Warp

You will love Warp if the terminal is where you live: Agent mode works beside your shell with your history and environment already in scope, and the Agent CLI runs the same agent without the app when you want it in a script or on a server. That is the daily trait that decides it: no new window. On a real repository it is strong at shell-heavy work and merely fine at large refactors.

Pick it for terminal and agent in one. Pick Claude Code for the strongest agent regardless of surface, and OpenCode if you want the same shape with your own keys from the first plan.

Two risks that compound. The sandbox is a container in Warp's cloud; the local agent runs in your shell outside it, so a wrong command is your problem, in your home directory. And bring-your-own-key, the setting that keeps code on your own provider account, is gated to a paid business tier, so the privacy-conscious configuration is the expensive one and the default is the leaky one.

The safe setup costs money and the cheap setup costs trust; pick knowingly. What it does right: cloud agents can be triggered from Slack, Linear, GitHub or webhooks, so the work that should not run on a laptop has somewhere else to run.

Two self-reported numbers. 1. SWE-bench Verified, 75.8 percent, September 2025, GPT-5 with "a light best@k wrapper" that proposes several candidate patches and selects one; the post does not state k, and the harness drives a desktop app a third party cannot rerun, so the figure is comparable to single-attempt leaderboard entries in no sense. 2. Terminal-Bench, 52 percent, June 2025, on a benchmark version since changed.

Both are marketing artifacts with footnotes, and the footnotes are the useful part. Quote them with the footnotes attached, or not at all. The observation: best-of-k measures the selector.

A company that shipped a terminal in April 2022 is now an agent vendor; open-sourcing the terminal, 64,754 stars, was a distribution decision: give away the surface, sell the inference. The business is credits, Build at $20 with 1,500 of them and Max at $200, the shape of a company whose margin is the spread between what it pays the labs and what it charges you.

Likely acquirer: a cloud or model lab wanting a terminal-shaped door onto developers. The pivot already happened, from terminal to agent. Position: watch the credit price; it is the only number that tells you how the spread is doing.

The demo is an agent fixing a build in the terminal. Procurement: Business is $50 per user with SAML SSO, capped at 25 seats, so sixty engineers means Enterprise and a custom quote. Overage is API rates plus 20 percent, which is a meter with a published markup, and I prefer that to a hidden one. Audit logs and retention terms are not on the page.

CI fit is real: the headless CLI runs in a pipeline. Onboarding is a day for anyone who has used a terminal. Approved with conditions: Enterprise contract, spend cap, retention terms in writing.

Warp going AGPL-3.0 was what I had waited for: I can read the Rust, build it, and file a fix instead of a ticket. The model layer takes a custom OpenAI-compatible endpoint, and MCP servers go in an mcpServers config with command or url, headers included, so my tool shelf plugs in as usual. The grievance: my local Ollama counts only after ngrok puts it on the public internet, because localhost and private addresses are rejected outright.

A fork survives the vendor now, and AGPL means a fork that ships has to stay open too, which I like. Grudging respect, upgraded to actual respect for the license.