↓ Skip to main content
9 – 17 UHR +49 8031 3508270 LUITPOLDSTR. 9, 83022 ROSENHEIM
DE / EN
Orchestrator Melty Labs, Inc. (Conductor from Meta) USA Desktop app, cloud, mobile, API

Conductor

Conductor: Mac app that runs Claude Code, Codex, Cursor and OpenCode in parallel in separate Git worktrees, with diff review, cloud sandboxes and iOS app.

Licence Proprietary
First release July 2025
Interfaces Desktop app (macOS), Conductor Cloud (Linux sandboxes), iOS app, API and MCP server (beta)
Platforms macOS

The seven building blocks in Conductor

How Conductor implements the building blocks of an AI harness

  1. 1

    Agent loop

    Conductor does not run an agent loop of its own. Every session uses a harness: Claude Code, Codex and OpenCode come bundled with the app, and Cursor Agent runs through the Cursor API. Per its FAQ, Conductor adds its own system prompts that explain the workspace and actions to the agent.

  2. 2

    Tools and execution environment

    Tools come from the respective agent; Conductor manages MCP servers for Claude Code, Codex and Cursor, runs setup, run and archive scripts per repository, and offers an API with an MCP server for cloud workspaces.

  3. 3

    Context management

    Every workspace gets its own Git worktree with a branch under `~/conductor/workspaces/`; files Git ignores, such as `.env`, can be carried over via "Files to copy" with `.worktreeinclude` patterns, and every local workspace gets its own range of ten ports.

  4. 4

    State and memory

    Conductor stores chats locally under `~/Library/Application Support/com.conductor.app`; before every agent response, a checkpoint saves the code in a private Git ref, so code and chat can be reverted to an earlier turn. Cloud chats are stored on Conductor's servers.

  5. 5

    Permissions

    Per its FAQ, local agents run without a sandbox with the permissions of your user account. Approvals for shell commands, file changes, MCP calls and web fetches are available as a user setting; the docs do not state a default. Cloud agents run in isolated Linux sandboxes.

  6. 6

    Verification and feedback

    A diff viewer, checks, todos as merge blockers and a PR view with review and CI status support the review; unresolved review comments and failed checks can be handed back to the agent directly.

  7. 7

    Orchestration

    Conductor's core contribution: many agents in parallel, one worktree per workspace. Since version 0.95.0 from 16 November 2026, agents in local workspaces can create workspaces themselves, start other agents, wait for their reply and follow up.

Models and model access

Which models the harness uses and how they are obtained

Models and model access
Models Model-agnostic within the four harnesses: you choose the model per chat in the agent that is started
Access routes Anthropic (Claude Code), OpenAI (Codex), Cursor (Cursor Agent), OpenCode providers, Amazon Bedrock, OpenRouter, Vercel AI Gateway (via environment variables)
Free choice of model Yes
Custom endpoint / LLM gateway Yes
Conductor does not call language models itself for agent work and, according to its docs, does not bill model usage; billing runs through your provider account or API key. Endpoints such as `ANTHROPIC_BASE_URL` or `OPENAI_BASE_URL` can be set in the settings or through managed configuration.
Features
MCP servers Sub-agents Skills Headless / CI Checkpoints Resume sessions
Agents supported by herdr

Enterprise use

Data processing, controls and pricing model according to the vendor

Data processing and location

Locally, messages go directly to the model provider according to the docs, and chats stay on the Mac. Cloud workspaces run in Vercel sandboxes in the us-east-1 region; Conductor stores session inputs and outputs on its own servers, whose location the privacy policy does not state.

Permissions

Local agents run without a sandbox with all the permissions of your user account; a Git worktree only separates files. Conductor itself recommends a separate machine or VM. Cloud agents see the environment variables and secrets stored there and, once permitted, can also act on your Mac.

Telemetry

Conductor sends analytics events to PostHog, for example when you create a workspace or send a message, with metadata, error messages, operating system and IP address. Conductor does not document a switch to turn this off. "Enterprise data privacy" disables AI-based extras such as generated chat titles, as well as custom MCP servers.

Identity and team

Team plans with an admin portal and roles are invite-only; Enterprise adds automated user provisioning, an SLA and finer security controls. Via `~/.conductor/settings.managed.toml`, the default model, environment variables and "Enterprise data privacy" can be enforced centrally; per the docs, the format is provisional.

Pricing

The local app is free according to the docs, and Conductor does not bill model usage. Cloud workspaces require a Pro, Team or Enterprise plan; usage-based pricing for cloud compute has been announced, and there is currently no public pricing page. The website states SOC 2 Type II.

Strengths
  • One workspace, one worktree: parallel agents work on separate branches
  • Claude Code, Codex and OpenCode bundled; Claude Pro/Max and ChatGPT plans work directly
  • Path from prompt to merge with diff viewer, checkpoints, checks and PR view
  • Conductor Cloud with isolated sandboxes, shared team workspaces and an iOS app
  • Usage view with tokens and estimated costs per model and workspace
  • Managed settings for centrally enforced endpoints and privacy options
Limitations
  • Desktop app for macOS only, Windows and Linux on the waitlist
  • Only four harnesses: Claude Code, Codex, Cursor Agent and OpenCode
  • Local agents run without a sandbox with full user permissions
  • Proprietary, no public pricing page; product roadmap after the move to Meta is open
  • Cloud workspaces for GitHub repositories only, hosted at Vercel in us-east-1
  • Telemetry to PostHog without a documented switch to turn it off

Use cases

Typical tasks for this harness

Distribute several tasks in parallel to Claude Code, Codex, Cursor Agent or OpenCode, each in its own worktree
Let one agent start and coordinate further agents in their own workspaces
Review results in the diff, hand review comments and failed checks back to the agent and merge
Keep agents working in cloud sandboxes while the laptop is closed
Steer agents on the go with the iOS app, local workspaces via remote control

Conductor is a Mac app that runs several AI coding agents such as Claude Code, Codex, Cursor Agent and OpenCode in parallel and gives every workspace its own Git worktree. It adds diff review, pull requests, Conductor Cloud with isolated sandboxes, an iOS app and an API. As an AI harness, Conductor mainly covers the orchestration building block. As of 22 November 2026.

What is Conductor?

Conductor is a proprietary tool by Melty Labs, Inc. from San Francisco that launched in July 2025. On 4 November 2026, the team announced with version 0.93.0 that it was joining Meta and initially planned to shut down Conductor Cloud on 20 November. With version 0.94.2 from 13 November, it reversed the shutdown (“Conductor is back”), and the website has since listed the product as “Conductor from Meta”. According to the changelog, the current version is 0.98.0 from 20 November 2026, with several releases per week. The desktop app runs on macOS only; there is a waitlist for Windows and Linux. According to the website, more than 100,000 developers use Conductor.

Conductor as an AI harness

Conductor is the workspace layer above the agents: according to its harness overview, it manages repositories, workspaces, sessions, terminals, diffs, checks and pull requests, while the harness runs the agent loop and tool calls. Every workspace gets its own Git worktree with a branch, named after a city. Several chats in the same workspace share branch and code, so Claude Code can implement while Codex reviews in a second tab. Setup scripts prepare new workspaces, checkpoints save the state before every agent response, and the PR view shows review comments and CI status.

Besides local operation, there is Conductor Cloud: every agent runs in its own Linux sandbox with 8 vCPUs and 16 GB RAM, and workspaces and chats can be shared with the organization and prompted together. The iOS app controls cloud workspaces and, experimentally via remote control, local workspaces on the Mac as well. An API in beta creates cloud workspaces and starts agents, also through an MCP server.

Models and model access

Within its four harnesses, Conductor is model-agnostic: Claude Code uses Claude models via subscription, the Anthropic API or third-party providers, Codex uses GPT models via a ChatGPT plan or API key, Cursor Agent uses the models of the Cursor account, and OpenCode the providers configured there. According to its docs, Conductor does not bill model usage. For Claude Code, the provider guide describes OpenRouter, Vercel AI Gateway, Amazon Bedrock and Z.ai via ANTHROPIC_BASE_URL or Bedrock variables, and for Codex OPENAI_BASE_URL. This also lets you route model access through a central proxy such as innFactory’s AI Gateway. Since version 0.98.0, a usage page shows chats, tokens and estimated costs of local workspaces per model and workspace.

Company use: data protection and governance

Companies should settle four points before a broad rollout. First, permissions: per its FAQ, local agents run without a sandbox with all the permissions of your user account. A worktree only separates files, not processes, network access or credentials; Conductor itself recommends a separate machine or VM for extra safety. Second, the cloud: per the cloud FAQ, cloud workspaces run in Vercel sandboxes in the us-east-1 region, and Conductor stores session content on its own servers. Third, telemetry: analytics events with metadata, error messages, operating system and IP address go to PostHog, and the privacy page does not mention a switch for it. Fourth, the product roadmap after the move to Meta, which remains open for long-term rollouts. For central policies, Conductor offers managed settings in ~/.conductor/settings.managed.toml, for example for endpoints and the “Enterprise data privacy” option, which disables AI-based extras and custom MCP servers.

Which agents does Conductor support?

Conductor supports exactly four harnesses: Claude Code, Codex, Cursor Agent and OpenCode. Claude Code, Codex and OpenCode come bundled; Conductor can use your own installation instead. Cursor Agent runs through the Cursor API with its own API key. You choose one harness and one model per chat tab. Claude Code’s experimental agent teams can be enabled via an environment variable. Further CLIs such as Gemini CLI, Amp or Pi are not supported as harnesses; Orca, Paradev or herdr cover those.

How it fits: Conductor, Orca, herdr and the AI Gateway

Conductor supports fewer agents than Orca or Paradev and is not open source, but offers a tightly guided interface and an extensive cloud with team features. herdr, by contrast, relies on long-lived terminals and barely intervenes in the agents. You will find a detailed comparison in the article herdr vs. Orca. Conductor’s usage page shows costs per machine; developers’ costs and agent time across teams and projects are captured by the AI Gateway in proxy and companion mode. For business users outside development, CompanyGPT is the right harness; our AI harness overview gives an overview.

From developer harness to enterprise harness

Coding harnesses are tools for developers. Business users, cost control and governance need two more layers.

CompanyGPT – the harness for business users

CompanyGPT gives any language model tools via MCP servers, integrations and n8n, company knowledge via companyRAG, roles from Entra ID or Keycloak, and audit logs – operated in your own cloud tenant (Azure, AWS or Google Cloud) or sovereignly on STACKIT.

Discover CompanyGPT
AI Gateway – cost and governance for every agent

The AI Gateway bundles the model access of your tools and coding agents behind an OpenAI-compatible API: sign-in via Entra ID, Keycloak or Cognito, budgets and hard limits per team and user, cost attribution per project and repository via the Companion, and logging of every request.

View the AI Gateway

Frequently Asked Questions

What is Conductor?

Conductor is a Mac app for running AI coding agents in parallel. It launches Claude Code, Codex, Cursor Agent or OpenCode in separate workspaces, each with its own Git worktree and branch, and brings diff review, checkpoints and pull requests. It is complemented by Conductor Cloud with isolated sandboxes, an iOS app and an API. It was built by Melty Labs from San Francisco; the website now lists it as "Conductor from Meta".

Is Conductor open source and free?

Conductor is not open source: the maker does not document public source code or an open-source license, and the terms of Melty Labs apply. According to the documentation, the local app is free, and Conductor does not bill model usage. Cloud workspaces require a Pro, Team or Enterprise plan. Usage-based pricing for cloud compute has been announced; there is currently no public pricing page.

Which coding agents does Conductor support?

According to the documentation, Conductor supports four harnesses: Claude Code, Codex, Cursor Agent and OpenCode. Claude Code, Codex and OpenCode come bundled with the app; if you prefer, Conductor uses your own installation instead. Cursor Agent runs through the Cursor API and needs a Cursor API key. Other CLIs such as Gemini CLI, Amp or Pi are not supported as harnesses.

Does Conductor use my Claude or ChatGPT subscription?

Yes. By default, Conductor uses the logins already stored on the Mac, including a Claude Pro or Max subscription. Since version 0.89.0 from 29 September 2026, a ChatGPT plan can be connected for Codex via "Continue with ChatGPT". Alternatively, you switch to API keys. If `ANTHROPIC_API_KEY` is set, Claude Code may use that key instead of the subscription, according to the docs.

Can Conductor be used in a GDPR-compliant way?

That depends on the operating mode. Locally, prompts and code go directly to the model provider according to the docs, and Conductor does not access the chats; analytics metadata does go to PostHog, though. Cloud workspaces run at Vercel in us-east-1, and Conductor stores session content on its own servers. Both belong in your records of processing and require an assessment of the third-country transfer; the agents' model providers need to be assessed separately.

Does Conductor work with innFactory's AI Gateway?

Yes, through the agents it launches. Conductor itself makes no model calls for agent work. In proxy mode, you set `ANTHROPIC_BASE_URL` or `OPENAI_BASE_URL` in the settings, or centrally via managed configuration. The gateway's companion measures tokens and agent time from the local session logs of Claude Code and Codex, including subscription usage. Sessions in Conductor Cloud do not run on your machine, so the local companion does not cover them.

Introduce coding agents safely in your company

We support selection, model access in the EU, permission design and cost control.