Changelog

What’s new in AgentDx

Every release, newest first.

1.2.0

November 4, 2026

Projects: work in Salesforce DX folders on your Mac, each tied to one org that stays in view, and deploy, retrieve, test and query from the editor with the commands you know from VS Code.

Projects

  • AgentDx opens on your projects, most recent first. Each one is a Salesforce DX folder on your Mac, tied to one org, and that org is named in the title bar, the status bar and the project list.
  • New Project creates the Salesforce CLI’s standard project layout in ~/AgentDx or a folder you choose, and can retrieve the org’s code into it.
  • Open Project… opens a project made with the Salesforce CLI or VS Code and suggests the org its .sf/config.json names.
  • Changing a project’s org asks first and updates .sf/config.json, so the Salesforce CLI and VS Code use the same org.
  • If you edited an org’s files in AgentDx 1.1, the first time 1.2 starts they are copied into a project named after that org, ending in “(from AgentDx 1.1)”.

Deploy and retrieve

  • Deploy, validate or retrieve a file, a folder, a manifest or the whole project from the Explorer, the editor or the command palette. Settings in sfdx-project.json, such as string replacements, apply as they do in the Salesforce CLI.
  • Deploy results list each component, and errors show on the lines they name and in the Problems panel.
  • Compare a file with the org’s copy before you deploy or retrieve it.
  • Deploy on save, off unless you turn it on, and never for a production org.
  • In a production org, a deploy or anonymous Apex needs the org’s alias typed first.
  • Deploys and anonymous Apex runs are recorded in the change journal, like the agent’s changes.

Commands from VS Code

  • A command palette (⇧⌘P) with SFDX commands named as they are in VS Code’s Salesforce Extensions.
  • Create an Apex class, Apex test, trigger, Lightning web component, Aura component or Visualforce page from the palette, or from the folder it belongs in.
  • Run anonymous Apex from the open file or a selection, and SOQL from a selection or the palette, with results in a table you can copy as CSV.
  • Run Apex tests from Run Test and Run All Tests above a test class, or pick them from a list. Results show each method, the coverage, and a link from each stack trace to its line.
  • Get Apex debug logs from the org and open them in the editor.
  • Go to File (⌘P), Find in Files (⇧⌘F), Save All, and New File, New Folder, Rename and Move to Trash in the Explorer.

The agent and the editor

  • Each chat works in the project’s folder, against the project’s org.
  • Completions and error checking for JavaScript, JSON, CSS and HTML files now work in the editor.

Approval and safety

  • A shell command is judged by its riskiest step, so a command chained after an sf command gets the approval it would get on its own.
  • A command or tool call that names another org is judged by the stricter of that org and the chat’s, so a production deploy started from a sandbox chat asks as one.
  • Skills from the bundled Salesforce plugins no longer let a shell command or file edit skip its approval.

Updates

  • When an update has downloaded, AgentDx offers to restart into it, and offers again each time it opens until it is installed. Later keeps you working, and quitting installs it.
  • An Updates view in the sidebar shows the version you have, the download’s progress and Check for Updates.
Download 1.2.0

1.1.1

November 3, 2026

Fewer interruptions: the language server’s checks never ask for approval, and Apex test runs ask only in production. The preflight check counts triggers per object, and each request to Claude carries about half the context.

Approval and safety

  • Apex and LWC diagnostics and SOQL checks from the language server run without an approval card.
  • Apex test runs, through the Salesforce MCP server or sf apex run test, run without asking in developer, sandbox and scratch orgs. In production they get a review card, without typing the org’s alias.
  • A test run no longer shows in the Changes panel as a change that can’t be undone: its data rolls back on its own.
  • An org without an alias is named on approval cards the way the org list names it, not by its 18-character ID.

The agent

  • When existing automation already does what you asked for, the preflight check’s instruction to stop now reaches the agent, instead of the raw report.
  • The preflight check counts active and inactive triggers for each object, so a request that touches Account and Contact no longer reads Contact’s triggers as Account’s.
  • Recent setup changes reach the agent in full, not just how many there were.
  • Each request to Claude carries about half the context it did.
  • After your first message, how hard the agent thinks only goes up within a session, which keeps the conversation in Claude’s cache.
  • “Yes, deploy it” and similar replies get the thought a deploy needs, instead of being taken for a question.

The app

  • Tool calls and approval cards show paths within the org’s source folder, such as main/default/classes/AccountHandler.cls, with the full path on hover.
  • Each tool call is named for what it does, such as “Check Apex for errors” or “Find files”, and a shell command by what it is for.

1.1.0

October 31, 2026

Claude is now included. Sign in, start the free trial, and the agent runs on Claude Sonnet 5.5 through AgentDx, with no Anthropic API key to set up.

Accounts and plans

  • Sign in with Apple, Google or a 6-digit code sent by email.
  • Yearly with a 7-day free trial, or monthly. The first 100 founding members pay half for as long as they stay subscribed.
  • Checkout and billing through Stripe in your browser; AgentDx unlocks as soon as the purchase goes through.
  • No Anthropic API key to create, paste in or pay for: the key setting is gone.

The agent

  • Claude Sonnet 5.5 for everything: the conversation, every subagent and background work.
  • AgentDx sets how hard the agent thinks: questions that only read run light, changes get more thought, and production or repeated failures get the most. The “think hard” and “quick” phrases and the /effort command are gone.
  • Long sessions are summarised automatically as they grow, so they can keep going.
  • Attached images and PDFs can be up to 2 MB each.

1.0.0

October 29, 2026

The first release. AgentDx reads your org before it writes to it, shows every change as a diff you approve, journals each one, and treats production like production.

Org grounding

  • Connect production, sandbox, developer and scratch orgs through your browser, using the Salesforce CLI’s own connected app and auth store.
  • A staged org index: the basics in seconds, objects and automation in a minute or two, code in the background.
  • A preflight check before every write, enforced by a deny-gate that holds even when tools may run without asking.
  • “What changed in this org”, answered from the Setup Audit Trail.
  • Decision memory: answers to questions AgentDx asked you, remembered per org.

Approval and safety

  • Approval cards with Apex line diffs, Flow element-by-element diffs and a before-and-after permission matrix.
  • Org-risk colours everywhere an org appears; an org AgentDx can’t identify is treated as production.
  • Typed confirmation of the org’s alias for permission changes, deletes and destructive metadata, and for larger changes in production.
  • Production metadata deploys run as a validation with RunLocalTests first.
  • Apex lint as code is written, Code Analyzer (PMD) findings at severity 1 or 2 block an Apex deploy, and a test gate with a 75% coverage floor.
  • Hard deletes switched off in every org.

Undo

  • A change journal written before every change runs and settled after.
  • One-click revert for metadata deploys and record updates, with a plain list of what a revert won’t restore.

The app

  • A workbench with a Monaco-based editor, file explorer and org browser.
  • Attach documents and images for context: Word, Excel and PowerPoint read as text; images and PDFs as they are.
  • Your Anthropic API key, stored encrypted with the macOS Keychain.
  • No telemetry from AgentDx or anything it runs.
  • Separate builds for Apple Silicon and Intel, signed, notarized and updating themselves.