sf-intelligence: offline Salesforce MCP server

= 20">

sf-intelligence

Know what breaks before you change Salesforce. Ask the AI assistant you already use about your org. sf-intelligence runs on your computer, never writes to your org, and every answer names the components it came from.

Website, demo and docs: [sfi.auditforce.cloud](https://sfi.auditforce.cloud) · Works in Claude Code, Claude Desktop, Cursor, VS Code + Copilot and Codex.

What it answers

The questions below were checked by an outside reviewer against a real production org (about 4,700 components) by reading the raw metadata. The answers were right.

A real answer from the built-in demo org (a solar installer called Verdant), unedited:

Can I delete `Project__c.Risk_Score__c`? Not safe to delete: blocking. ProjectTriggerHandler writes it and ProjectTriggerHandlerTest reads it (parsed from the Apex); the Residential page layout shows it (declared). Checked: Apex, flows, workflow rules, validation rules, layouts. Not checked: assignment, auto-response and escalation rules, because they were not retrieved.

Try it in three commands

npm install -g sf-intelligence

claude mcp add --scope user sf-intelligence-demo -- npx -y sf-intelligence demo # a built-in demo org, no Salesforce login; then ask your assistant a question cd your-sfdx-project && sfi init --target-org my-org-alias # then: connect your own org (read-only retrieve)

sfi demo on its own starts an MCP server and waits for a host; it is not a command to run in a terminal.

Then register the server in your AI host; the install guide has the exact block for each host. Or one click: Try the built-in demo org in Cursor in one click (no Salesforce login). AI agents installing this server: see llms-install.md.

By default your assistant sees 16 tools: the eight that answer the questions above and eight that help it find its way (sfi.resolve, sfi.search_components, sfi.get_component, sfi.health_check, sfi.list_analyses, sfi.describe_analysis, sfi.run_analysis, sfi.live_consent). Every other analysis sits behind one gateway tool, sfi.run_analysis, and is listed by sfi.list_analyses. The vault is a local folder, answers are served offline, and the optional live read-only plane is off until you turn it on. What your AI assistant does and does not see is in Privacy.

For engineers: how the vault, the dependency graph and your assistant fit together is in How it works; what is deliberately not modeled is in What it does NOT do.

Ask it like a person

You don't need to know the exact API name of anything. A typo-tolerant resolver turns messy phrasing into the right component, and when it isn't sure it asks you a clarifying question instead of guessing:

You: where's the emale field?

It: I found several matches for "email" — which did you mean?

· Email__c on Account

· Email__c on Contact

· Alternate_Email__c on Lead

You: what about the paymnet object?

It: Payment__c — one confident match.

You: is there a "warranty fee" field?

It: Nothing matched that confidently. Your vault was last refreshed on

2026-05-20 — want to pull fresh metadata from the org (/sfi-refresh) in case

it's new, or rephrase?

It never silently commits to a guess. When several components are plausible it asks which you meant; when nothing matches it offers to refresh from the org or stop. A high match score is string similarity, not proof.

How it works

One read-only refresh turns your org into a local vault: the raw metadata, one Markdown page per component, and a dependency graph. From then on every question is answered offline. Your AI assistant resolves the name you used, calls one of the 16 default tools (or sfi.run_analysis for any other analysis), and builds its answer only from what came back, citing each component.

flowchart TB

subgraph REFRESH["Refresh — once, and whenever the org changes"] direction LR ORG[("Salesforce org")] -->|"sf project retrieve (read-only)"| SRC["source/ raw metadata"] SRC -->|"extract + parse Apex / Flow / XML"| VAULT[("Local vault Markdown + DuckDB graph")] end

subgraph ASK["Ask — every question, offline on your machine"] direction TB Q["Your question (plain language)"] --> HOST["Your AI assistant (Claude, Cursor, Copilot, Codex)"] HOST -->|"1 · sfi.resolve"| RESOLVE["The component you meant (or a clarifying question)"] RESOLVE --> HOST HOST -->|"2 · a core tool, or sfi.run_analysis for the rest"| TOOLS["sfi.* analysis"] TOOLS -->|"read"| VAULT TOOLS -.->|"opt-in, read-only, capped"| LIVE[("Live org")] TOOLS -->|"cited components + confidence

• what was not checked"| HOST HOST --> ANS["Grounded answer"] end

Every box on the Ask path runs on your machine; the dotted edge to the live org is the only one that can touch Salesforce, and only after you opt in.

How it stays honest

Every answer is built from tool output, never from general Salesforce knowledge. Each relationship carries how it was found: declared (stated in the metadata), parsed (from Apex, Flow or formula source) or heuristic (pattern matching, so spot-check before acting). Each answer says which metadata families were checked and which were not, so "nothing references this" always means "no evidence in what was retrieved". A request to change the org is refused and answered with the read-only check that fits instead. For "what does this imply" questions, a deterministic, offline concept model of general Salesforce truth (never your org's data) adds cited structural claims, such as a master-detail cascade delete or two flows with an undefined order. The details: routing and refusal, architecture and the trust boundary, asking questions, ADR-001 confidence-tagged edges and ADR-008 deterministic reasoning.

What you can ask

Every area answers in plain language; ask "what can you do?" for the live map.

All of these are read-only briefings; none of them write to the org. Every component an answer names is backed by a tool call and cited with its canonical id (CustomObject:Account, CustomField:Account.Industry__c), and every relationship with its confidence (declared, parsed or heuristic). The full catalog with the live counts comes from sfi.capabilities (through sfi.run_analysis) or the tools page; the counts are pinned by eval/product-manifest.json, never typed by hand.

What it does NOT do

Boundaries are explicit, and the product tells you plainly when it hits one rather than papering over the gap with general Salesforce knowledge:

• Offline by default. Vault tools never call Salesforce mid-conversation. Run /sfi-refresh (or sfi refresh) to update metadata. The opt-in live tools (sfi.live_*) run read-only queries with strict caps and label their answers live_org. They stay off until you grant consent once per org with sfi.live_consent { grant: true } (bound to the org and user, scoped, expiring, stored locally) or set SFI_LIVE_PLANE_ENABLED=1. A per-call liveEnabled: true is intent only, not consent. Hybrid answers combine the vault and the live org and say so. The product never picks which org to query on its own.
• No record-level data. The vault stores schema and source, not rows. "How many Opportunities closed last quarter" is a question for your org directly.
• Static analysis, not runtime. Dependency edges come from metadat