Warp Terminal: Open-Source Client, Agent Orchestration, and Oz Pricing Explained

Warp is a Rust-based, GPU-accelerated terminal that groups each command and its output into a discrete, selectable block rather than a continuous scroll — and, as of April 28, 2026, its client is open source under a dual AGPL-3.0/MIT license, with OpenAI signed on as the GitHub repository's founding sponsor. The repository passed 60,000-plus stars within months of the release. Warp now describes itself less as a terminal and more as an agentic development environment: it runs its own Agent Mode for multi-step coding tasks, and — distinctively among terminal-native AI tools — orchestrates other vendors' coding agents, including Claude Code, Codex, and Gemini CLI, inside the same window rather than locking a developer into one vendor's CLI.

For a team evaluating it, the deciding factor is whether multi-vendor agent orchestration inside a genuinely open terminal client is worth adopting Warp specifically for that role, given that the open-source release covers the terminal only — Oz, the cloud platform that actually runs and coordinates background agent work, remains closed-source and proprietary.

··········

WHAT'S OPEN, WHAT'S NOT, AND WHAT THE TERMINAL ACTUALLY DOES.

License boundaries, block-based UI, and cross-agent orchestration define the product beyond the open-source headline.

The April 2026 release put the Warp terminal client — the Rust codebase, the GPU rendering pipeline, the warpui and warpui_core UI crates — under AGPL-3.0, with the UI crates specifically under the more permissive MIT license. That split matters mechanically: AGPL requires that anyone distributing a modified version as a service release their changes too, while MIT on the UI layer allows freer reuse of just the interface components. What did not move into the open repository is Oz, Warp's cloud agent orchestration platform, which reacts to webhooks, CI events, or Slack messages and runs agents in the background without anyone at a keyboard — that layer stays proprietary and vendor-hosted.

Inside the terminal itself, Warp Blocks solve a specific, common problem: a long session's output becomes searchable, individually selectable, and shareable rather than one continuous stream a developer has to scroll through to find where a command started. Agent Mode translates natural-language requests into commands, edits, and test runs directly in the shell, and MCP server support lets the agent reach external tools the same way Claude Desktop or other MCP clients do. The distinguishing feature is that Warp doesn't require picking its own agent over a competitor's — Claude Code, Codex, and Gemini CLI can run inside the same terminal window, which makes Warp a host environment for multiple agent vendors rather than a single-vendor lock-in.

........

Component	License / status	Function
Terminal client (Rust, GPU rendering)	AGPL-3.0	Open-source core; modifications distributed as a service must be released too
UI crates (warpui, warpui_core)	MIT	More permissive reuse of the interface layer specifically
Oz (cloud agent orchestration)	Proprietary, closed-source	Runs background agents reacting to webhooks, CI events, Slack messages
Warp Blocks	Built into the open client	Groups each command and its output into a searchable, selectable unit
Agent Mode	Warp's own agent, cloud-dependent	Runs multi-step coding plans directly in the shell
External agent support	Orchestrated, not self-hosted	Runs Claude Code, Codex, and Gemini CLI inside the same terminal window

........

··········

WHY "OPEN SOURCE" HERE DOESN'T MEAN SELF-HOSTABLE AI.

The terminal is inspectable; the AI pipeline that makes it useful for agentic work still depends on Warp's paid infrastructure.

Making the client AGPL-licensed lets developers inspect exactly what data flows to Warp's servers and how the agent communicates with LLM providers — a direct response to earlier criticism that Warp was the only AI-first terminal keeping its code closed while asking developers to trust its telemetry and data handling. That transparency is real, but it doesn't extend to running Warp's AI features independently of Warp's own infrastructure: Agent Mode, Warp Drive collaboration, and Oz's cloud orchestration all depend on hosted services that are not part of the open-source release and cannot be self-hosted, regardless of the client's license.

Security and compliance features follow the same pattern as the AI features rather than the license: SOC 2 compliance, SSO, and Zero Data Retention are available, but gated to the Business and Enterprise tiers rather than included with the open-source client or the free plan. A team drawn to Warp specifically for the open-source angle should be clear on what that angle actually covers — code transparency and the ability to fork or audit the terminal client, not a path to running the agentic features on infrastructure the team controls.

··········

WHAT WARP'S CREDIT-BASED PRICING ACTUALLY COSTS.

Free covers light use; the paid tiers price AI actions by credit rather than by a flat per-seat rate.

Warp's free tier includes 75 AI credits per month, four concurrent cloud agents, and three indexed codebases — enough to trial Agent Mode and Oz without committing, but not enough for daily heavy use. The paid tiers scale by credit allowance rather than by feature unlocks alone: Build runs $18 per month for 1,500 credits, and Max runs $180 per month for 18,000 credits, with command generation, agent turns, and codebase indexing queries all drawing from the same pool.

........

Plan	Price	Credits / month	Notable inclusions
Free	$0	75	4 concurrent cloud agents, 3 indexed codebases
Build	$18/mo	1,500	Same AI feature set as Free, higher usage ceiling
Max	$180/mo	18,000	Same feature set, enterprise-scale usage headroom
Business/Enterprise	Custom	Custom	SOC 2, SSO, Zero Data Retention

........

Warp does not publish a per-action credit cost, so estimating whether 1,500 monthly credits covers a given team's usage means running the tool for a billing cycle rather than calculating it in advance — a materially different budgeting exercise than a flat-subscription competitor like Claude Code's Pro or Max plans, where the AI access is bundled into a fixed price rather than metered per action.

··········

THE DECISION RULE FOR EVALUATING WARP.

Warp earns a spot in a developer's stack in proportion to how much value multi-vendor agent orchestration inside one terminal window actually delivers over running Claude Code, Codex, or Gemini CLI separately in a plain terminal — for a developer who has already standardized on one agent vendor, that orchestration layer adds cost without adding capability. The open-source client matters most to developers who specifically wanted auditability of the terminal's data handling and rendering pipeline; it does not unlock a self-hosted path to the AI features themselves, since Agent Mode and Oz remain dependent on Warp's proprietary cloud regardless of what license covers the client. The credit-based pricing is the detail most likely to surprise a team moving from a flat-subscription coding assistant — running Warp for a full billing cycle before committing to Build or Max is the only reliable way to know whether 1,500 or 18,000 monthly credits actually match real usage, since Warp itself does not publish the per-action cost needed to calculate that in advance.

·····

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org