Superset Review

Superset skips the crowded fight to build a better coding agent. Instead it orchestrates Claude Code, Codex, OpenCode, Gemini CLI, Copilot, and Cursor Agent from one dashboard, each task isolated in its own Git worktree and branch. Smart positioning: let the model vendors race, own the workflow layer above them.

The honesty is refreshing. Superset states plainly that worktrees separate files but don't sandbox processes or stop merge conflicts. That's a real tradeoff, not a marketing gap.

Pricing is legible: free for one user, $15/month Pro with unlimited users and automations, Enterprise adds SOC 2 Type II and SSO. Local-first storage, source-available under Elastic License 2.0. No cloud lock-in, no proxy fees on top of your own API keys. Windows isn't supported yet, which cuts out a chunk of dev teams. Still, for teams already running multiple CLI agents, this solves a real coordination problem cheaply.

The architecture decision here is the MCP server plus TypeScript SDK plus CLI, all three exposed, not just the desktop app. That means Superset is designed to be scripted and embedded into a broader agent pipeline, not just clicked through. If we adopt it, in three years we've built automations and cron-scheduled runs on top of its API — that's a real dependency, and Elastic License 2.0 source-availability is the mitigation, not a guarantee.

The OAuth 2.1 with PKCE and dynamic client registration for programmatic auth tells me someone on this team has built developer-facing infra before; that's not a checkbox feature teams bother with for a v1.

What concerns me long-term is the vendor's own admission that worktrees don't sandbox processes. That's an honest disclosure, but it means the isolation model is filesystem-level, not security-level — fine for solo repos, a harder sell once compliance asks about SOC 2 scope on the Enterprise tier.

Pro is $15/month, unlimited users, unit unspecified. That's not per-seat pricing — it reads like a flat account fee. If that holds at scale, 50 developers cost the same as 5. TCO math nobody else in this category offers.

Free tier caps at 1 user, local workspaces only. Fine for solo evaluation, useless for team piloting — you hit the wall the moment you need Remote access or Automations, both Pro-gated.

Enterprise adds SAML SSO, SCIM, audit logs, SOC 2 report, uptime SLA. No price shown, annual billing only, contact sales. Standard enterprise gate. Everything below that tier is priced in public — procurement can model Free and Pro without a call. Billing model, invoicing terms, and auto-renewal window aren't stated anywhere I could find. Ask before signing.

The port management piece in the diff dashboard is the tell. Anyone running more than two agent sessions against the same repo knows the real fight isn't the diffs, it's remembering which worktree bound to which dev server port before you start debugging a phantom 404. That's a day-3 problem, not a demo feature, and it's addressed here directly.

The Agent Skills install (npx skills add superset-sh/skills) tells me they've thought about how an agent discovers Superset's own capabilities, not just how a human clicks around it. Combined with the A2A agent card, this reads like infrastructure meant to be queried, not just operated.

Ceiling is filesystem-only isolation — fine for branch hygiene, useless if an agent runs something destructive at the process level. And no Windows build yet means half your team's WSL setups are locked out for now. For a Mac/Linux shop running Claude Code and Codex in parallel, $15/month is a trivial ask against the coordination tax it removes.

First ten minutes: you're not learning a new coding agent, you're learning a dashboard that watches your existing ones. That's a smaller ask than most tools in this category, and it shows in how fast the setup question answers itself — pick Claude Code, Codex, whatever, point Superset at your repo, go. The diff review dashboard and port management give you one screen instead of six terminal tabs, which is the actual daily win here.

Where it gets interesting is month three, when you've got a dozen worktrees spawned from automations running on a schedule and you have to remember which branch belongs to which forgotten cron job. Nothing about the tool stops that pileup from becoming its own chore. Terminal access inside workspaces helps when you need to intervene directly, but it's still you doing the cleanup.

No mobile app that I could find, so 'monitor your agents from anywhere' stops at your laptop lid. Fine for a solo dev, less fine if you wanted to glance at a runaway agent from your phone. At $15 a month for the Pro tier, cheap enough that the learning curve is the real cost, not the invoice.

"Orchestrate them all" is the tagline. Fair enough — the SOC 2 badge shows up twice on the homepage title tag alone, which reads more like a compliance sales pitch than a dev tool front page. That's a tell worth sitting with, not a dealbreaker.

The Q&A is refreshingly blunt: worktrees don't sandbox, don't stop conflicts, you review and merge yourself. I've seen fewer vendors admit the ceiling of their own architecture than claim it. That honesty buys trust.

Biggest open question: Superset doesn't proxy model calls or own the agents it orchestrates. If Claude Code or Codex change terms tomorrow, Superset's value shifts with them. It's a coordination layer sitting on other people's roadmaps — solid idea, dependent fate.