# sf-intelligence [Health: Active]

**Category:** 👤 Customer Data Platforms  
**Repository:** https://github.com/PranavNagrecha/Salesforce-Intelligence  
**GitHub Stars:** 3  
**npm Downloads (last month):** 2183  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/sf-intelligence

## Description
Offline, read-only MCP knowledge base for one Salesforce org's metadata, deps & impact.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "sf-intelligence": {
    "command": "npx",
    "args": ["-y","sf-intelligence"]
  }
}
```

## Documentation

## What sf-intelligence does

sf-intelligence gives AI assistants a local view of one Salesforce org's metadata, dependencies, and impact relationships. It is designed for questions where changing or removing a component could have consequences, such as finding every reference to a field, checking whether a field appears safe to delete, or determining what runs when a record is saved.

The server is read-only. It does not make changes to the Salesforce org, and its normal question-answering path works from a local vault rather than making network requests. Answers identify the components used as evidence and distinguish relationships found directly in metadata from those parsed from source or inferred by pattern matching.

## How it works

A refresh retrieves Salesforce metadata through a read-only Salesforce project operation. The retrieved source is converted into a local vault containing raw metadata, one Markdown page per component, and a dependency graph backed by DuckDB. Subsequent questions are handled on the local machine.

An AI host first resolves the component referred to in a natural-language question. The resolver tolerates misspellings and asks for clarification when multiple components could match instead of silently choosing one. The host then calls a focused analysis tool or routes the request through the general analysis gateway. Results include cited components, confidence information, and the metadata families that were or were not checked.

An optional live plane can be enabled for read-only access to the org. It is off by default, and the local vault remains the source for offline analysis. A deterministic concept model can add structural Salesforce explanations, such as master-detail cascade behavior, without using org data.

## Setup and configuration

Install the package globally with npm on Node.js 20 or newer:

```bash
npm install -g sf-intelligence
```

To explore the built-in demo, register the demo command with an MCP host:

```bash
claude mcp add --scope user sf-intelligence-demo -- npx -y sf-intelligence demo
```

The demo uses a bundled example org and does not require Salesforce authentication. Running `sfi demo` directly starts an MCP server and waits for an MCP host; it is not intended as a standalone terminal command.

For an actual Salesforce project, run the initialization command from the project directory and provide the target-org alias:

```bash
cd your-sfdx-project
sfi init --target-org my-org-alias
```

The server can then be registered with an MCP-compatible host such as Claude Desktop, Claude Code, Cursor, VS Code with Copilot, or Codex.

## Tools and capabilities

The default MCP surface contains 16 tools. Eight support the main analysis questions, while the remaining tools help the assistant locate components and manage analyses. Named utility tools include:

- `sfi.resolve` for matching natural-language references to components
- `sfi.search_components` and `sfi.get_component` for vault lookup
- `sfi.health_check` for checking server or vault state
- `sfi.list_analyses` and `sfi.describe_analysis` for discovering available analyses
- `sfi.run_analysis` as the gateway for analyses beyond the default set
- `sfi.live_consent` for enabling the optional live, read-only plane

Supported analysis areas include field and component usage, deletion-impact verdicts, record-save automation order, unused Apex detection, and effective visibility or edit permissions. The server also supports finding objects and fields, inspecting validation rules, and understanding metadata relationships through plain-language questions.

## Limitations and notes

Results describe evidence in the metadata that was retrieved. A statement that no reference was found is not proof that no reference exists outside the checked metadata families. Each relationship is labeled as declared, parsed, or heuristic; heuristic matches should be reviewed before taking action.

The deletion-impact analysis reports what it checked and what it did not check. The read-only design means requests to modify the org are refused rather than executed. The local vault must be refreshed when the Salesforce org changes, and live access remains opt-in. General Salesforce concepts supplied by the deterministic model are structural explanations, not data from the connected org.

_Full upstream README: https://allmcps.com/mcp/sf-intelligence/readme_