Salesforce orgs diverge hard — custom objects, renamed fields, per-company conventions — and a generic connector guesses. Running the learn step once means later queries use names that exist. The backup and time-machine pair is the other reason to look: it is unusual for an MCP server to give you a recoverable historical view rather than only the current record.
An MCP server for Salesforce that adapts to whatever org you connect: it discovers your schema dynamically, so custom objects and fields work like standard ones. Beyond CRUD it backs up data and files, and can read an earlier backup as it stood at a point in time.
- Records — `salesforce_query`, `salesforce_create`, `salesforce_update`, `salesforce_delete`
- Schema — `salesforce_describe` for object and field detail, and `salesforce_learn`, which analyses your complete Salesforce setup so later answers use your org's real objects and field names
- Backup — `salesforce_backup` covers data and files including ContentVersions, Attachments and Documents with their metadata; `salesforce_backup_list` shows what exists; `salesforce_time_machine` reads a point in time out of them
- Setup and diagnostics — `salesforce_auth` and `salesforce_installation_info`
- Authentication is browser-based OAuth with automatic token refresh, and Claude detects when it is needed rather than making you run a terminal step
Node.js 18+ and a Salesforce Connected App with OAuth enabled. Set the callback URL to `http://localhost:9876/callback` and select the api and refresh_token/offline_access scopes; keep the Consumer Key and Consumer Secret. The recommended install is npx — put `npx @aiondadotcom/mcp-salesforce` in your client config and use the setup tool on first run to supply client ID, client secret and your instance URL. Credentials are written to `~/.mcp-salesforce.json` with 600 permissions and the learned schema cache to `~/.mcp-salesforce-cache/`. Package version 1.2.5.
One command plus a key — npx @aiondadotcom/mcp-salesforce, then supply credentials