Conductor

Conductor (conductor.build) is a Mac desktop app for running parallel coding agents in isolated git worktrees. Unlike older descriptions that framed it as Claude-Code-only, the 2026 product runs Claude Code, OpenAI Codex, and Cursor agents side by side, with experimental OpenCode support. You add a repository, spawn agents into separate worktrees, watch status from a unified dashboard, then review diffs and merge when the work is good. Local parallelism remains free; you bring your own model subscriptions and API keys. Paid tiers unlock Conductor Cloud (Vercel sandboxes), multiplayer, team admin, SSO, and the Conductor API. The problem Conductor solves is sequential agent bottlenecks and messy multi-window setups. One Claude Code session is capable but serial. Manually juggling several terminals, checkouts, and PR tabs does not scale when you have a backlog of independent bugs and features. Conductor makes isolation the default: each workspace is a git worktree, so agents do not overwrite each other's uncommitted files. You see what each agent is doing, steer or interrupt, and collect reviewable branches instead of hoping scrollback tells the truth. Pricing as published on conductor.build/pricing (August 2026): Free at $0 for parallel agents on local Mac workspaces with BYO subscriptions and keys. Pro at $50/month adds Conductor Cloud workspace hours, multiplayer, teams up to 5, the Conductor API, and a mobile app marked as coming soon. Teams at $60/month per user is for organizations of 5+ with admin portal, centralized billing, and SSO. Enterprise is custom with DPA, purchase-order billing, SCIM, custom security controls, SLA, and dedicated support. Model spend is separate, Anthropic Claude Max is commonly recommended when you run heavy Claude Code parallelism; Codex and Cursor costs follow those vendors' plans. Cloud compute is included for now on paid plans, with Conductor noting future usage-based cloud pricing may arrive. Cloud workspaces (Pro/Teams/Enterprise) provision 8-core / 16GB Amazon Linux 2023 sandboxes via Vercel in us-east-1, preloaded with Node.js 24, npm, pnpm, Python 3, git, GitHub CLI, ripgrep, git-lfs, tmux, and vim. Agents keep working after you close the Mac app, unlike local sessions that die when the machine sleeps. Privacy differs: local Conductor sends messages to your model provider and keeps session data on device; cloud sessions store chat inputs and outputs on Conductor's servers because the sandbox is managed. Company context: Conductor is Melty Labs / YC S24, founded by Charlie Holtz and Jackson de Campos, and raised a $22M Series A in 2026. Release release rhythm has been aggressive, changelog entries through 0.79.0 (Cloud Polish, early August 2026) after 0.78.0 introduced Conductor Cloud, with earlier 2026 milestones for Cursor support and Dispatcher, OpenCode, multiplayer/API early access, and current model pickers spanning Claude Opus 5 / Sonnet 5 / Fable 5, GPT-5.6, and Grok 4.5 class stacks. Competitors worth naming: CodexMonitor (open-source Codex-only Tauri command center, maintenance paused), Claude Code alone (same quality, no orchestrator), Cursor background agents (Cursor-native parallelism), and Meta Muse Code (terminal agent narrative rival). Conductor's bet is Mac-native UX plus multi-runtime worktrees, now extended into cloud and multiplayer for teams willing to pay. Honest fit: Solo Mac developers with Claude Max or Codex seats get useful throughput on Free. Agencies and startups that want always-on sandboxes, shared workspaces, and admin controls should evaluate Pro or Teams. Windows/Linux-only shops are out of scope for the desktop app today. Security-conscious teams should read the local-vs-cloud data FAQ carefully before moving secrets into cloud workspaces. Extended guidance for engineering managers: treat Conductor as throughput infrastructure. It does not invent review quality. Parallel agents amplify good tickets and amplify vague ones. Require scoped prompts, acceptance checks, and human review before merge. Budget model spend separately from the $50 Pro seat, four Opus-class agents can burn a Max pool quickly. Prefer local workspaces for proprietary code when policy forbids vendor-stored transcripts; use cloud when laptop lid-close would otherwise kill overnight jobs. Pair with branch protections and CI so agent PRs cannot skip tests. Revisit seat count quarterly as Muse Code, official Codex desktop, and Cursor evolve. Procurement checklist: (1) confirm macOS fleet, (2) map who needs Free vs Pro vs Teams, (3) decide cloud data handling, (4) set Anthropic/OpenAI/Cursor spend caps, (5) standardize worktree cleanup, (6) train on review-before-merge, (7) keep a bake-off against CodexMonitor and vendor-native agents so Conductor stays a deliberate choice.