Cursor iOS app messaging desktop agents, Claude Haiku 5.5 costs tenth of Haiku 4.5, Slack bots handle release checks like users
Your daily Cursor briefing. What shipped, what changed, and what it means for the things you are building.
Edition date: 2026-10-09
Chapters:
00:00 Cursor's iOS app can now read and message agents running on your own computer
01:03 Claude Haiku 5.5 arrives in Cursor at about a tenth of Haiku 4.5's cost on short requests
01:55 One team hands release checks to Slack bots (automated chat helpers) and 10+ AI agents that use the build like real users
02:49 A permissions.json file in a project's .cursor folder breaks Cursor's agent, one user reports
03:37 Cursor user reports $22.95 in extra charges despite a $5 spending limit
Cursor HQ · cursorhq.ai
Maya: Hey builders! Here's what's on the radar today on Cursor HQ. Cursor's iOS app can now message your local agents, Claude Haiku five point five landed cheap, a team handed releases to Slack bots, a permissions file broke the agent, and a billing overrun. I'm Maya.
James: And I'm James. Let's get into it.
Maya: So the Cursor iOS app can now list the local agents running on your computer, show what each one is doing, and let you send a message. The changelog calls it Remote Control, and it's on by default for everyone except Enterprise.
James: Right, and the setup is short, Maya. Download the app, sign in, and your computers appear automatically. Tap one, approve the pairing in the desktop app, and your agents show as a list you can open and answer.
Maya: Exactly, and nothing moves to the phone. Cursor says the agents keep running on your machine, so they work even if your phone loses signal. The flip side is your computer has to stay on and online.
James: That's the catch, Maya. There's a Keep this computer awake setting, but the changelog says the machine needs to be plugged in with the lid open. A laptop in a bag is not covered.
Maya: Onto pricing. Cursor added Anthropic's Claude Haiku five point five to its model list on October seventh. Cursor says that on shorter requests it costs ten X less than Claude Haiku four point five. To use it, open Settings, then Models, and turn it on.
James: And the price is ten cents per million input tokens, fifty cents per million output, Maya. There's a separate tier above one hundred thousand input tokens, where rates climb to fifty cents input and two dollars fifty output.
Maya: Worth flagging, James: Cursor's posts don't define shorter requests, and they don't say how that ten X comparison applies beyond that wording. So keep an eye on your own usage before assuming the savings.
James: Fair, and one more. Cursor cut cache reads for Claude Sonnet five point five to ten cents per million, down from twenty. Cache reads are the charge for reusing context, so that halves the line item for Sonnet users.
Maya: This next one is a real workflow, James. A developer automated their team's release process using two Slack bots built with Grok Bot. One, called sandcastle, acts as release manager. The other is an engineer bot that takes the bug fixes.
James: And the flow is concrete, Maya. The developer tells sandcastle to cut a release. It DMs every contributor with links to their PRs, kicks off the build, and monitors it. Contributors can DM back with blockers.
Maya: Then the fuzz swarm, which is my favorite part. Sandcastle starts a group of agents, usually ten or more, on Grok four point seven at xhigh. They click through the build like real users, and a few act as chaos monkeys clicking at random.
James: Right, and when the swarm finds an issue, the engineer bot spins up a Cursor Project to triage and fix the high-priority ones. Depending on severity, the fix gets cherry-picked and a patch release cut. This is one team's setup.
Maya: Now a rough one. A user on the Cursor forum reports that a file named permissions dot json in a workspace's dot cursor folder makes every agent action fail. Even a valid empty curly brace triggers it, and removing the file fixed it immediately.
James: And the failures are specific, Maya. Running a shell command, reading a file, searching with Glob, calling an MCP tool. Each returns Invalid arguments, field Required, and the user says the arguments are stripped before they reach the tool.
Maya: Exactly, and it persisted across new chats, a restart, and an update. The setup was Cursor three point two two point one two on Windows with PowerShell, using GLM five point three Flash Low. They couldn't even retrieve a Request ID.
James: Last one, Maya. A Cursor user says the editor recorded about twenty two dollars ninety five in on-demand usage, even though their spending limit was set to five dollars. On Reddit, they describe working in Auto mode, with expensive models they had disabled showing in the breakdown.
Maya: And the timing is what confuses them most, James. They say the usage was recorded while their laptop was offline and idle during dinner, and they found no suspicious GitHub logins. They couldn't reach support, which only accepts work email, and nothing in the post shows how the charges were generated.
Maya: That's your Cursor HQ daily brief. Check cursorhq.ai for the full stories. Keep building.