sf-intelligence

A grounded, fail-closed backend for AI assistants working in one Salesforce org — answers come from the org's real metadata, not a guess.

sf-intelligence is an offline-first, read-only, MCP-first knowledge base for a single Salesforce org. You run one sf project retrieve; it builds a local Markdown vault and a DuckDB dependency graph, then answers questions locally through an MCP server (the sfi.* tools) — no network egress for vault answers. It is not a standalone chatbot: a semantic router advises — it turns each plain-language question into a meaning-ranked shortlist of the sfi.* tools that can answer it, tagged with the plane it needs (offline vault / opt-in live / hybrid) and a confidence band — and your host LLM decides which tools to run. It fails closed: write imperatives, prompt injection, and record-value exfiltration are refused by shape (with a read-only alternative offered), unanswerable asks get an honest gap instead of a lookalike tool, and genuine ambiguity gets a clarifying question instead of a guess. Terse follow-ups resolve through an optional host-passed context.previous param — the server itself stores no conversation state. An opt-in live read-only plane can answer record counts and samples. MIT + Commons Clause.

Upgrading to 0.3.0 (breaking)

0.3.0 changes defaults. If you are coming from 0.2.x, read this before upgrading — full detail in CHANGELOG.md.

• SFI_TOOL_PROFILE now defaults to core. 19 tools are advertised and directly invokable; the other ~190 are reached with sfi.run_analysis { name: 'sfi.<tool>', args }. Set SFI_TOOL_PROFILE=full to restore the previous advertise-and-invoke-everything behavior.
• liveEnabled: true no longer opens the live plane. Grant standing consent with sfi.live_consent { grant: true } (OrgId + principal bound, scoped, 7-day expiry) or set SFI_LIVE_PLANE_ENABLED=1.
• Existing on-disk live grants stop working — v1 consent records are dropped. Re-grant once.
• The update check is now opt-in. Set SFI_UPDATE_CHECK=1 (default network mode is off).
• Every success envelope gains a contentPolicy block (~280 bytes) marking org metadata as untrusted data for hosts.

Install

Requires Node.js 20+ and an authenticated Salesforce CLI (sf).

npm install -g sf-intelligence

(or run it ad hoc with npx -y sf-intelligence … — no global install needed)

Register the MCP server

Claude Code (from your Salesforce DX repo):

claude mcp add --transport stdio --scope project sf-intelligence -- npx -y sf-intelligence mcp

Claude Desktop, or any other MCP client — add to the client's MCP config:

{
  "mcpServers": {
    "sf-intelligence": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "sf-intelligence", "mcp"]
    }
  }
}

First run

From your Salesforce DX repo (the directory with sfdx-project.json):

sfi init                               # create the local org-kb/ vault
sfi refresh --target-org my-org-alias  # retrieve metadata, build the vault
sfi status                             # freshness, source-tree hash, counts
sfi doctor                             # diagnose sf CLI / vault / auth issues

Then ask anything in your MCP client — "what fields does Account have?", "what breaks if I delete this field?", "why can't this profile see Opportunities?", "give me a tour of this org."

Boundaries

Read-only and offline by default. Static analysis, not runtime. No business record data in the vault. The product names its limits plainly rather than guessing.

Documentation

Full guides, capabilities, the tool catalog, and configuration: https://sfi.auditforce.cloud

• Getting started
• Capabilities · All tools
• Configuration · FAQ
• Quality & trust

License

MIT + Commons Clause — see the LICENSE file shipped in this package, or https://sfi.auditforce.cloud/licensing.html.