Conductor

Detailed report

Conductor organizes parallel coding work on a Mac, but does not support your own headless-server fleet. Development is continuing after a reversed wind-down announcement. Cloud availability and commercial terms remain unclear.

Research checked November 13, 2026. Undated documentation was retrieved on that date.

Product status

The product is Conductor at conductor.build, developed by Melty Labs and now part of Meta. Its official brand kit identifies it as a Mac app for parallel coding agents and provides its letter mark and wordmark. This is distinct from Netflix’s Conductor workflow engine and Google’s Gemini CLI extension named Conductor.

The status claim has a factual basis, but needs a timeline:

• November 4, 2026: Conductor announced it was joining Meta. Its changelog scheduled Cloud to become read-only November 13 and close November 20, with local model updates and critical fixes continuing until early 2027.
• November 9: version 0.94.0 appeared, demonstrating continued maintenance after the acquisition announcement.
• November 12: reporting described the founder’s reversal. The official announcement now says, “Plans have changed! We’ll still be developing Conductor,” and says it will keep working normally. November 12 reporting dates that reversal but explicitly notes uncertainty over Cloud’s deadlines.

Verdict

“dead” is no longer accurate. “Development revived. Cloud availability and commercial terms unresolved” is defensible. The conflicting official pages prevent a stronger assurance.

What it offers

Conductor turns parallel agent work into an organized development workflow. Each task gets files, a branch, terminal, conversation, diff, and review path. You can delegate several independent changes without manually juggling terminal windows and Git worktrees, then inspect and merge the results in the same application. The benefit is an integrated workflow. This research does not measure model intelligence. The documentation describes the complete task-to-review loop.

For a solo engineer already paying for Claude or Codex, the appeal is practical: spend less attention remembering which agent owns which branch, and more attention evaluating the work. Its clearest use case is several concurrent, reasonably independent tasks on a Mac.

Agents, models, and billing

Conductor supports four agent integrations: Claude Code, Codex, Cursor Agent, and OpenCode. The official provider guide documents authentication for each integration: Claude subscriptions, CLI login or API keys, Codex sign-in or API keys, Cursor API keys, and OpenCode provider credentials.

OpenCode exposes additional provider choices, including OpenRouter, Baseten, Cerebras, and Vercel AI Gateway. Claude-compatible endpoints are also configurable. Model availability depends on the integration and account. This is not an unrestricted guarantee that every provider model works.

Conductor explicitly does not resell or bill model usage. Your subscription limits or API charges remain separate. A useful billing warning: setting ANTHROPIC_API_KEY can cause Claude Code to use API billing instead of subscription usage. Cursor’s documented API integration should not be assumed to consume an ordinary Cursor editor subscription.

Devices, remote execution, and persistence

The desktop application supports macOS, with Windows and Linux explicitly unavailable. There is no documented installable headless Linux worker for your servers.

Two different remote architectures matter:

Your Mac

Experimental Remote Control lets an iPhone browse/create local workspaces, read chats, send prompts, select models, and stop agents. Pairing uses a QR code and the same account. The Mac must remain awake, online, and running Conductor. Traffic is end-to-end encrypted through Conductor’s relay. Devices reconnect when remote control is re-enabled. This is documented phone-to-Mac control. I found no equivalent documented laptop-to-laptop control or automatic synchronization of local sessions.

Vendor infrastructure

Conductor Cloud runs agents in Vercel-hosted Linux sandboxes. The Cloud FAQ explicitly says using your own cloud machines is not yet supported. Self-hosting was an Enterprise roadmap item. Its published configuration is Amazon Linux 2023, eight vCPUs, 16 GB RAM, and 32 GB ephemeral storage in us-east-1.

Cloud supports SSH into its sandbox, port forwarding, and one-way cloud-to-Mac file synchronization. That is materially different from attaching Conductor to your existing SSH server. Ordinary shell commands could call SSH, but treating that as a supported remote execution architecture would be inference.

Persistence has limits: Cloud files and chats survive sleep, processes do not. Workspaces sleep after 20 inactive minutes, and sandboxes stop after 23 hours 50 minutes even when active. Local sessions end when the app closes or Mac shuts down. Saved conversations are therefore not durable background execution.

Isolation, orchestration, and review

Local workspaces use Git worktrees: separate checkouts sharing Git history and repository data. They isolate edits, not operating-system permissions, and do not eliminate eventual merge conflicts.

Conductor supports independent workspaces or multiple agents sharing one workspace. The latter enables implementation/review collaboration but shares files and branch state. Testing uses saved actions, terminals, workspace-specific ports, or Spotlight testing from the original repository root. The Checks panel brings together Git status, PR details, CI, deployments, and review threads.

A beta HTTP API and MCP integration can create Cloud workspaces and supervise conversations. Its documented scope is Cloud, so it does not establish a general local fleet-control API.

Setup, security, and ownership

Setup involves installing the Mac app, authenticating GitHub and an agent, importing a repository, and configuring dependencies and ignored files. Shared setup/actions can live in .conductor/settings.toml; .env files and databases need explicit handling. The first-workspace guide makes those requirements clear.

Local agents run with your user permissions, without a security sandbox. Security documentation says model traffic goes directly to your provider. Privacy documentation distinguishes local chats from vendor-stored Cloud conversations and discloses analytics/crash collection. Provider policies still apply.

Conductor is proprietary under its November 3, 2026 terms. Its public GitHub repository contains release information and issues, not an open-source application you can maintain yourself.

Pricing: distinguish historical prices from current offers

The live pricing URL now redirects to the homepage. A cached official page records the previous monthly prices:

Cloud hours were described vaguely as “a bunch”. No exact allowance was published. Model charges were additional. November 8 reporting says sales stopped November 4, existing Pro features became free, and active subscribers received refunds. The reversal has not produced a clear replacement price sheet. Those historical figures are not verified purchasable offers on the research date.

Development and user reports

Launched in July 2025, Conductor has a substantial feature surface and frequent releases, but lifecycle stability is unsettled.

Independent reports show both attachment and frustration. In the November 4 acquisition discussion, users praised the integrated workflow and described being distressed about losing it. Others criticized migration notice. A thread displayed as “2mo ago” reports regressions involving PR instructions and repository settings. I could not verify its exact posting date.

More concrete issue reports include July 8’s background-job loss after restart and August 20’s unanswered-question lifecycle failure. Both remained open when retrieved, but concern older versions and do not prove current reproducibility. A November 9 user report describes losing Cloud access before the announced deadline.

These are anecdotes and issue reports, not consensus. Competitor promotion appears in these discussions. Vendor homepage testimonials are separate marketing evidence.

Fit and a practical trial

Choose Conductor for a Mac-centered workflow with frequent parallel branches and a preference for integrated review. Teams can benefit from shared Cloud conversations, handoffs, and administration, subject to renewed availability assurances.

Avoid relying on it as your primary control interface for headless Linux servers, mixed operating systems, or restart-safe unattended pipelines. The main concern is architectural fit, compounded by uncertain Cloud continuity.

Trial one repository with two independent tasks, Claude implementation and Codex review, existing tests, and a complete PR cycle. Pair an iPhone. Disconnect/reconnect it. Restart the app during disposable work. Then export and restore the conversation. Before adopting Cloud, verify actual access, billing, deadlines, and export behavior.