Salesforce knowledge base — FAQ

The questions people ask before they point sf-intelligence at a production Salesforce org.

Does it write to or change my Salesforce org?

No. It is read-only and has no write path. The first refresh runs sf project retrieve, which only reads metadata. No tool deploys, edits, or deletes anything in your org.

Does it send my org data to the cloud or to an AI provider?

Vault answers make no network calls. The metadata index lives on your machine in an org-kb/ directory, and the MCP server reads those local files to answer. The only times it talks to Salesforce are during a refresh (to retrieve metadata) or when you explicitly enable the opt-in, read-only live plane for record counts and samples. What you index stays yours.

How is this different from asking ChatGPT or Claude about Salesforce?

A general model answers from generic Salesforce training data, which is often wrong for your org. sf-intelligence answers only from the metadata you actually retrieved, cites every component with its canonical ID (like CustomField:Account.Industry__c), and asks a clarifying question when a name is ambiguous instead of guessing.

What is the difference between the offline and live planes?

The offline plane answers from your last refresh's local vault — schema, dependencies, permissions, source — and is the default. The live plane runs capped, read-only SOQL against the connected org for record-level facts like counts and samples, and is off until you grant consent once per org. Hybrid answers fuse both and disclose each provenance. Live answers never silently backfill a stale vault claim.

Can it tell me how many records are in an object?

Only with the live plane enabled. The offline vault stores schema and source, not rows, so record-level questions need a read-only live query (sfi.live_count) or a direct sf data query against the org.

How much does it cost?

Free and open source under the Apache-2.0 license, distributed on npm as sf-intelligence.

What do I need to run it?

Node.js 20+ and an authenticated Salesforce CLI (sf) pointed at the org you want to index. It runs in Claude Code, Claude Desktop, or any MCP client. See the getting-started guide.

Are its answers always correct?

Answers are grounded in real metadata, but each is tagged so you know how much to trust it. Confidence is declared, parsed, or heuristic — heuristic findings from the regex Apex scanner can have false positives and should be spot-checked. It also performs static analysis, so dynamic SOQL and reflective Apex are invisible: a “no references found” result means “no static evidence”, not “definitely unused”.

How do I keep the answers up to date?

Run sfi refresh (or /sfi-refresh in Claude Code) after you deploy changes. sfi.health_check reports whether the vault is healthy, degraded, or unhealthy and tells you when a rebuild is due.

Still have a question?

The getting-started guide covers setup end to end, and the capability map shows everything it can answer.