwise. Team, Enterprise and API are excluded from that change ([anthropic.com](https://www.anthropic.com/news/updates-to-our-consumer-terms)). OpenAI says data sharing for model improvement is enabled by default on personal ChatGPT workspaces, while Business, Enterprise and Edu are excluded by default ([help.openai.com](https://help.openai.com/en/articles/8983130-what-if-i-want-to-keep-my-history-on-but-disable-model-training)). The safe setup exists, but it is a business contract plus settings, not the default tool an employee already has open.

[Is it GDPR compliant to use AI agents with customer data?](https://clonepartner.com/blog/blog/is-it-gdpr-compliant-to-use-ai-agents-with-customer-data) covers plans, transfers and BAAs in depth.

## Where AI genuinely helps in a CRM migration

None of this makes AI useless for CRM migration work. The value sits in a different place than "connect Claude to both CRMs and tell it to move everything".

The best documented example is **the dlt team's own move from HubSpot to Attio**: "One working student, Claude Code, one stakeholder call, and 2 weeks" ([dlthub.com](https://dlthub.com/blog/migrate-hubspot-attio)). The AI drafted the data model, the transforms and the tests. People reviewed every change, and the rollout went from a sandbox workspace to a dry run, then to a live sync with limits turned down.

The most useful detail is what the AI got wrong. It proposed a Slowly Changing Dimensions Type 2 design that keeps old rows as versions. That conflicted with GDPR deletion rules, because a contact deleted in the source has to be deleted in the destination too. The team caught it. As the post puts it, nothing in the schema or the HubSpot API told the AI about that rule.

The pattern that worked: **AI drafts, a person reviews, and deterministic code runs the job.**

Salesforce's own `getObjectSchema` MCP tool is another good fit. It returns picklist values and admin-written guidance that speed up a mapping draft before anyone writes load code ([developer.salesforce.com](https://developer.salesforce.com/docs/platform/hosted-mcp-servers/guide/sobject-all.html)).

For a deeper look at the risk model for AI-generated migration scripts, see [the data migration risk model](https://clonepartner.com/blog/blog/why-ai-migration-scripts-fail).

## Who owns the outcome when an agent corrupts data?

Anthropic's Commercial Terms say it is the customer's responsibility to evaluate whether outputs suit its use case, and provide the services "AS IS" ([anthropic.com](https://www.anthropic.com/legal/commercial-terms)). OpenAI's Services Agreement says "Customer is solely responsible for all use of the Outputs", also "AS IS" ([openai.com](https://openai.com/policies/services-agreement/)). These terms are reasonable for a general-purpose tool. They also mean that when an agent-run migration corrupts data, the customer owns the outcome and the clean-up.

Once users work in the destination, a bad load cannot simply be dropped and re-imported. New live records mix with migrated ones, automations have fired and emails have gone out. Most of the cost of a redo is engineering time, user disruption and lost trust in the new system.

## When to use an agent and when not to

| Scenario | Agent via MCP? | Why |
|---|---|---|
| Move 200 contacts from Pipedrive to HubSpot, one time | Workable | Small enough to review every record. No associations, no attachments. |
| Migrate 15,000 deals with multi-object associations and stage history | No | Exceeds the search cap, needs a persistent ID map and audit-field permissions |
| Move data that includes PHI or financial records | No | Needs a BAA, deterministic processing and an audit trail. The model adds a processor under GDPR Article 28. |
| Generate a first-draft field mapping for review | Yes | Schema discovery, picklist analysis and transform code are where AI helps most |
| Write and test transformation scripts an engineer will review | Yes | The dlt pattern: AI drafts, a person reviews, code runs |
| Any migration that needs a consistent delta | No | The delta must use the exact mapping reviewed in the sample |

## How ClonePartner handles what MCP servers cannot

ClonePartner runs CRM migrations as a managed, engineer-led service. A dedicated migration engineer owns each project.

The differences that map to the constraints above:

- **A persistent source-to-destination ID map** stored in a database outside any session. The same frozen mapping runs the sample, the full migration and the delta, so the production run matches what you reviewed.
- **Parent-before-child load order and idempotent loads by external ID**, so a retry updates instead of duplicating. The engine uses the history-preserving paths, such as Salesforce's audit-field permissions, and pauses destination automations before the load.
- **Multi-stage validation:** record-count parity per object, field-level checks, referential integrity checks and checksums. You get a reconciliation report, not a chat transcript.
- **Unlimited sample migrations** until the data is right, included in the fixed price, and re-migration at no additional cost.
- **SOC 2 Type II and ISO 27001**, independently audited, with GDPR, CCPA and HIPAA safeguards. Data moves through a deterministic pipeline, and fields are copied from source to destination without a language model reading customer records. NDAs and DPAs are available on request, and a BAA is signed before any PHI is handled.

The bulk historical load runs in the days before go-live while your team keeps working in the old CRM. Before the working day starts on Monday, the old system is frozen and a final delta of only the changed records takes about 30 minutes. Most migrations complete in 1 to 5 business days, depending on data volume, complexity and API rate limits. Each project is one fixed price, starting at $500.

If you want AI in the loop, use it where it is strongest: drafting mappings, transforms and tests. If you want a CRM migration that survives scale, attachments, history and Monday morning, keep the execution path deterministic and accountable.

> Book a free 30-minute scoping call. We'll map your source and target systems, flag the edge cases, and send you a fixed-price proposal, with no obligation and no hourly billing.
>
> [Talk to us](https://clonepartner.com/talk-to-us?duration=30&utm_source=blog&utm_medium=button&utm_campaign=demo_bookings&utm_content=cta_click&utm_term=demo_button_click)

## Frequently asked questions

### Can I use Claude or ChatGPT to migrate my CRM?

Yes for a small, reviewed, one-off move of a few hundred records. No for full history, attachments, associations, audit fields or regulated data. Salesforce's hosted MCP server writes one record per call and allows at most 50,000 records per SOQL transaction. HubSpot's CRM search stops at 10,000 results, and its Create date is read-only on every object except deals. A migration also needs a source-to-destination ID map that lives outside the chat session.

### What are the limits of Salesforce's MCP server for data migration?

The hosted MCP server creates one record per tool call, allows at most 50,000 records per SOQL transaction, returns at most 2,000 records from SOSL search, and has no upsert, QueryAll or undelete tool. Setting the original CreatedDate needs the org's Set Audit Fields upon Record Creation setting and the matching user permission, which an administrator grants.

### Does HubSpot's MCP server support migrating attachments?

No. HubSpot's remote MCP server lists no attachment or file tool for CRM records. It also treats carts, invoices, orders, subscriptions and conversations as read-only, has no CRM delete tool, and relies on a CRM search capped at 10,000 results per query and 200 per page.

### Why do AI-run CRM migrations break at scale?

Three main reasons. The source-to-destination ID map has to persist across thousands of writes, retries and a later delta, which a chat session is not built to hold. Model outputs are not fully deterministic, so a mapping regenerated for production can differ from the one reviewed in the sample. And MCP servers do not call the bulk endpoints, so the job becomes thousands of one-record writes.

### How much does a managed CRM migration cost with ClonePartner?

ClonePartner charges one fixed price per project, starting at $500, with no hourly billing, no hidden fees and no change orders. Most migrations complete in 1 to 5 business days, depending on data volume, complexity and API rate limits.
