Commercial Agent Tools: November 3, 2026

1. Momentic Launches Mo, a QA Agent Harness That Bug-Bashes Web and Mobile Apps With Parallel Sub-Agents

Momentic (YC W24) launched Mo on Hacker News on November 2, a hosted agent harness that plans, explores and bug-bashes user flows in web, iOS and Android apps (desktop is listed as coming soon) instead of generating Playwright or Selenium scripts. Mo builds a directed acyclic graph as it explores, forks large product areas into parallel sub-agents that each get their own browser or device, splits work between planner agents that reason over screenshots and video and locator agents that resolve plain-English targets from page structure, and coordinates overlapping fixtures through “using” and “modifying” leases; a tunnel exposes private and local apps to the hosted infrastructure. On Momentic’s own QA-Bench (four open-source apps including Gitea and Metabase, 1,396 human-labeled bugs, 60 minutes, no source access) Mo found 269 bugs at 19.3% recall, 84% precision and $2.67 per bug, versus 86 bugs at 6.2% recall and $0.80 per bug for Codex with Playwright MCP and 59 bugs at 4.2% recall and $4.09 per bug for Claude Code; the dataset and harness are published on Hugging Face and GitHub. Pricing: usage-based, with $250 of Mo spend for new signups and paid plans that add infrastructure capacity, included usage and top-ups; PR testing for Vercel and Cloudflare deploy previews just shipped. Source Launch HN QA-Bench

2. LinearB Rebuilds Around Agentic Studio, Feeding Delivery History to Coding Agents Over MCP

LinearB announced Agentic Studio on November 2, rebuilding its engineering-intelligence platform around a unified context layer that normalizes data from Git repositories and project trackers so engineering leaders can ask in plain language why delivery has stalled, and so coding agents can receive rework rates, past production incidents and code ownership through MCP before they modify a pull request. The company cites its own benchmark of 2.7 million pull requests across 250 organizations: teams using AI coding tools nearly doubled merged changes year over year, but AI-assisted pull requests are about 2.6 times larger, wait more than five times longer for review, and merge 79% of the time at “elite” organizations versus 92% for human-written code. The first agent skill, an early-access “effort skill,” scores a pull request’s complexity, rework history and risk to set how much reasoning and testing a coding agent spends on it, with code review, merge governance and token cost skills planned. Kraken used the MCP integration to trace slow reviews to timezone misalignment and cut review times by more than half in four months. Pricing was not disclosed in the announcement. Source

3. PhoneUse Rents Cloud Android Phones and Robot-Tapped Real iPhones to Agents Through MCP, an API or a CLI

PhoneUse, posted to Show HN on November 3, gives agents such as Claude Code, Codex, Cursor and Windsurf a phone through 12 MCP tools including phone.acquire, screen.read, tap, type, app.open and sms.read, with the same functions available over an API and CLI. Cloud phones are virtual Android devices that boot in about 4 seconds; real phones are retail iPhones and Androids in a San Francisco rack with carrier SIMs, operated by a physical stylus and read by an overhead camera with no jailbreak or helper app, so they receive SMS verification codes and pass device integrity checks such as Play Integrity and App Attest that emulators fail. One parameter switches a task between the two, and listed uses include app QA across many devices, mobile-only ordering apps and posting on social apps where browser bots are detected. Pricing: on-demand cloud phones at $0.024 per minute, a dedicated cloud phone from $20 per month, a real iPhone at $149 per month, and custom enterprise fleets. Source Show HN

4. Cloudbrowser Gives Each Agent a Persistent Cloud Chrome Profile That Can Be Paused, Forked and Handed Between People and Agents

Cloudbrowser, posted to Show HN on November 2 and labelled a preview, hosts isolated Chrome browsers whose profiles (cookies, local storage and downloads) persist across pause and resume, so a person can sign in through a live view link and then hand the same browser to an agent. Agents connect through a scoped CDP URL with Playwright, Puppeteer, agent-browser or similar tools, browsers are managed through a REST API and a Go CLI, and a paused browser can be forked into independent copies for parallel tasks, though in-memory page state is not guaranteed to survive a resume. Pricing: a free tier with 1 browser-hour per month and one running browser; Basic at $20 per month for 250 hours and 10 concurrent browsers ($0.08 per extra hour); Pro at $99 per month for 1,250 hours and 100 concurrent browsers ($0.06 per extra hour). Source Show HN

5. ArchDocs Adds Design Docs That Coding Agents Read Over MCP, With Agent Reports Kept Apart From Reviewed Evidence

ArchDocs, an architecture diagramming tool available since late 2024, introduced Design Docs on Show HN on November 2: a spec, open questions, an implementation plan and a work log that people and coding agents (Claude, Cursor, Codex, VS Code) share through a remote MCP server with OAuth. Agents receive the open questions with the plan and are instructed to ask rather than decide, request a task’s context before starting and leave a handoff (done, blocked, next) when they stop, and the evidence they submit stays labelled as agent-reported until a person reviews it. A plan-versus-recorded view flags planned changes with no recorded work and work done outside the plan, with pull request and CI evidence from GitHub on the Business tier and above. Pricing: Free at $0 for up to 3 seats (10 architecture docs, the MCP server and one design doc), Team at $16 per seat per month billed annually, Business at $39 per seat per month billed annually with a five-seat minimum, and custom Enterprise. Source Pricing Show HN

6. Smalt Lets Developers Talk to Claude Code Hands-Free on a Mac, With Speech Recognition on Device

Smalt, posted to Show HN on November 2, is a Mac voice assistant that connects to Claude Code: after running /smalt talk, the user speaks, Claude works on the user’s own plan while each step appears on an on-screen pill, and the answer is read aloud, with the option to cut in by talking over it. Speech is transcribed on the Mac’s Neural Engine with Parakeet (25 European languages, or 99 with Whisper) and audio is never saved or uploaded; the app also handles system-wide dictation, meeting notes, live translation and answers from local files, and Codex support is listed as coming. It requires an Apple Silicon Mac on macOS 14.4 or later, and the site links no source repository. Pricing: free, with no account. Source Show HN