Multi-Agent

iOS Simulator control: boot, install, tap, swipe, type, screenshot, accessibility audit, crash logs, leaks, xcodebuild and xcresult, App Store and TestFlight validation.

A real browser page: navigate, click, type, read, evaluate, screenshot, an accessibility snapshot whose refs the other tools accept, console and network capture with request detail, plus one-shot scraping with a shared cache and change tracking, sitemap-aware crawling with llms.txt, request mocking, offline mode, the clock and ARIA snapshot assertions.

Android emulator and device control over adb: install, launch, input, UI tree, logcat, meminfo, locale and font scale, APK audit, screen recording.

App Store Connect and Google Play Console from one family: apps, builds, reviews and replies, subscription prices per country, vitals, purchase checks and a bundle dry run, plus search, schema and call over every other store operation. Writes return a preview until confirmed, and keys stay in the Keychain.

Virtual iPhones through the vphone-launchpad CLI you install: host readiness, machine list, start and stop of existing machines, screenshot, tap, swipe, key, accessibility tree and an allowlist of read-only guest calls. Destructive methods are refused.

Language-server intelligence over a checkout: definitions, references, hover, diagnostics, workspace and document symbols, without loading files into context.

Mock-mode versus Figma: launch the app in mock state, enumerate the variants, compare pixels, spacing, typography and colour, and report what differs.

Read-only access to the product's other repositories at their latest integration branch: resolve to a pinned sha, grep, read, list, log, and look up join keys across up to twenty repos at once, with file text fenced as untrusted and tokens sent only to the hosts setup bound them to.

A local voice host over a Unix socket: speak, ask and hear the answer, stop, act and answer permission prompts by voice. The socket must be the user's own, and speech text is never logged.

The same offloaded file, also indexed: SQLite FTS5 ranks passages with bm25 and answers in about 50-100 tokens which part is about what, instead of grepping for which lines contain a string. Long text can also be paged into images, with the measured token trade-off.

Apple Wallet passes: build, sign, inspect and validate, with the signing passphrase resolved from the credential store rather than an argument.

Offloaded output: a large tool result is written to a file and queried, so the transcript never carries it.

Synthetic passport, residence permit and visa fixtures for MRZ scan tests: ICAO 9303 lines with correct check digits, rendered on a plain canvas with a permanent TEST SPECIMEN label and nothing that imitates a real document.

Simple text work handed to free-tier models instead of Claude: summaries, extraction against a JSON Schema, labels, translation and drafts, over Groq, Mistral, Gemini, OpenRouter free models, Z.AI and a local Ollama. Every answer is checked against its source before it is returned, and when none passes the caller does the task itself.

Provider-backed search (Brave, Perplexity, arXiv) and cited answers with quick, pro, reasoning and deep modes, recency and academic filters and follow-up threads. Keys are read from the environment at call time and go into a header; a literal key is not an accepted argument.

Offline security arithmetic: a CVSS 3.1 base score and band computed from the vector, and a lockfile parsed into an ecosystem, name and version inventory. Neither tool reaches the network; a CVE lookup is a separate step.

Key frames out of a screen recording, near-duplicates dropped, so a video the toolkit just made becomes something a model can look at.

Resolves a locator by id, role, label or text on iOS, Android or the web and says whether it matched exactly one element, none or several, without acting on it.