Skip to content

SFWIZ

SFWIZ is a terminal-based agent harness for Salesforce development, inspired by Claude Code but purpose-built for the platform.

Core trick

Routes each turn to one of six persona subagents with per-persona tool scopes and gates deploys behind a mandatory user confirmation.

Tags

Description

SFWIZ is a terminal-based agent harness for Salesforce development, inspired by Claude Code but purpose-built for the platform. When you work on Salesforce day to day, you end up juggling org auth, metadata deploys, anonymous Apex runs, SOQL probes, and release notes that shift every season. Generic LLM agents don't really know the platform, and they forget the moment a new release ships. SFWIZ tries to fix both problems at once. Under the hood there's a single orchestrator running on the Anthropic SDK with streaming and prompt caching. It routes each turn to one of six persona subagents through the Claude Agent SDK: org-admin, designer, developer, deploy-manager, reviewer, and qa. Each persona has its own strict tool-scope, so the reviewer is read-only and only the deploy-manager can actually push metadata. Anything destructive (deploys, scratch creation, permset assignment) goes through a mandatory ask_user gate that blocks the agent until you confirm. For lifecycle work it shells out to the sf CLI rather than reimplementing what's already there, and for runtime queries it uses jsforce. A background Bun Worker quietly scrapes the Apex reference, the LWC guide, and the last few release seasons into a local qmd RAG store, so the agent gets smarter about Salesforce over time instead of staying frozen at its training cutoff. The whole thing ships as a single compiled binary.
arufian/sfwiz★ 60 forksTypeScriptApache-2.0Created Apr 24, 2026 · last push Jun 11, 2026

README excerpt

<p align="center">
  <img src="assets/logo.png" alt="sfwiz logo" width="320" />
</p>

```
 ███████╗███████╗██╗    ██╗██╗███████╗
 ██╔════╝██╔════╝██║    ██║██║╚══███╔╝
 ███████╗█████╗  ██║ █╗ ██║██║  ███╔╝
 ╚════██║██╔══╝  ██║███╗██║██║ ███╔╝
 ███████║██║     ╚███╔███╔╝██║███████╗
 ╚══════╝╚═╝      ╚══╝╚══╝ ╚═╝╚══════╝
 Salesforce × Claude — terminal-native AI harness
```

**sfwiz** is a Claude-Code-style interactive TUI harness for the Salesforce ecosystem — built for Apex developers, LWC engineers, and Salesforce admins who want AI-assisted workflows without leaving the terminal.

> Hackathon project — powered by Anthropic Claude.

## Demo

![SFWiz — full pipeline run in the terminal](docs/screenshot.png)

## Architecture

- **Orchestrator** (`@anthropic-ai/sdk` `messages.stream()`): streaming tool-use with manual dispatch.
- **6 persona subagents** (`@anthropic-ai/claude-agent-sdk` `query()`): org-admin · designer · developer · deploy-manager · reviewer · qa. Each runs as an isolated subagent with its own tool-scope and model (Opus 4.7 for reviewer + designer; Sonnet 4.6 for the rest). Structured JSON returned via the final `result` message is injected back into the orchestrator as a tool-result.
- **Tools**: filesystem (read/edit/write/grep), shell, jsforce (SOQL/describe), `sf` CLI (deploy/scratch/permset/retrieve/tests/apex), and `ask_user` for confirmations.
- **Continuous-learn worker** (Bun Worker, opt-in): daily scraper → qmd embed of Apex reference, LWC guide, and 

Team