Open Source Agent Tools: November 10, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. iphone-use-app Releases an On-Device Agent That Controls the Same iPhone It Runs On, Without a Mac at Run Time
iphone-use-app (EliotAndres/iphone-use-app, AGPL-3.0) published its first release, v0.1, and v0.2 less than ten minutes later on November 9; it is an iOS app in which a vision model operates the phone the app is installed on. The app mounts a developer disk image and uses a loopback VPN to take screenshots and send taps, swipes, long presses, typing and app-open actions; each step sends a screenshot to a model through OpenRouter or any OpenAI-compatible endpoint that accepts images and tool calls. A Mac is needed once, to create a pairing file, and the project requires iOS 26 or later with Developer Mode on; the unsigned build is sideloaded through SideStore or AltStore with LocalDevVPN for free Apple IDs, and it cannot go on the App Store. Version 0.2 makes Claude Haiku 5.5 the default model and stops sending the installed-app list, cutting the prompt from about 14k to about 1.7k tokens. The README benchmark replays 11 recorded steps twice per model and reports 91% correct decisions for Haiku 5.5 at a median of 1.3 seconds and about $0.0003 per decision, against 95% for Opus 5 at about $0.014. Runs can start from the chat, Siri, a Shortcut or an iphoneuse://run?prompt= link. Release Show HN
2. agent-email-guard Puts Budgets, Loop Detection and a Kill Switch in Front of Any Agent’s Email Send Tool
agent-email-guard (agentisend/agent-email-guard, MIT), a repository created on November 9 and posted to Show HN the same day by the maker of the AgentiSend email API, wraps an agent’s send function and returns one of allow, hold for approval or refuse for each message. By default it refuses the fourth near-identical message (same recipient, subject and at least 0.85 trigram body similarity) within 60 minutes and refuses sends that collapse onto one or two inboxes; hourly, daily and monthly budgets, a per-minute rate ceiling, a suppression list, approval holds for named or new recipients and a killSwitch.pause() call are off until configured. It ships for Node, Python, PHP and Go against one specification and shared conformance vectors, takes limits from a single agent-email-guard.json policy file, and has examples for nodemailer, Amazon SES, Resend, Postmark, SMTP, the OpenAI Agents SDK and the Vercel AI SDK. The README states that the in-process guard is advisory because code the agent controls can remove it and in-memory budgets are not shared across processes, so it recommends the Redis or Postgres store, and notes that AgentiSend’s hosted API enforces the same rules where the agent cannot bypass them. The maker’s demo sends 400 attempts of one receipt, with 5 delivered and 395 refused. Repository Show HN