One-click installer and configurator for the Local SF Architect engine — a fully local, offline Model Context Protocol (MCP) server that gives your AI assistant deterministic Salesforce architecture tools (governor-limit math, Apex blast-radius analysis, architecture pattern search, diagrams, scoring, and lint).
This extension does not replace the engine — it installs it from PyPI and wires it into whichever AI agent you use:
| Agent | How it is wired |
|---|---|
| GitHub Copilot / VS Code | Native MCP provider API (registerMcpServerDefinitionProvider, VS Code 1.101+); falls back to .vscode/mcp.json |
| Cursor | Merged into ~/.cursor/mcp.json (mcpServers) |
| Claude Code | claude mcp add --scope user (falls back to ~/.claude.json) |
Requirements
- VS Code (or Cursor) 1.101+
uvon your machine (the extension guides you if it is missing)- Internet access once for the engine install + the ~210 MB embedding and reranker models
What it does on first run
- Detects your host (VS Code vs Cursor) and installed agents.
- Installs the engine:
uv tool install --upgrade sf-local-architect. - Downloads the embedding model and seeds the local knowledge base.
- Wires the engine into every detected agent (existing configs are backed up).
After setup, open your AI chat in Agent mode and ask questions like:
- "What breaks if I refactor AccountService.cls?"
- "Will fetching 45,000 rows in a loop hit governor limits?"
- "Score the architecture health of this project."
Commands
All under the Local SF Architect category in the Command Palette:
- Run Setup (Install + Configure)
- Install / Update Engine
- Download Model + Seed Knowledge Base
- Configure GitHub Copilot / VS Code · Cursor · Claude Code · All Detected Agents
- Rebuild Local Stores
- Show Logs · Show Status
Settings
sfArchitect.enginePackage— PyPI package (defaultsf-local-architect)sfArchitect.autoSetupOnActivate— run setup automatically on first activationsfArchitect.autoConfigureAgents— wire detected agents without per-agent promptssfArchitect.autoDownloadModel— download the embedding model during setupsfArchitect.uvPath— absolute path touv(empty = auto-detect)
Privacy
The engine runs locally over stdio. Your code never leaves your machine. The only network calls are the one-time PyPI install and embedding-model download.
Note: a marketplace
icon.png(128×128+) should be added undermedia/and referenced frompackage.jsonbefore publishing.