Is DevThrottle an AI agent?

No. DevThrottle does not write code and is not an agent - it is the control room for the agents you already run. It lines up Claude Code, Codex, and your other tools on one board so you can see what each is doing at a glance and answer them by voice. You bring the agents and their subscriptions, and running them through DevThrottle costs nothing.

The app is free forever, and your code never leaves your computer. We host nothing of yours, so it costs us nothing when you use it. The only thing for sale is prepaid credits, which pay for hosted AI features like transcription at a published rate card. The free version is not a trial and never expires.

Which coding agents does it work with?

It works with the AI coding agents you already use - Claude Code, Codex, Aider, Gemini, and others - across different tools at the same time. You keep your existing agent and subscription; DevThrottle runs and watches them all on one board.

Browser Harness review - Browser Use coding agent

Browser Harness is Browser Use's open-source browser-control tool for coding agents: a thin, editable harness built directly on the Chrome DevTools Protocol that connects an LLM to the user's real running Chrome. In the project's own words, it is "one websocket to Chrome, nothing between" - no framework layer, no selector engine, no wrapper API standing between the agent and the browser. It is MIT-licensed, written in Python, and installs as a skill for coding agents such as Claude Code and Codex. It is the browser-automation tool DevThrottle uses and recommends for agent-driven browser work.

Key capabilities

Compositor-level coordinate clicks — Instead of resolving DOM selectors, the agent takes a screenshot, reads the target's pixel, and clicks at that coordinate. CDP mouse events are hit-tested in the browser's compositor, so clicks pass through iframes, shadow DOM, and cross-origin boundaries - the exact surfaces where selector-based automation breaks down on iframe-heavy enterprise apps.

Attaches to the user's real browser — The normal local flow attaches to an already-running Chrome/Chromium debugging endpoint rather than launching a fresh instance, so the agent works inside the user's actual logged-in sessions instead of re-authenticating in a sterile profile. Input is dispatched through the CDP Input domain - the same path as real user input, not JavaScript-injected synthetic events - for both reading and writing.

Self-healing helpers — The core harness is deliberately tiny (about 1k lines across 4 core files). When the agent needs a capability the harness does not have, it writes the missing helper into an agent-editable workspace file during execution, so the harness improves itself every run.

Domain skills — Reusable, community-contributed per-site playbooks (GitHub, LinkedIn, Amazon, and more) capture the selectors, flows, and edge cases of specific sites, so agents do not rediscover them on every task. The skills are written by the harness itself as it works.

Cloud browsers — For headless servers, parallel sub-agents, or isolated work, the same harness can drive remote Browser Use Cloud browsers, with a free tier that includes concurrent browsers, proxies, and captcha solving.

Why DevThrottle recommends it

Command-line coding agents constantly need a browser - to verify a deploy, read a dashboard, or drive a web app - and most browser tools fail exactly where real work happens: behind logins and inside cross-origin iframes. Browser Harness solves both by design. Coordinate clicks reach anything visually on screen regardless of frame or shadow boundaries, and attaching to the running browser means the agent operates in the user's existing sessions. That combination is why it is the default browser tool in DevThrottle's own agent workflows.

Autonomy level

Level 3 (supervised agent): the harness executes whatever its driving coding agent decides, in a screenshot-act-verify loop that keeps each step observable. It is a capability layer for an agent session rather than an unattended long-horizon system, though remote cloud browsers raise the effective autonomy for headless deployments.

Strengths

Coordinate clicks pass through iframes, shadow DOM, and cross-origin boundaries at the compositor level
Attaches to the user's real running Chrome, reusing existing logged-in sessions instead of re-authenticating
Input goes through the CDP Input domain, the same path as real user input, for reads and writes
Self-healing: the agent writes missing helpers during execution, so the harness grows with use
Minimal, auditable core (about 1k lines) with an MIT license and no usage fees
Ships as a paste-one-prompt skill install for Claude Code and Codex
Scales to remote Browser Use Cloud browsers for headless, stealth, and parallel sub-agent work

Limitations

Coordinate clicking trades selector robustness for boundary robustness: targets are pixels, so layout changes require re-screenshotting rather than relying on stable selectors
Requires Chrome/Chromium with remote debugging enabled, including a per-attach permission popup on Chrome 144+
Chrome/Chromium only; no Firefox or WebKit support
Young project (first public releases in mid-2026, v0.1.x), so APIs and conventions are still settling
Not a test runner: no assertions, tracing, or CI-oriented test surface - scripted regression testing is better served by a framework like Playwright