The first problem with a coding agent is easy to notice: it can make a mistake. The second is stranger. Give a developer several agents and the question becomes who is doing what, in which files, and whether any of it is ready to ship. The machine can work all night. The human still has to remember which branch contains the useful part.
Conductor has made that awkward middle job its business. Its Mac app runs Claude Code, Codex, Cursor and OpenCode in parallel. Every task gets its own workspace, branch, files, terminal and diff. The developer can start one agent on a feature, another on a bug, and review each result before opening a pull request. There is no new model to learn. Conductor organizes the agents people already use.
- Start separate agent tasks without making them share a checkout.
- Keep the branch, files, chat and review attached to each task.
- Move work into cloud sandboxes when it needs to outlive a laptop or involve a team.
- Pay for coordination and cloud features; bring your own model subscription or key.
A branch for every ambition
Charlie Holtz and Jackson de Campos met playing Ultimate Frisbee at Brown. Holtz later worked at Point72 and led growth at Replicate; de Campos built machine-learning infrastructure at Netflix. They founded the company in 2024 and joined Y Combinator's Summer batch. Their legal company name, Melty Labs, is a reminder that naming a business is sometimes less tidy than naming a Git branch.
The original insight was practical. A single agent often leaves the developer waiting or switching between terminals. Several agents in one repository can collide. Git already had an answer to the file problem: worktrees, separate working directories tied to the same repository. Conductor made that mechanism the default shape of an agent task, then added the setup scripts, chat, diff, pull-request state and archive controls around it.
The design also exposes its own limit. A worktree keeps agents from changing the same files on disk; it is not a security sandbox. Local agents still run with the user's Mac permissions unless stricter controls are added. Ignored configuration files do not automatically appear in a fresh worktree, so a project may need copied files or setup scripts. Conductor saves the repeated choreography, but the project still has to be prepared for parallel work.
“A huge amount of my work now is wildly parallelized, in a way I've never had to worry about before.”An engineer quoted by Conductor
The laptop stopped being big enough
In March 2026, the founders announced a $22 million Series A led by Spark and Matrix. They said usage had grown tenfold since January. On today's site, Conductor says more than 100,000 builders trust the product, and points to engineers at companies including Linear, Vercel, Notion and Ramp. These are company-reported claims, useful as a measure of its ambition, not a count of paying teams.
The funding announcement contained a more revealing line: the first version ran small AI teams on MacBooks, but the founders wanted a platform available from a phone to a private cloud. That explains the product's subsequent turn. Conductor Cloud now places agents in isolated cloud sandboxes, with repositories and dependencies prepared in advance. Work continues when the laptop closes. A teammate can open the same workspace, watch a live chat, add a prompt or take responsibility for the task.
This is the difference between personal productivity and team infrastructure. A local worktree solves a developer's collision problem. A shared cloud workspace solves the manager's visibility problem: which tasks exist, which person owns them, and what an agent has actually done. September's Routines extend the idea to work triggered on a schedule or by GitHub events. The agent can begin before anyone opens the app; someone still needs to judge the result.
The bill for an extra pair of hands
The local Mac app is free. Conductor's Pro plan is $50 a month and includes cloud workspaces, multiplayer for up to five Pro users, the API and mobile access. Teams costs $60 per user each month and adds broader collaboration, centralized billing and administration. Enterprise pricing is custom. Users bring their own coding-agent subscriptions or API keys; those model costs sit beside the Conductor bill. The company says it plans usage-based pricing for cloud compute in the future.
There is a privacy distinction hiding inside that price table. In local sessions, Conductor says prompts go directly to the model provider and session data stays on the device. In cloud sessions, the company stores chat inputs and outputs on its servers to run the shared service. For a team choosing between the two, persistence and collaboration come with a different data path. Conductor says it holds a SOC 2 Type II attestation, but organizations still have to decide where their own code and conversations may run.
What the human is for
Conductor operates in a crowded neighborhood. Cursor has its own agent. GitHub has pull requests and review. A determined engineer can assemble worktrees, terminals and shell scripts without paying anyone. Conductor's argument is that the repeated handoffs between those pieces deserve an interface of their own. It integrates several agent providers instead of insisting that one model do every job.
That bet can fail in an ordinary way: agents may produce code faster than people can assess it. Running six branches at once does not grant a sixth pair of eyes. Tests, setup and review discipline become more important as the agent count rises. The company appears to understand the tension. It has spent product effort on diffs, checks, workspace status, pull-request flows and shared chats, not merely on adding more prompt boxes.
A small sign of how Conductor learns arrived in April. Guillaume Sabran, an early power user who had complained about bugs in its feedback portal, brought his AI-for-Xcode project cmd into the company. The account on Conductor's own blog is charmingly unvarnished: a demanding customer became a colleague. The method is easy to copy, even without a venture round. Watch where an expert user finds friction; if that person has built a good answer, work with them.
The larger lesson is almost embarrassingly managerial. Assign work clearly. Give it a place to happen. Make the outcome legible. Decide what earns its way into the main branch. Conductor has wrapped those habits around AI agents. The developer's new title may be manager, but the job still ends with a judgment call.