Warp

Operator's take

There's a moment when agent fleets stop being a personal productivity trick and become an organizational question: who can start an agent, on what spend, running where, and who's allowed to see what it did. Warp — the company that spent years building the agentic terminal — answers that moment with Oz, a cloud platform for running fleets of coding agents, and its framing is the most enterprise-shaped in the lane: "a software factory that automates and orchestrates software development."

The pieces: agents start from anywhere — an @warp mention, an event trigger, or the API (warp agent run-cloud, a one-call agent.run for a prompt like "fix LIN-482"). The platform routes tasks across frontier and open-weight models automatically to balance quality and cost, and you can power the agents with Claude Code, Codex, or Warp's own, switching at any time. A single pane shows every agent running across the team, with one-click joins to audit a session; agent archetypes like FOREMAN and REVIEW structure the work; and self-improvement loops feed a persistent Agent Memory the site explicitly says is yours to keep even if you leave. It runs self-hosted or Warp-hosted, with prebuilt Docker dev environments published on GitHub. The proof point the company leads with is a customer whose Warp-built agent "Rex" lives in Slack, connects to Jira, and writes 54% of its own code — 35,000 lines a week.

The terminal itself is open-source (Mac, Linux, and Windows) and connects to repos via MCP; the platform around it is not. Pricing runs in tiers: a real-but-rate-limited free tier (cloud agents and codebase indexing), Build at $18/month with 1,500 monthly AI credits and frontier model access, Max at $180/month with twelve times the credits, and Business at $45/user/month with team admin controls. The remaining catch is the shape of the bet: this isn't a tool you adopt, it's a control plane your organization standardizes on. For a solo operator the local-first open-source tools are lighter and cheaper; for a team that needs governance over everyone's agents, that's precisely what Oz is selling.

What it's good at

• One control plane for every agent — Claude Code, Codex, or Warp agents, started from mentions, events, or the API, all visible in one team-wide view.
• Automatic model routing — tasks route across frontier and open-weight models to balance quality and cost, with per-agent harness choice that can change at any time.
• Auditability built in — join any running session with one click, sort by who originated the task, and see how the team actually uses agents.
• Memory that outlasts the platform — persistent Agent Memory the company says is yours to keep even if you stop using Warp, plus self-improvement loops.
• Self-hosted or Warp-hosted — run cloud agent workloads on your own infrastructure, with prebuilt Docker dev environments on GitHub.
• An agent CLI that works anywhere — Warp's coding agent installs into any terminal, separate from the Warp terminal itself.
• Open-source terminal, three platforms — the terminal is open source and runs on Mac, Linux, and Windows; agents get codebase context through MCP.
• Clear pricing tiers — free (rate-limited cloud agents and indexing), Build $18/mo (1,500 AI credits), Max $180/mo (12× credits), Business $45/user/mo (admin controls).

What it's not

• Not a solo-operator tool first — the value is team governance, routing, and observability; a single user gets less from the control plane than a team does.
• Not open source where it counts — the terminal is OSS, but the Oz platform is closed even when self-hosted; only the dev-environment images are public.
• Not transparently priced (yet) — Oz pricing didn't surface in capture; Factories is early-access with $10,000 usage credits for select teams.
• Not a worktree manager — it orchestrates cloud agents; the local diff-review loop belongs to other tools in the lane.