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.
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
AI Agent Data Migration via MCP: The 2026 Guide
Can Claude or ChatGPT, connected to your CRM or help desk through MCP servers, migrate your data? They can help you plan it. They are the wrong engine for the load itself.
The models are capable. They draft field mappings, write transformation code and check samples well. The difficulty is the job. A data migration is a bulk, one-time and largely irreversible move of every record, and vendor MCP servers are built for conversational work on a few records at a time. What an agent-run migration lacks is a migration engine and someone accountable for the result.
This guide is the hub for our series on AI agents and data migration. Each section summarises one way an agent-run migration goes wrong, and links to the post that covers it in depth.
Last verified: 17 November 2026. The MCP server facts in this guide were checked against each vendor's own documentation on that date. Vendor MCP servers change often, so check the linked pages before you plan a migration.
What is MCP data migration?
MCP data migration means connecting an AI agent to the source and destination systems through Model Context Protocol servers, and having it read records from one and write them into the other. MCP gives the agent a set of tools and a way to call them. It does not give it a bulk loader, an ID map or a transaction.
The protocol itself reflects that. The 2026-07-28 revision removed protocol-level sessions and stream resumability, and there is no protocol-level transaction (modelcontextprotocol.io). If a response stream breaks, the client re-issues the request as a new request. If the first write had already succeeded, a retry without idempotency creates a duplicate.
Three execution paths get confused in planning:
- MCP tool calls. The agent calls a vendor's MCP tools, which are designed for a few records per conversation. Salesforce's
createSobjectRecord, for example, creates one record per call (developer.salesforce.com). - Vendor bulk and import APIs. Separate endpoints for volume and history, such as Salesforce Bulk API 2.0 (developer.salesforce.com) and Zendesk's Ticket Import API (developer.zendesk.com). The MCP tools do not call these paths.
- A migration engine. Deterministic code that calls the bulk and import APIs, keeps an ID map outside any chat session, retries safely and ends with a reconciliation report.
Most agent-run migration problems come from using the first path for work that needs the second and third.
Four routes to migrate business data in 2026
The routes differ in who owns the execution engine and who owns the outcome.
| Route | How it works | Suits | Where it runs out of room |
|---|---|---|---|
| Agent + vendor MCP servers | Claude or ChatGPT calls vendor MCP tools to read and write records | Lookups, schema discovery, a few hundred records reviewed by hand | One-record writes, search caps, action allowances, no import endpoints |
| Agent-written script | An AI coding agent writes a script against the vendor's REST, bulk or import APIs, and an engineer runs it | Engineers who review generated code and test it on a sample | Retry safety, edge cases and accountability rest on whoever ran it |
| Self-serve migration tool | A product with pre-built connectors for supported pairs, run by the customer | A listed source and destination pair | Pairs, custom objects or compliance needs outside the tool's scope |
| Engineer-led managed migration | A migration engineer scopes, maps, samples, runs and reconciles the move | Custom fields, relational records, regulated data, unlisted pairs | A fixed project fee instead of your own engineering time |
AI agent vs migration tool breaks down the ten components a migration engine has, and what happens when a general-purpose agent works without them.
Where an agent-run migration goes wrong
None of these problems is about model quality. They come from API design, protocol limits and the data itself.
| Failure family | What happens | Covered in depth |
|---|---|---|
| Volume | Search caps stop exports early. One-record writes and action allowances stretch a load over days. | MCP server limits |
| Lost history | Records carry the migration date and the connected user. | Keeping created dates |
| Side effects | Triggers fire, and customers get emails about old issues. | Importing without notifications |
| Run mechanics | Retries create duplicates. Mappings drift between runs. | AI agent vs migration tool |
| Safety controls | Approvals get switched to "always allow". | Is it safe? |
| Data protection | Customer records pass through extra processors. | GDPR and AI agents |
| Accountability | The AI service terms leave the outcome with you. | Who is responsible? |
1. Volume: a conversational pipe for a bulk job
Vendor MCP servers are designed for a person asking about a few records, and each vendor keeps a separate path for volume. Salesforce's hosted MCP server creates one record per call (developer.salesforce.com). The HubSpot CRM search that the MCP search tool relies on stops at 10,000 total results for any query, and no error is raised unless someone checks the total (developers.hubspot.com). Freshdesk's MCP integration allows 100 actions a minute and 12,000 included actions a year on Enterprise (support.freshdesk.com).
Our own arithmetic, as an illustration: at 100 actions a minute, writing 100,000 records one per action takes at least 1,000 minutes, and the year's allowance runs out after 12,000 of them. MCP server limits that break data migrations has the full table for eight platforms.
2. Lost history: created dates and authors
Most platforms treat created date, author and closed date as system fields. A record created through the normal endpoint carries the migration date and the connected user. Zendesk sets original dates only through its admin-only Ticket Import API (developer.zendesk.com). 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 (developer.salesforce.com). HubSpot's Create date property is read-only for every object except deals (knowledge.hubspot.com).
Keep created date when migrating covers the history-preserving path on each platform.
3. Side effects: triggers and notifications
Records created through normal endpoints behave like new activity. 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" (developer.zendesk.com). Its Ticket Import API does not run triggers on imported tickets with a status other than closed, and CC'd users get no notifications (developer.zendesk.com). An email cannot be recalled by deleting the ticket.
How to import tickets without sending notifications lists what to pause on each platform before a load.
4. Run mechanics: duplicates and mapping drift
A migration runs more than once: a sample, the full load and a final delta. Without protocol-level transactions, a retried write can create a duplicate unless the destination accepts an idempotency key. Zendesk ticket creation accepts one, but "Keys expire after two hours" (developer.zendesk.com).
The field mapping also has to stay the same across runs. Anthropic's documentation says results are not fully deterministic even at temperature 0 (platform.claude.com). A mapping regenerated from a prompt for the delta can differ from the one reviewed in the sample. A migration engine freezes the mapping in a file and keeps the source-to-destination ID map in a database. The migration engine breakdown walks through each of these components.
5. Safety controls that cannot scale
Every AI vendor's safety guidance puts a person in front of consequential actions. OpenAI advises to "always enable tool approvals so end users can review and confirm every operation" when using MCP tools (developers.openai.com). The MCP tools specification says there should always be a human in the loop who can deny tool invocations (modelcontextprotocol.io). A migration makes hundreds of thousands of writes, which nobody can approve one at a time. In practice, approvals get switched to "always allow".
Is it safe to use an AI agent to migrate data? maps the advice of Anthropic, OpenAI, Microsoft and Google to the migration control that satisfies each point.
6. Security and data protection
A migration agent holds what Simon Willison calls the "lethal trifecta": access to private data, exposure to untrusted content and a way to send data out (simonwillison.net). It reads every record, much of that text was written by outsiders, and it holds write access to a second system. Research demonstrations, not customer breaches, show the risk. Invariant Labs used a malicious GitHub issue to make an agent copy private repository data into a public pull request (invariantlabs.ai). General Analysis used a support ticket to make an AI coding assistant with a privileged Supabase MCP connection expose a secrets table (generalanalysis.com).
Sending every customer record through a model's context also adds a processor under GDPR Article 28 (gdpr-info.eu). For health data, Anthropic's BAA page says data sent to third parties through MCP and connectors "isn't covered under Anthropic's BAA" (privacy.claude.com). Is it GDPR compliant to use AI agents with customer data? covers plans, contracts and BAAs.
7. Accountability: who owns the result
Anthropic's Commercial Terms provide the services "AS IS" and exclude liability for lost data (anthropic.com). OpenAI's Services Agreement says "Customer is solely responsible for all use of the Outputs" (openai.com). 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.
AI agent deleted your data: who is responsible? covers what that means in practice.
If an agent migration has already gone wrong
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 issued and linked, and automations have fired. On Salesforce, deleted records stay in the Recycle Bin for up to 15 days, and the hosted MCP server has no undelete tool (developer.salesforce.com).
AI migration went wrong? The engineer's triage guide covers how to stop writes, take stock and repair.
By platform
CRM. Can Claude or ChatGPT migrate your CRM? goes through the Salesforce, HubSpot and Pipedrive MCP servers, and why the source-to-destination ID map belongs outside the chat.
Help desk. Can an AI agent migrate your help desk? covers Zendesk, Intercom and Freshdesk. For example, the Intercom MCP server has no tool to create conversations or contacts (developers.intercom.com).
Where AI genuinely helps in a migration
AI is a strong drafting tool inside a human-reviewed process:
- Schema discovery and first-draft field mappings. Salesforce's hosted MCP
getObjectSchematool returns picklist values and admin-written field guidance (developer.salesforce.com). - Transformation code and tests that run outside the model. The dlt team migrated its own CRM from HubSpot to Attio with an AI drafting the data model, transforms and tests, while humans reviewed every change and ran a staged rollout (dlthub.com). The AI proposed a design that conflicted with GDPR deletion rules, and a human caught it.
- Generate, run and score loops on real records. Quandri's LLM wrote data adapters over about 40 iterations, with a domain expert reviewing (quandri.io). The team notes that a score cannot see a value placed in the wrong field.
- Reviewable mapping files with a dry-run mode. Directus generated mapping files marked as auto-generated, and a human approved them before a script applied them (directus.com).
- Spot checks on small samples, knowledge-base rebuild drafts and duplicate suggestions.
The common shape of every published success: AI drafts, a person reviews, and deterministic code runs the job. Anthropic's engineering team recommends the same split. It suggests that agents process large data in code outside the model, because "models may be more likely to make mistakes when copying data between tool calls" (anthropic.com).
How to choose a route
| Situation | Agent + MCP | Agent-written script | Self-serve tool | Managed migration |
|---|---|---|---|---|
| A few hundred records, no history needed | Workable with review | Workable | Workable for a listed pair | More than you need |
| Thousands of records or more | Search caps and action limits apply | Needs bulk APIs, safe retries and an ID map | Workable for a listed pair | Workable |
| Original dates and authors required | Generic create tools stamp today's date | Must use each platform's import path | Depends on the tool | Uses the import paths |
| Regulated data (PHI, EU personal data) | Check BAA and processor coverage | You own compliance | Check the vendor's certifications | DPA on request, BAA before any PHI |
| Who owns the outcome | You | You | You, with the vendor's tool | A dedicated migration engineer |
Five questions drive the decision:
- Record volume. How many records, across how many objects? The table in the MCP limits reference shows where each vendor's caps start.
- History. Do original created dates, authors and closed dates matter for SLA reports and audits? If yes, the load must use each platform's import path.
- Regulated data. Does the data include PHI or EU personal data? If yes, every processor in the chain needs a contract, and PHI needs a BAA.
- Downtime. Can your team stop working during the move? If not, you need a delta. Load the history first while the team keeps working, then freeze the old system and move only the records that changed.
- Ownership. Who pays for a redo if the load goes wrong? Decide that before the first write, not after.
How ClonePartner runs a migration
ClonePartner has run 1,500+ successful migrations and built 500+ integrations. There are two ways to buy, and you choose.
ClonePartner Express is a self-serve migration tool you run in the browser. You connect your accounts, run a free sample on your real data before spending anything, and then run the full migration. It covers four destinations today, each with its own source list.
The managed migration is for everything else. A dedicated migration engineer owns the project from scoping to cutover. Engineers use AI for the drafting work described above. The load itself runs through a deterministic pipeline, and fields are copied from source to destination without a language model reading customer records. The mapping is frozen after the sample is reviewed, and the same mapping runs the full migration and the delta. Validation covers record-count parity per object, field-level checks, referential integrity and checksums, and ends in a reconciliation report.
| ClonePartner Express | Managed migration | |
|---|---|---|
| Who runs it | You, in the browser | A dedicated migration engineer |
| Pairs | Gorgias (from 17 help desks), Front (from Intercom or Zendesk), Superhuman Docs (from Notion or Slab), Google Docs (from Confluence or Quip) | Any source and destination |
| Before you pay | A free sample on your real data: 100 tickets, 10 pages or 10 documents | A free 30-minute scoping call and a fixed-price proposal |
| Samples | The free sample | Unlimited sample migrations, included in the price |
| Price | Shown on each Express site | One fixed price per project, from $500, with no hourly billing, hidden fees or change orders |
| Suits | A listed pair you want to run yourself | Custom fields, relational records, permissions and compliance requirements |
For a managed migration, the bulk historical 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 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. Re-migration costs nothing extra.
ClonePartner holds SOC 2 Type II and ISO 27001, and applies GDPR, CCPA and HIPAA safeguards. Data is encrypted with AES-256 at rest and TLS 1.2 or higher in transit. NDAs and DPAs are available on request, a BAA is signed before any PHI is handled, and staging databases are purged within 30 days of project close (trust.clonepartner.com).
Two ways to run it. If your source and destination are a pair ClonePartner Express supports, you can run the migration yourself: Gorgias, Front, Superhuman Docs or Google Docs, each with a free sample on your real data first. For any other pair, or custom fields, relational records, permissions or compliance requirements, talk to a migration engineer about a fixed-price managed migration.
The practical answer for 2026: use AI for discovery, mapping drafts, code, tests and spot checks. Run the load itself through deterministic code, with a person accountable for the result.
Frequently Asked Questions
- Can I migrate CRM data with Claude or ChatGPT using MCP?
- You can move a small number of records and review each one by hand. A full migration runs into vendor limits. Salesforce's hosted MCP server creates one record per call, and the HubSpot CRM search behind its MCP search tool stops at 10,000 results for any query. Original created dates need each platform's own import path, which generic create tools do not use. Use the agent to draft the mapping, and run the load through the vendor's bulk and import APIs.
- What is MCP data migration?
- MCP data migration means connecting an AI agent to the source and destination systems through Model Context Protocol servers, and having it read records from one and write them into the other. MCP gives the agent tools to call. It does not give it a bulk loader, an ID map or a transaction, so it is a tool layer, not a migration engine.
- Is it safe to use AI agents for HIPAA-regulated data migration?
- Check your BAA coverage first. Anthropic's BAA page says data sent to third parties through MCP and connectors is not covered under its BAA. OpenAI does not offer a BAA for ChatGPT Business. A deterministic pipeline that copies fields without a model reading them keeps PHI out of the model's context.
- Where does AI actually help in a data migration?
- AI helps inside a human-reviewed process. It drafts field mappings from schema discovery, writes transformation code and tests, runs generate, run and score loops on sample data, and suggests duplicates. The pattern that works is that AI drafts, a person reviews, and deterministic code runs the job.
- How long does a managed data migration take?
- Most ClonePartner migrations complete in 1 to 5 business days, depending on data volume, complexity and API rate limits. For comparison, here is our own illustration of the agent route. At Freshdesk's Enterprise MCP rate of 100 actions a minute, writing 100,000 records one per action takes at least 1,000 minutes before any retries.