Warp Unbundles Its Coding Agent Into a Standalone CLI

Warp unbundled its coding agent into a standalone CLI on August 4, per the company's blog. It delegates tasks across Claude Code, Codex and other agent harnesses.

That price matters more than the feature list. Warp charges $18 a month for $20 of inference credit, with $10 ad-hoc credits or a bring-your-own-key option. The same subscription can route tasks to Claude Code, Codex or an open-weight model depending on the job. For anyone already paying for more than one of those tools, that's a budget decision, not a discount.

The CLI runs in Ghostty, iTerm2, VS Code, Windows Terminal or Mac Terminal, so switching editors doesn't mean switching agents. Underneath it, Warp built a tmux-like layer between the agent and the shell that keeps a session alive across directory changes, the company says. That layer can also drive full-screen apps like vim, sqlite and Python REPLs, not just line-based commands. It also runs remote SSH sessions without installing a binary on the far side.

Task complexity gets auto-routed across frontier and open-weight models, with YAML files for defining custom routers. The orchestrator can also hand a task to a subagent running a different harness entirely, including Claude Code or Codex.

That's the part worth arguing with. Model routing is close to table stakes for agent CLIs now. The harder problem is the session layer, the piece that keeps vim and SSH alive while the agent switches model backends underneath. That's what decides whether this replaces a terminal multiplexer or just sits on top of one.