Warp Review

Block-based terminal, Git worktrees, a built-in editor. Warp isn't just another AI wrapper on bash — the interface itself is different, and that's a real bet, not a feature checkbox.

The pricing is the tell. $18/month buys 1,500 credits, roughly $20 of agent usage at API rates. That's usage-based cost dressed as a subscription. Fine for a five-person pilot, murky for budgeting at 200 seats.

Competes directly with Cursor and GitHub Copilot's CLI push, plus raw Claude Code and Codex, which Warp itself lets you run inside it. That's smart hedging — they're not betting the house on one model vendor. Terminal-based agentic coding is still early. No public funding data here, so runway is a guess, not a fact.

Rebuilding the terminal around structured blocks instead of scroll-text is a real architectural choice, not a skin. It changes how commands, output, and agent responses get referenced and reused, which matters once you're running Claude Code, Codex, and Gemini CLI side by side instead of one vendor's harness.

The credit model ($18/mo for 1,500 credits, 12x at $180 for Max) ties cost directly to agent usage at API rates. That's honest pricing versus flat-seat tools like Cursor, but it also means a heavy refactor month hits the bill directly — plan for variance, not a fixed line item.

Oz is the real long-term bet: self-hosted cloud agents, an OpenAPI spec, bring-your-own-LLM at Enterprise. If that orchestration layer matures, you've adopted a software factory, not a terminal. If it doesn't, you're left with a very good shell.

Free plan exists, real terminal, bring-your-own-inference. Build is $18/month, 1,500 credits, roughly $20 of agent usage. Credits run out, you reload. That's the real cost driver, not the sticker.

50 seats × $18 × 12 = $10.8K/year, base case. Add reload credits once teams burn through 1,500/month — category norm is heavy users blow past included usage by 2-3x. Year 3 could land $25-30K, not $10.8K. Max tier at $180/month (12x Build's credits) exists precisely because Build undersizes for power users.

Business tier is $45/seat, capped at 25 seats, adds SAML SSO — no separate SSO tax, unlike some competitors in this category. Enterprise is custom, contact sales, unlimited seats. Compare to GitHub Copilot Business at $19/seat flat — no credit metering, easier budget forecast. Warp's ROI story ties to Oz's PR review and incident response, but credit consumption makes cost per outcome hard to pin down.

The block model is the actual pitch: commands, output, and agent responses become discrete objects you can reference and reuse instead of grepping through scrollback. That's a genuine daily-workflow change, not a skin. Git worktree support baked in means agent-driven refactors don't collide with your main working tree — something tmux-plus-Copilot setups handle badly.

Day-3 friction shows up in the credit system. Build tier gives 1,500 credits (~$20 of usage) — burn through that mid-sprint on a multi-agent refactor and you're either reloading or throttled. That's a metered-usage anxiety Cursor and Claude Code subscribers don't deal with the same way. Multi-agent support (Oz, Claude Code, Codex, Gemini CLI) is a strong integration story on paper, but running four agent backends through one orchestration layer is more surface area to debug when something silently fails mid-task.

Docs cover Oz's SDK/CLI/API paths, which reads like engineers wrote it, not marketing. Power-user ceiling is high — Agent Kits, self-hosted cloud agents, Analytics API — but onboarding a solo dev into that depth takes real ramp time.

Block-based input/output is the real idea here. Commands and outputs become discrete chunks you can reference and reuse instead of squinting at scrollback, which is the kind of change that sounds small until you've lost an hour to copy-pasting terminal history. That's the stuff that actually saves time on day one, not the AI pitch.

The agent side is where it gets busy. Warp Agent, Oz, Agent Kits, multi-agent support for Claude Code, Codex, Gemini CLI — that's a lot of surface area for week one. Compare it to something like Cursor, which mostly hides its complexity behind a chat box. Warp asks you to understand orchestration concepts pretty early.

Credits are the sneaky part. Build is $18/month but only $20 of usage, and Max jumps to $180. Mobile isn't even in the conversation, this lives on your laptop only. No mobile parity to judge means no afterthought either, but three months in, your bill is the thing you'll actually be thinking about.

Three tells before pricing. One: title tag lists OpenAI four times — scraped noise, or SEO stuffing, unclear either way. Two: 'software factory' for Oz. That's the kind of superlative that ages poorly — Devin said similar things in 2024. Three: credit-based pricing ($18/mo for 1,500 credits, $180/mo for 18,000) means your real cost floats with model usage, not the sticker price.

The terminal itself is credible. Block-based I/O, Git worktrees, a built-in editor — concrete, shippable, differentiated from raw Claude Code or Codex CLI usage. That's the strong half.

Oz is the risky half. Orchestrating Claude Code, Codex, and Gemini CLI as a neutral layer is smart positioning, but it's also a bet that Anthropic or OpenAI don't just absorb this into their own CLIs. Exit is fine — it's still a terminal, you can walk away. Enterprise self-hosting is a real signal of viability, not just a checkbox.