An AI coding assistant writing Apex against your org has one enormous handicap: it does not know your org. It knows Salesforce in general, so it invents field names that sound right, assumes objects exist, and writes SOQL against a schema it imagined.
Model Context Protocol is the standard that fixes this. It is a common interface for exposing tools and data to an AI client, and it means an agent can ask your org a question rather than guessing at the answer.
What changes in practice
Without MCP, asking Claude Code to “add a validation rule preventing discounts over 20% on closed opportunities” produces plausible code referencing Discount__c, which may not be what your field is called.
With an org connected over MCP, the agent describes the Opportunity object first, finds the actual field, checks what automation already fires on stage change, and then writes the rule. The difference is not that the model got smarter — it is that it stopped guessing.
What the orgadmin.ai MCP server exposes
Roughly forty tools across schema, metadata, code, query, data, insights, projects, and org knowledge. In day-to-day use the ones that earn their keep are:
- Describe objects and fields, including picklist values and relationships
- Search every stored metadata type in one call
- Read Apex classes, triggers, and LWC source
- Summarise everything that fires on an object — triggers, flows, workflow rules, validation rules
- Run SELECT-only SOQL against live data
- Read the org context notes and learnings your team has already captured
Access is scoped. A personal access token carries an explicit set of scopes — org:read, org:write, deploy, data:write, files:delete, admin — so a token minted for a coding assistant can be read-only and physically unable to deploy or write records, regardless of what the agent decides to try.
Connecting a client
The server speaks streamable HTTP at a single endpoint and authenticates with a bearer token, which covers Claude Code and Codex CLI directly. For Claude Desktop there is a packaged extension bundle, since Desktop expects a stdio bridge rather than an HTTP endpoint.
The setup is the same shape in each case: mint a token with the narrowest scopes that do the job, add the server to your client’s MCP configuration, and confirm the tools appear.
Two caveats worth knowing up front
The MCP server is off unless the deployment enables it. It is not on by default, and admin-level tools sit behind a second switch again.
There is no OAuth authorisation server yet, which is what the ChatGPT and Claude web connector directories require. Token-authenticated clients — Claude Code, Claude Desktop, Codex CLI — work today; the web directories do not.
Why ground the agent at all
The failure mode of ungrounded AI on Salesforce is not that it refuses to answer. It is that it answers confidently and wrongly, in a way that looks correct until it is deployed. Every tool above exists to replace an assumption with a lookup — and on a mature org with a decade of accumulated automation, that is the entire difference between useful and dangerous.