Launched:self-serve migrations intoSuperhuman Docs (Coda)
Try it now
01Agent-first
Runs where you already work
Plug it into Claude, ChatGPT or Cursor. Describe the move in plain English; the agent runs it.
02Engineer-led
Our production engine, unlocked
The pipeline our engineers use on managed enterprise migrations — the same code, now something you can drive yourself.
03Pricing
Try 10 pages free, then $1 a page
Credit-based, pay-as-you-go. No scoping call, no quote — sample it on your own docs before you spend anything.
04Sources
NotionSlabConfluenceSoonGoogle DocsSoon
Skip to content

MCP Server Limits That Break Data Migrations (2026)

A reference table, verified on 17 November 2026, of what the Salesforce, HubSpot, Zendesk and Intercom MCP servers and four other platforms can and cannot do at migration scale.

Abdul Aleem Abdul Aleem · · 11 min read
MCP Server Limits That Break Data Migrations (2026)
TALK TO AN ENGINEER

Planning a migration?

Get a free 30-min call with our engineers. We'll review your setup and map out a custom migration plan — no obligation.

Schedule a free call
  • 1,500+ migrations completed
  • Zero downtime guaranteed
  • Transparent, fixed pricing
  • Project success responsibility
  • Post-migration support included

MCP Server Limits That Break Data Migrations (2026)

Official vendor MCP servers are built for conversational work on a few records at a time, not for data migrations. If you are evaluating whether an AI agent can move your CRM or help desk data through a vendor's MCP server, the practical answer is: useful for discovery and spot checks, wrong for the bulk load.

Info

Last verified: 17 November 2026. Every row in the table below was checked against the vendor's own documentation on that date. Vendor MCP servers change often. Tools arrive, limits shift and early-access programmes graduate, so open each vendor link to confirm the current state before you decide.

Across Salesforce and HubSpot, Zendesk, Intercom and Freshdesk, Atlassian, Notion and Zapier, the pattern is the same. MCP servers expose tools for chat-driven work on a few records. Each vendor keeps a separate bulk or import path for large volumes, and that path is not what the MCP tools call.

What each platform's MCP server can do at migration scale

None of the official servers below is a full historical migration path on its own. A dash means the point was not part of our 17 November 2026 check.

Platform What the tools can write Read and rate limits Notes for a migration
Salesforce hosted MCP platform/sobject-all: create, read, update, delete, query, search and relationship tools for every standard and custom object. createSobjectRecord creates one record per call. Scoped servers: SObject Reads (query only), SObject Mutations (create and update, no delete), SObject Deletes. soqlQuery: max 50,000 total records per transaction across all queries. find (SOSL): max 2,000 records. Every operation is limited by the user's field-level security, object permissions and sharing rules. No upsert, QueryAll or undelete tool. Deleted records go to the Recycle Bin for up to 15 days. Deleting a master record also deletes its detail records. Separate production and sandbox server URLs. Large loads use Bulk API 2.0.
HubSpot remote MCP Contacts, companies, deals, tickets, leads, line items, products, appointments, courses, listings, projects, services and custom objects. Activities (calls, emails, meetings, notes, tasks). Pipelines, custom properties and static segment membership. Quotes are writable in beta with an extra scope. Read-only: carts, invoices, orders and subscriptions (beta), and conversations. get_crm_objects: max 100 IDs per request. search_crm_objects: max 200 results per page. CRM search: 10,000 total results per query, five search requests per second per account. No CRM delete tool. No attachment or file tool for CRM records. With the Sensitive Data setting on, conversation data is blocked from the MCP server (not from the standard CRM APIs). A search export stops at 10,000 records, with no error unless someone checks the total.
Zendesk Bridge (early access) An interaction layer between Zendesk and external AI agents. A catalog of more than 400 Zendesk tools. A pre-built integration is capped at 150 tools. Rate limits are layered at tenant and agent level. Numeric limits are confirmed at enrollment. Early access starts 10 November 2026, by enrollment. Bulk exports and destructive skills are not available during early access. Human approval applies where the account's own policies and thresholds require it.
Intercom MCP Add an internal note to a conversation. Create an article. Update an article. The note tool cannot send customer-visible replies. Search tools return up to 150 items per page. US and EU hosted workspaces are supported. AU hosted workspaces are not yet supported. No tool to create conversations or contacts. Articles are attributed to the admin who authorized the MCP connection, and the author cannot be set per call.
Freshdesk MCP 37 tools: 20 read and 17 write. One bulk tool, createTicketBulkUpdate, updates multiple tickets. 25 / 50 / 100 actions per minute on Growth / Pro / Enterprise. Included MCP actions per account per year: 1,200 / 6,000 / 12,000. Generally available from 10 September 2026. Authentication is by API key only.
Atlassian Rovo MCP (Jira, Confluence) Delete tools are off by default, and deletions are permanent. — Jira Cloud's documented way to keep an original created date is a CSV import through External System Import.
Notion API — About three requests per second on average. Rich-text items are capped at 2,000 characters. Notion-hosted file URLs expire after one hour.
Zapier MCP Each tool call is an operation in an app connected to Zapier. Each successful tool call consumes two Zapier tasks, a fixed rate. —

The Notion row covers Notion's public API limits, not an MCP tool list.

Several patterns stand out:

  • Search caps look like complete results. HubSpot's CRM search stops at 10,000 records for any query. An agent asked to export "every contact since 2019" through search gets 10,000, and no error is raised unless someone checks the total.
  • The core migration objects are often missing. The Intercom MCP server cannot create conversations or contacts, the records a help desk migration produces.
  • Allowances are sized for conversations. Our own arithmetic, as an illustration: a 100,000-record job needs more than eight times Freshdesk's 12,000-action Enterprise allowance for a year.

Why vendors keep separate bulk and import paths

Vendor MCP servers are designed for conversational work on a few records at a time. Each vendor keeps a separate API for moving data at volume or with its history intact.

Salesforce documents Bulk API 2.0 as the path for large loads (developer.salesforce.com). Zendesk's Ticket Import API is admin-only. It can set solved_at, updated_at and created_at on tickets and created_at on comments. Triggers do not run on imported tickets with a status other than closed, and CC'd users receive no notifications. The bulk import endpoint accepts up to 100 tickets per request (developer.zendesk.com).

A ticket created through Zendesk's normal create endpoint is stamped with the current time, so the history is re-dated to the migration day. Zendesk warns that tickets created through Create Many Tickets "may be affected by your business rules, which can include sending email notifications to your end users". It points to the bulk import endpoint for historical imports or for creating more than 1,000 tickets (developer.zendesk.com).

Import paths handle what conversational tools do not:

  • Timestamp preservation, so SLA reports, ageing views and first-contact metrics survive the move.
  • Authorship, so history does not collapse onto the connected user.
  • Quiet loading, so customers do not receive emails about issues resolved years ago. An email cannot be recalled by deleting the ticket.
  • Volume, through endpoints that take many records per request.

The generic create tools an agent reaches for do none of this. Records arrive re-dated, attributed to the connected user and treated as new activity by the destination's automations.

Scripts that bypass MCP have a deadline of their own on Zendesk. Accounts created on and after 28 July 2026 cannot create or use API tokens. From 27 October 2026 no account can create new API tokens, and on 30 April 2027 all remaining tokens are deactivated permanently. Zendesk says OAuth replaces API tokens (support.zendesk.com).

How the MCP protocol itself limits bulk jobs

The limits above are vendor-specific. The MCP protocol itself has gaps that surface at scale (modelcontextprotocol.io).

No transactions. There is no protocol-level transaction. Our illustration: a batch of 500 creates is 500 independent calls. If call 347 fails, 346 records exist and 154 do not, and recovery is entirely on the client.

No protocol sessions or stream resumability. The 2026-07-28 revision removed protocol-level sessions. Servers that need state across calls must mint their own handles and pass them as tool arguments. It also removed stream resumability. A broken response stream loses the in-flight request, and the client must re-issue it as a new request with a new ID. If the original write had already succeeded, a re-issued write without idempotency creates a duplicate.

Warning

Duplicate risk on retry. Some destination APIs accept idempotency keys. Zendesk ticket creation accepts an Idempotency-Key header, but "Keys expire after two hours", and a duplicate key sent two hours or more after the original request creates a new ticket (developer.zendesk.com). A job that pauses overnight cannot rely on those keys.

Long-running work is optional. The revision moved long-running work to an optional Tasks extension, and client support varies. Client limits add to this. Claude Code warns when MCP tool output exceeds 10,000 tokens and limits it to 25,000 tokens by default, configurable with MAX_MCP_OUTPUT_TOKENS. A per-server tool timeout is a hard limit that progress notifications do not extend (code.claude.com).

100,000 records at 100 actions per minute: our illustration

This is our own arithmetic, not a vendor figure. Take Freshdesk's Enterprise MCP allowance, the highest per-minute rate in the table.

  • Rate limit: 100 MCP actions per minute (support.freshdesk.com)
  • 100,000 records ÷ 100 actions per minute = 1,000 minutes, about 16.7 hours of continuous, error-free writes
  • Annual allowance on Enterprise: 12,000 actions, enough for 12% of the job
  • A real migration also needs reads for lookups and ID mapping, and retries, so the true action count is higher

For HubSpot, a search-based export stops at 10,000 records. For Zapier, 100,000 successful writes would consume 200,000 tasks (docs.zapier.com).

The arithmetic shows the design intent: these servers are sized for dozens of records in a conversation, not hundreds of thousands in a pipeline.

Why system fields and authorship break first

System fields such as created date, creator and closed date are where an MCP-based migration stops being faithful to history. Each platform has one path that can set them, and it is usually not the generic create tool.

HubSpot. "The Create date property is read-only for all objects except deals" (knowledge.hubspot.com). Companies created through the API "will not be deduplicated by the Company domain name property" (knowledge.hubspot.com), so a retried company create can leave duplicates.

Zendesk. The admin-only Ticket Import API sets the original dates. Ticket created_at or updated_at values before 1970 are rounded up to 1970, and a comment created_at cannot be before 1970 or in the future. Metrics and SLAs are not supported for imported tickets, and Zendesk recommends tagging them and excluding them from reports (developer.zendesk.com).

Intercom. The MCP server cannot create conversations, and articles created through it carry the authorizing admin as author. Intercom's Create Conversation API accepts created_at for migrating past conversations, outside MCP (developers.intercom.com).

Salesforce. Audit fields (CreatedDate, CreatedById, LastModifiedDate, LastModifiedById) can 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 inactive users needs the "Update Records with Inactive Owners" permission (developer.salesforce.com).

Jira. The documented way to keep an original created date is a CSV import through External System Import (support.atlassian.com).

Attachments. A Zendesk attachment upload returns a token that must be attached to a ticket within 60 minutes (developer.zendesk.com). Notion-hosted file URLs expire after one hour (developers.notion.com). An agent that copies file URLs instead of the files ships links that break.

Load order and the ID map

Associations need the IDs of both records. On HubSpot, parents must exist before children can be linked (developers.hubspot.com). A safe load runs in this order:

  1. Load parent objects first (organisations, accounts, companies) and record each new destination ID.
  2. Keep a source-to-destination ID map in a file or database, outside any chat session.
  3. Load child objects, using the map to replace source parent IDs with destination IDs.
  4. Create association records last.

On Salesforce, upsert by external ID makes the load idempotent and lets child records reference parents by external ID (developer.salesforce.com). The hosted MCP server has no upsert tool.

Security and compliance risks of migrating through a model

A migration needs read access to everything in the source and write access in the destination. Simon Willison's "lethal trifecta" is access to private data, exposure to untrusted content and a way to communicate externally (simonwillison.net). A migration agent has all three by design. It reads every record, much of the text was written by outsiders, and it holds write access to a second system.

The MCP tools specification says there should always be a human in the loop who can deny tool invocations (modelcontextprotocol.io). OWASP's LLM06:2025 Excessive Agency recommends a human approve high-impact actions (genai.owasp.org). A migration makes hundreds of thousands of writes. Nobody approves them one at a time, so approvals get switched to "always allow".

Passing every customer record through a model's context also adds a processor to a flow that does not need one. Anthropic's engineering team recommends having the agent write code that processes data outside the model: "The agent sees five rows instead of 10,000" (anthropic.com).

What MCP servers are good for in a migration project

MCP servers are strong where conversational, schema-aware access to a live platform is what you need.

Before the migration:

  • Schema discovery. Salesforce's getObjectSchema returns picklist values and admin-written guidance, and its tool description tells agents to call it before creating records (developer.salesforce.com). An agent can use it to draft a first-pass field mapping for a person to review.
  • Sample record inspection. Pull a handful of representative records to find edge cases: empty fields, unusual picklist values and odd association patterns.
  • Permission checks. Confirm the migration credential can see every object and field the job needs.

Writing transformation code. Have the agent see a few sample rows and write the transformation logic. Run that logic over the full dataset as ordinary code.

After the load. Spot-check migrated records against the source, and fix a few outliers by hand.

The pattern that works: the agent discovers and drafts, and deterministic code outside the model's context moves the data. The mistake is treating the MCP surface as the bulk mover.

For a broader look at preparing data for AI features in the destination, see CRM and help desk data prep for AI agents.

When to use a purpose-built migration

Use an MCP server when an agent needs to work with a handful of records in a live platform. Use a dedicated migration approach when any of these apply:

  • Volume exceeds a few hundred records. The caps and allowances in the table make larger jobs impractical through MCP.
  • Original timestamps must survive. SLA reports, ageing views and first-contact metrics need the platform's import path, not the generic create tool.
  • Authorship matters. Records created through normal endpoints are attributed to the connected user.
  • You cannot afford duplicates. Without protocol-level transactions, a retry after a network failure can create a duplicate unless the load is idempotent.
  • Triggers and notifications must stay silent. Records created through normal endpoints can fire automations, start timers and email customers.
  • Compliance needs a chain of custody. A model in the data path is one more processor to contract with and document.

Once users work in the destination, a bad load cannot simply be dropped and re-run. New live records mix with migrated ones, IDs have been linked and automations have fired. A redo means finding the bad records, fixing them without touching live work, re-importing and re-checking.

Your options are to write your own import scripts against each vendor's bulk and import APIs, use a purpose-built migration tool, or work with an engineer-led service that already knows which API path each platform requires.

ClonePartner uses AI where it helps, for schema discovery, mapping drafts and transformation code. The data itself moves through a deterministic, validated pipeline against the import APIs described above. A dedicated migration engineer runs sample migrations on your real records for review, then freezes the mapping before the production run. The bulk load runs in the days before go-live while your team keeps working. Before the working day starts on Monday, the old system is frozen and a final delta of changed records takes about 30 minutes. Most migrations complete in 1 to 5 business days, and each project is one fixed price, starting at $500.

Frequently Asked Questions

Can I use a vendor MCP server to migrate data between platforms?
You can create records one at a time, but vendor MCP servers are built for conversational work on a few records. Every vendor keeps a separate bulk or import path, such as Salesforce Bulk API 2.0 or Zendesk's Ticket Import API, and the MCP tools do not call it. Beyond a few hundred records, search caps, rate limits and action allowances make MCP impractical.
Can an AI agent preserve original created dates during a migration?
Not through most generic create tools. HubSpot's Create date property is read-only for all objects except deals. Zendesk's normal create endpoint stamps the current time, and original dates need the Ticket Import API. The Intercom MCP server has no conversation-create tool. Salesforce allows audit fields only on insert, and only when the org enables Set Audit Fields upon Record Creation and the loading user has that permission.
What happens if an MCP connection drops during a bulk write?
The 2026-07-28 MCP revision removed stream resumability. A broken response stream loses the in-flight request, and the client must re-issue it as a new request with a new ID. If the original write had already succeeded, the retry creates a duplicate unless the destination API accepts an idempotency key and the client uses it.
Are there rate limits on Freshdesk's MCP integration?
Yes. Freshdesk allows 25, 50 or 100 MCP actions per minute on Growth, Pro and Enterprise, with 1,200, 6,000 and 12,000 included actions per account per year. Our own arithmetic, as an illustration, is that 100,000 one-action writes at 100 a minute take at least 1,000 minutes, and the Enterprise allowance covers 12,000 of them.
What are MCP servers actually good for in a migration?
Schema discovery, first-draft field mappings, sample inspection, spot checks and writing transformation code. Salesforce's getObjectSchema tool returns picklist values and admin-written field guidance. Use MCP for discovery and QA, and run the data move through deterministic code against the vendor's bulk or import APIs.

More from our Blog

AI Agent Data Migration via MCP: The 2026 Guide
CRM/Help Desk/Migration Guide

AI Agent Data Migration via MCP: The 2026 Guide

AI agents are useful for drafting migration mappings and code. Vendor MCP servers are built for a few records at a time, so the bulk load needs a migration engine and someone accountable for the result.

Nachi Raman Nachi Raman · · 12 min read