HubSpot, Salesforce, Pipedrive, Salesloft, Outreach and Gong have all shipped servers for the Model Context Protocol (MCP), the open standard that lets AI assistants such as Claude, ChatGPT and Microsoft Copilot query outside systems. For SDR and AE teams, that moves account research, call prep and note-logging out of the CRM screen and into a chat window. The vendors' implementations differ more than the shared protocol suggests.
What MCP is
Anthropic introduced MCP in November 2024 as an open standard for connecting AI assistants to the systems where data lives (Anthropic, 2024). In December 2025 it donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation that Anthropic co-founded with Block and OpenAI (Anthropic, 2025).
A vendor runs an MCP server that publishes a list of tools, such as "search deals" or "update contact". The assistant, acting as the MCP client, reads that list, decides when to call a tool and feeds the result into the model's context. The server decides what the signed-in account can see and change.
Why sales vendors are shipping it
Anthropic's launch post described the problem MCP was meant to solve: "Every new data source requires its own custom implementation" (Anthropic, 2024). For a CRM vendor, one MCP server now covers every assistant that supports the protocol, instead of a separate integration for each.
How the main approaches differ
Read-write CRM servers. HubSpot moved its remote MCP server out of beta to general availability for all accounts on 13 April 2026, with read and write access to CRM records and activities, some objects kept read-only, and existing HubSpot user permissions applied throughout (HubSpot, 2026). Salesforce made its hosted MCP servers generally available on 29 April 2026 for Enterprise Edition orgs and above. Every transaction runs as the authenticated user, so a record the assistant updates shows that person's name in the audit trail (Salesforce, 2026). Pipedrive's server, launched on 30 June 2026, can create and update records and turn meeting notes into CRM entries, follows each user's existing permissions and records every action in Pipedrive's change logs (Pipedrive, 2026).
Read-only insight servers. Gong's server is read-only and does not create, update or delete data in Gong or the CRM. It returns AI-generated insights rather than raw transcripts through three tools, ask_account, ask_deal and generate_brief, and Gong tech admins decide which assistants can connect (Gong, 2026).
Two-way engagement platforms. Outreach runs both an MCP server and an MCP client. The server lets tools such as Claude, Microsoft Copilot and ChatGPT sequence prospects, update records and run automations in Outreach, and the client pulls outside data into Outreach's agents (destinationCRM, 2026). Salesloft announced in July 2026 that its server would add Clari Copilot call intelligence, Clari forecasting data and write-back through late summer (Clari + Salesloft, 2026). Buyers should confirm which of those are live.
Admin-governed workflow servers. Clay exposes workflows that admins package, which it calls Functions. A Function stays hidden from MCP until an admin enables it, and admins can set a workspace credit limit with per-user overrides (Clay, n.d.).
What it changes for SDR and AE work
For an SDR, the change is research without switching tabs. An assistant connected to a CRM server and an enrichment server can combine account history with fresh contact data in one answer, then draft a first message from it.
For an AE, the main uses are a brief before the call and logging afterwards. Salesforce's own example is a rep preparing for a quarterly business review who asks the assistant for account history, open opportunities, recent cases and the stakeholder map in one conversation (Salesforce, 2026). Logging notes and next steps back is a write action, which is where the controls below matter most.
What to check before rolling it out
- Whose permissions apply. Confirm whether the server acts with each rep's permissions or a shared credential. Salesforce and Pipedrive document per-user permissions. Gong offers both personal access and a shared, organisation-wide token in which every connected user sees the same data (Gong, 2026).
- Which objects are writable. HubSpot keeps some objects read-only, and Gong's server cannot write at all (HubSpot, 2026; Gong, 2026).
- Whether app roles carry over. Clay notes that a user with its Viewer role can still trigger runs through MCP, so restricting the web app does not on its own restrict the assistant (Clay, n.d.).
- Whether changes are traceable. Edits made by an assistant should land in an audit trail tied to a named user, as Salesforce and Pipedrive describe.
What commonly goes wrong
Prompt injection is a particular concern for sales data. Inbound emails, call transcripts and prospect web pages all reach the model as tool results, and the model can act on instructions hidden inside them. Developer Simon Willison set out the problem in April 2025, including "tool poisoning", a technique documented by Invariant Labs in which instructions hidden in a tool's description steer the model without the user seeing them (Willison, 2025). An assistant that reads untrusted email and can also write to the CRM gives an attacker a route to change records.
The protocol's own guidance addresses part of this. The MCP security best practices forbid "token passthrough", where a server accepts and forwards tokens that were not issued to it, and require proxy servers to obtain per-client consent to prevent "confused deputy" attacks (Model Context Protocol, 2026).
A second risk is plain bad data. An assistant that logs notes or updates fields across many accounts can create duplicates or overwrite values reps maintained by hand. Starting with read-only access, then enabling writes one object at a time, limits how far a mistake spreads.
Sources
- Anthropic. "Introducing the Model Context Protocol." 2024. https://www.anthropic.com/news/model-context-protocol
- Anthropic. "Donating the Model Context Protocol and establishing the Agentic AI Foundation." 2025. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
- HubSpot. "Remote HubSpot MCP server is now generally available." 2026. https://developers.hubspot.com/changelog/remote-hubspot-mcp-server-is-now-generally-available
- Salesforce. "Salesforce Hosted MCP Servers Are Now Generally Available." 2026. https://developer.salesforce.com/blogs/2026/04/salesforce-hosted-mcp-servers-are-now-generally-available
- Pipedrive. "Pipedrive launches native MCP server, bringing CRM workflows directly into AI assistants." 2026. https://www.pipedrive.com/en/newsroom/pipedrive-launches-native-mcp-server-bringing-crm-workflows-directly-into-ai-assistants
- Gong. "About Gong MCP server." 2026. https://help.gong.io/docs/about-gong-mcp-server
- destinationCRM. "Outreach Launches MCP Suite and Agentic Ecosystem Marketplace." 2026. https://www.destinationcrm.com/Articles/CRM-News/CRM-Across-the-Wire/Outreach-Launches-MCP-Suite-and-Agentic-Ecosystem-Marketplace-175095.aspx
- Clari + Salesloft. "Salesloft Sets the Pace on MCP, Bringing Live Revenue Data to Every Major AI Ecosystem." 2026. https://www.businesswire.com/news/home/20260709845453/en/Salesloft-Sets-the-Pace-on-MCP-Bringing-Live-Revenue-Data-to-Every-Major-AI-Ecosystem
- Clay. "MCP Security & Privacy." Undated, accessed November 2026. https://university.clay.com/docs/mcp-security-privacy
- Model Context Protocol. "Security Best Practices." 2026. https://modelcontextprotocol.io/specification/latest/basic/security_best_practices
- Simon Willison. "Model Context Protocol has prompt injection security problems." 2025. https://simonwillison.net/2025/Apr/9/mcp-prompt-injection/
