Most Salesforce teams lose hours every week to the same small jobs: checking what changed in an org, preparing a release, answering "is this field used anywhere?" I built an internal assistant to take on that work, using Claude models connected to our tools through the Model Context Protocol (MCP).
What MCP actually does
MCP is an open standard that lets an AI model call tools in a structured way. Instead of pasting data into a chat, the model can ask a server to run a query, read metadata or fetch a work item, and get structured results back. A Salesforce MCP server exposes org operations; other servers can expose Azure DevOps or your release tooling.
The architecture
- Claude as the reasoning layer that plans the steps and writes the answer.
- Salesforce MCP Server for SOQL queries, metadata lookups and org details.
- Azure DevOps for work items, pull requests and pipeline status.
- Gearset for comparisons, deployment history and release packages.
The rule I followed: the assistant can read anything, but every change to an org still goes through the normal pipeline and a human approval.
What it handles today
- Impact analysis before a deployment: which components changed, and what depends on them.
- Release notes drafted from merged pull requests and linked work items.
- Quick admin answers, like who has a permission set or where a field is referenced.
Lessons learned
Start read-only. Give the model narrow, well-named tools instead of one "do anything" tool. And log every tool call, because when an answer looks wrong you want to see exactly what the model asked for.
In upcoming posts I'll go deeper into automated pull request reviews and using Claude Code for pre-deployment checks.