Does Claude keep the data it reads through a connector?
TJ Coyle, Co-founder and CEO·4 min read
TL;DR
- A connector fetches what one question needs, live, and doesn't build a store of your data for the next question.
- MCP, the protocol connectors run on, carries a request over and an answer back; its current revision removed protocol-level sessions [MCP specification, 2026].
- What a connector returns becomes part of your conversation, so Anthropic's rules for conversations apply. Those differ by plan [Anthropic Privacy Center, 2026].
- For a lookup in one app, that's exactly what you want. For questions across systems, you need data kept current and linked.
Key facts: what connectors keep
- What MCP is: An open standard, created by Anthropic, for AI apps to connect to tools and data [Anthropic Help Center, 2026]
- Protocol-level sessions: Removed from Streamable HTTP in the 2026-07-28 revision [MCP specification, 2026]
- Deleted chats on consumer plans: Removed from history immediately, from back-end storage within 30 days [Anthropic Privacy Center, 2026]
- API inputs and outputs (commercial): Deleted on the back end within 30 days, with stated exceptions [Anthropic Privacy Center, 2026]
- Enterprise custom retention: Set by owners, minimum 30 days [Anthropic Privacy Center, 2026]
The short answer
A connector doesn't keep a copy of your data to answer future questions. Each time you ask, Claude calls the connector, the connector fetches what that question needs from one app, and the answer comes back into your chat. Whatever comes back is part of the conversation, so Anthropic's retention rules for conversations apply to it. Nothing is joined to your other systems along the way.
How a connector works
You switch on the HubSpot connector and ask about a deal. Here's what happens:
- You ask Claude a question about the deal.
- Claude calls the HubSpot connector with a request.
- The connector fetches that deal from HubSpot, live.
- The result comes back into your chat, and Claude answers.
Ask again tomorrow and it fetches again.
Underneath, the connector is an MCP server. MCP is "an open standard, created by Anthropic, for AI applications to connect to tools and data" [Anthropic Help Center, 2026]. It carries a request to a tool and the result back. The protocol's current revision removed protocol-level sessions from its HTTP transport [MCP specification, 2026], and anything remembered between calls is left to whoever builds the server.
That last point matters. The connector's own operator, HubSpot or whoever runs the server, sets its own policy on what it logs. Check theirs too.
Quotable fact: a Claude connector fetches what each question needs from one app at a time; it builds no store of your data and joins nothing across systems.
What Anthropic's policies say about your conversations
Because a connector's results land in your chat, the question becomes how long Anthropic keeps chats. Anthropic publishes this, and it differs by plan:
- Consumer (Free, Pro, Max): What the policy says: Deleted chats leave your history immediately and back-end storage within 30 days. If you allow model training, de-identified data from new or resumed chats after the setting is on may be kept up to 5 years; Source: [Anthropic Privacy Center, 2026]
- Claude for Work (Team, Enterprise): What the policy says: Chats are kept in the product until you delete them; deleted chats leave back-end storage within 30 days; Source: [Anthropic Privacy Center, 2026]
- Enterprise: What the policy says: Owners can set custom retention periods, minimum 30 days; Source: [Anthropic Privacy Center, 2026]
- Anthropic API: What the policy says: Inputs and outputs deleted on the back end within 30 days, with stated exceptions; Source: [Anthropic Privacy Center, 2026]
Anthropic also keeps flagged content longer where its Usage Policy is involved [Anthropic Privacy Center, 2026]. Read the current pages for your plan rather than relying on a summary.
When a live connector is the right tool
If your question lives in one app with a good connector, use the connector. It's free, it's live, and it works.
- "What's the status of the Acme deal?"
- "Find the Slack thread about the pricing change."
- "Summarize this Notion page."
When you need data kept current
The trouble starts when the answer needs more than one system. "Which customers have an open ticket and an overdue invoice?" needs the help desk and billing. Each connector sees its own half, and nothing tells the AI that "ACME Inc." in one is "Acme" in the other.
That's where keeping data starts to matter. altspark keeps a synced copy of what you choose, current on the schedule you set, in altspark's hosted service on Azure. Your source stays the system of record.
On top of that copy, altspark links the same customer across the CRM, billing and contracts, putting any guessed match in front of you to confirm, and joins synced tables in one question. Your project's own MCP server carries all of it.
- A live lookup in one app: A connector: Yes; Data kept current (altspark): Also works
- An answer across two or more systems: A connector: Each sees its own half; Data kept current (altspark): Linked and joined
- A system with no connector, like a database on the office server: A connector: Not reachable; Data kept current (altspark): Reached read-only through altbot
- A copy kept outside the source: A connector: No; Data kept current (altspark): Yes, of what you choose

