orgadmin.ai

MCP Server

A Salesforce MCP server for the AI client you already use

Paste one URL into ChatGPT or Claude, sign in, and your existing AI client can describe objects, read Apex, run SELECT-only SOQL and audit the org — through the same tenant checks, workspace roles and audit log as the web app. The reasoning runs on your subscription, not on tokens metered here.

Start free trialSee pricing
You want your own AI client working against the org without handing it the keys to change anything. A token mid-mint with one scope ticked — read org metadata and data — and deploy, record DML and file deletion left off and marked Destructive. Scopes decide which tools are registered at all, so a client holding this token never sees a write tool in tools/list. The command below is the whole of the Claude Code setup.

What you can do

  • Add it to ChatGPT or Claude as a custom connector by pasting one URL, then sign in with OAuth — no token to copy
  • Set each org to full access, read only or off per connection — read production while the agent builds in a sandbox
  • Connect any other MCP client that speaks streamable HTTP — Claude Code, Codex CLI, Cursor and others — by signing in or with a scoped personal token
  • Install into Claude Desktop from a generated .mcpb bundle, or bridge any stdio-only client with a dependency-free script
  • Describe objects and fields, search metadata, read Apex, LWC and flow source, and run SELECT-only SOQL
  • Read debug logs and past deployment failures so the agent fixes the actual error rather than guessing
  • Drive a full 14-module security review or a tech debt sweep from the client, landing in the same records the in-app features write
  • Build Development Studio projects from the client: create, pull the components a change touches, read, edit and version them
  • Read large files without truncation — tools window by line range, search without reading, and append in guarded parts

How it works

1

Paste the server URL

Add it to ChatGPT or Claude as a custom connector, or to Claude Code with one command. Settings → MCP access has it with a copy button.

2

Sign in and choose scopes

The app opens an orgadmin.ai consent screen. Read-only is ticked by default; anything that changes Salesforce you tick deliberately. Clients that cannot sign in use a scoped personal token instead.

3

Ask from your own client

Your agent gets the org tools alongside everything else it can already do — your repo, your terminal, your files.

4

Revoke whenever

Every call re-reads the owner from the database, so disconnecting an app, revoking a token or deactivating an account cuts the client off mid-conversation.

Step by step: Connect your AI client to your orgs (MCP) · Connect ChatGPT, Claude, Claude Code & Codex

Why it's different

Scopes hide tools, not just block them

A token’s scopes decide which tools are registered at all, so tools/list shrinks with the grant — an agent holding a read-only token cannot see a write tool, let alone call one. Anything genuinely destructive sits behind a further gate: deploy and record DML still require the server-side write flag and an owner or admin role, exactly as they do in the web app. And because the reasoning happens in your client, a review driven over MCP spends none of your plan’s AI allowance.

Try asking

  • “Why did last night’s deployment fail?” — from the terminal, with your repo open
  • “Run a security review on the sandbox and summarise the criticals.”
  • “Which Apex references Discount__c? Pull those classes into a project and fix the rounding bug.”
  • “Explain this debug log and tell me which SOQL is slowest.”
  • “Draft a permission set that matches what this profile actually uses.”

MCP Server — questions

ChatGPT and Claude (web, desktop and mobile) as custom connectors — paste the URL and sign in through OAuth. Anything else that speaks MCP over streamable HTTP works too, Claude Code, Codex CLI and Cursor among them, by signing in or with a personal token. Claude Desktop can also install from a generated .mcpb bundle, and any stdio-only client can use the included bridge script. It is not listed in the ChatGPT or Claude app directories; you add it yourself by URL.

Salesforce’s hosted servers expose CRM record operations — create, update, delete and query across sObjects — activated per org in Setup. This exposes the investigation surface instead: metadata, Apex and LWC source, automation summaries, debug logs, deployment failures, plus the security review and tech debt engines. A new connection here is read-only by default, and the two can sit side by side in the same client.

No. The model doing the thinking is whichever one your client runs, billed by your own subscription. Runs driven over MCP spend no allowance here, which is why a bring-your-own-key plan plus MCP is the cheapest way to use this product.

Only if you grant a scope that allows it, and only when the server-side gates are open. The default grant is read-only. Deploy and validate are deliberately left out of the tool catalog entirely — an agent can read why a deployment failed and edit the fix, but a human ships it.

Works even better with

AI Org Chat

Ask what fires on Closed Won. It checks the triggers, the flows, and the Process Builder nobody deleted.

Learn more

Development Studio

Writes Apex, LWC and Flows, reads the compile errors, tries again.

Learn more

Security Review

Fourteen modules, severity-ranked, exportable.

Learn more

Replacing something?

  • orgadmin.ai vs Agentforce

    Salesforce’s platform for customer-facing AI agents. Compare when what you wanted was AI for the people who run the org.

Read next

Put MCP Server to work on your org

Every AI feature on your own provider key, from A$14.99 per user per month.

Start free trialSee pricing
  • 14-day free trial
  • No credit card required
  • Cancel any time