Local SF Architect (VS Code extension)

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).

It also bundles sf-compress (sfc) — a dev-support output compressor that wraps your sf CLI commands, stores the full raw output on disk, and feeds your agent only a compressed view (errors, failures, summaries, counts). A 200k-token deploy report or debug log no longer floods the context window, and nothing is lost — every output ends with an id you can expand on demand.

This extension does not replace the engine — it installs it from PyPI and wires it into whichever AI agent you use:

Requirements

• VS Code (or Cursor) 1.101+
• uv on 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

1. Detects your host (VS Code vs Cursor) and installed agents.
2. Installs the engine: uv tool install --upgrade sf-local-architect.
3. Downloads the embedding model and seeds the local knowledge base.
4. Wires the engine into every detected agent (existing configs are backed up).
5. Installs the local-sf-architect and sf-compress Agent Skills into the workspace so your agent reliably routes Salesforce questions through the engine's tools and compresses verbose sf CLI output.

After an extension update, existing installs automatically upgrade the engine and refresh the skill once per release — no manual steps needed.

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."
• "Deploy force-app and show me only the errors." (routed through sf-compress)

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

Dev-support output compression (sf-compress / sfc)

Verbose Salesforce CLI output is highly compressible — deploy reports, retrieve preview tables, --json payloads, and Apex debug logs shrink dramatically while keeping every error and summary. The bundled sf-compress engine is exposed three ways, all through this one extension/engine install:

• MCP tools on the same sf-local-architect server: sfc_status, sfc_on, sfc_off, sfc_run, sfc_show, sfc_stats, sfc_graph. Your agent calls these automatically when it recognizes dev-support work.
• CLI: sfc <args> (standalone) or sf-architect compress <args> (alias), e.g. sfc run -- sf project deploy start -d force-app.
• Agent Skill: sf-compress, installed into your workspace during setup alongside local-sf-architect.

On by default. sf-compress is active out of the box for dev-support loops (deploy/validate, test runs, debug logs, retrieve/preview, data queries). It stores the full raw output under a git-ignored .sfc/ folder and returns an id you can expand later (sfc show <id>). During analysis/architecture work — where full detail matters — turn it off per project with sfc off (or the sfc_off tool); sfc run then stays fully dormant.

sf-compress has no extra VS Code commands or settings of its own yet — it is driven by your AI agent (MCP tools) and the sfc CLI. State is project-local under .sfc/.

Settings

• sfArchitect.enginePackage — PyPI package (default sf-local-architect)
• sfArchitect.autoSetupOnActivate — run setup automatically on first activation
• sfArchitect.autoConfigureAgents — wire detected agents without per-agent prompts
• sfArchitect.autoDownloadModel — download the embedding model during setup
• sfArchitect.uvPath — absolute path to uv (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. sf-compress is local too — it executes your sf commands and stores their output on your machine only; it adds no network calls.