What is Conductor? The Mac app for parallel coding agents

The short definition

Conductor is a native macOS application that lets you run a number of coding agents at the same time without them colliding. You point it at a repository, create a workspace, and Conductor builds a git worktree for that workspace, starts an agent process inside it, and gives you a chat thread, a file tree, a terminal, and a diff view scoped to that one branch. Create five workspaces and you have five agents working five branches concurrently, each reviewable on its own.

The part that surprises people coming from Cursor or Devin is that Conductor supplies no intelligence of its own. It has no model, no proprietary agent, and no inference bill. Its own harness documentation is unusually direct about this: Conductor does not bill or resell model usage, and harness usage is billed through the provider account or API key. What you are buying, when you buy anything, is orchestration.

Who builds it

Conductor is made by Melty Labs, a San Francisco company in Y Combinator's Summer 2024 batch. Charlie Holtz is chief executive and Jackson de Campos is chief technology officer; the team page lists seven people. The product first appeared as a Show HN post in July 2025, described then as a Mac app that lets you run a bunch of Claude Codes at once, and it collected a few hundred points and a long comment thread on that framing alone.

In March 2026 the company announced a 22 million dollar Series A from Spark and Matrix, led by Ilya Sukhar at Matrix, who had also led the seed and joined the board, with Nabeel Hyatt at Spark, Y Combinator, and the founders of Notion and Linear participating. That brings total disclosed funding to roughly 24.5 million dollars. In April 2026 the team absorbed cmd, an AI-in-Xcode tool, and its founder Guillaume Sabran joined as a founding engineer.

The company states more than 100,000 builders on its marketing site and names large engineering organizations as users. Treat both as vendor claims. What can be checked independently is the release cadence, which is genuinely fast: the changelog carries several substantive releases per month through August 2026, including single-sign-on and SCIM support in mid August and a cloud sandbox startup optimization the week before.

How workspaces and worktrees actually work

Conductor's core mechanic is git worktrees, used the way the feature was designed to be used. A worktree is a second working directory attached to the same repository object database, so creating one is fast and cheap and does not re-download history. Conductor places them under a predictable path, one directory per repository and one below that per workspace, and names each workspace after a city so the branches are memorable rather than numbered.

The consequence is real isolation of the thing that usually breaks when you run agents in parallel: the working tree. Two agents editing the same checkout produce a merge of two half-finished thoughts. Two agents in two worktrees produce two branches you can review separately, and the loser costs you nothing but a deleted directory. This is the same reason parallel Claude Code sessions are usually recommended with a worktree per session rather than a shared repository.

The four harnesses, and how unevenly they are supported

Conductor runs first-party agents rather than reimplementing them. As of August 2026 that means Claude Code, Codex, Cursor, and OpenCode. Claude Code, Codex, and OpenCode ship bundled inside the app, in an application support directory, and you can point Conductor at your own installations instead. Cursor runs through the Cursor Agent API and needs a Cursor API key. Gemini is not a first-party harness; it is reachable only indirectly, through OpenCode's provider list.

Feature support is not uniform across the four, and Conductor publishes the matrix rather than hiding it. This matters more than it sounds, because the features that vary are the ones people choose an orchestrator for.

Model coverage tracks the harnesses closely. Claude Opus 5 support arrived within days of release in late July 2026, and Grok 4.6 arrived in August through the Cursor harness rather than as a separate integration. That is the pattern to expect: Conductor gets a model when one of its four harnesses gets it, which is fast for frontier Anthropic and OpenAI models and indirect for everything else.

What it costs, and what it costs on top of

Conductor publishes four tiers: Free at 0 dollars, Pro at 50 dollars a month, Teams at 60 dollars per user per month, and a custom Enterprise tier. Billing is monthly and no annual option is shown. Teams is invite-only, so the 60 dollar figure is a published list price you cannot buy without talking to the company. The full breakdown, including what the cloud hours allowance does and does not say, is in Conductor pricing.

The number that decides most evaluations is not on the pricing page. Because Conductor resells no inference, a Pro seat is additive to whatever you already pay for the agent. A developer on Claude Max and Conductor Pro is paying both bills every month, and the orchestrator is the smaller of the two. That is a defensible trade if the orchestration is what unblocks you, and a poor one if you mostly wanted a nicer window onto a plan you already own.

Where Conductor does not go

A fair explainer has to name the boundaries, because several of them are load-bearing for a purchase decision and none of them are hidden by the vendor.

• macOS only. The documentation says Windows and Linux are not available yet and offers a waitlist with no ship date. This is not a soft limitation you can work around; there is no build.
• No general web application. An account portal exists for API keys and billing, but the workspace interface is the Mac app.
• Mobile is in private beta. Marketing says the iOS app is coming very soon, while the changelog contains real mobile bug fixes, so a working beta clearly exists. It is not generally available, and there is no Android signal at all.
• No user-installable CLI. A conductor command exists inside cloud sandboxes so agents can orchestrate themselves, and there is a beta HTTP API and a hosted MCP server, but there is nothing to install on your own machine. Packages on npm with similar names are third-party.
• No audit log. Enterprise advertises fine-grained security controls and custom privacy settings, but audit logging is not named anywhere in the documentation or changelog. Ask before assuming.
• No spend controls, by construction. Conductor meters no usage, so there is nothing for it to cap. Budget enforcement lives in your Anthropic or OpenAI console, or in a workbench that does sit in the billing path.

One further risk is worth understanding rather than fearing. Conductor runs Claude Code through the Claude Agent SDK, which put it inside the scope of an Anthropic policy change on subscription usage in third-party tools announced in May 2026. Anthropic delayed that change indefinitely in June 2026, and Conductor published a note saying so. Delayed is not cancelled. Any product whose economics depend on your personal subscription being usable inside someone else's application carries that dependency, and it is worth pricing when you plan a year ahead.

Questions people ask

Keep reading

Continue through Agentic coding with Is SuperGrok worth it? The honest math for each tier.