Maya: Hey builders! Here's what's on the radar today on Cursor HQ. Turso agents branch databases with no token, a git setting escaped Cursor CLI's sandbox, Grok Bot hands coding to Cursor agents, a .cursorrules snapshot protocol, and Delta MCP. I'm Maya.
James: And I'm James. Let's get into it.
Maya: So Turso, the database company, just put a plugin in the Cursor Marketplace. You install it by typing slash add-plugin turso and signing in. After that, your agent can create, branch, and query databases directly from the editor.
James: And the part that stands out, Maya, is credentials. Turso says there's no API token to copy. You sign in instead, which removes that whole step of pasting a secret key into your agent setup.
Maya: Exactly right, and they posted a demo of an agent giving every user their own database, though the announcement doesn't spell out how it's built. Turso says it also works with Claude Code and Codex, linking a repo called turso-mcp. Announcements dated October eighth.
James: Now this next one is a security write-up. Cursor CLI wraps the shell commands its agent runs in a sandbox that blocks writes to your home folder. Researchers at accomplish.ai say the tool's own git commands skipped that zone.
Maya: And here's the mechanism, James. Their workspace had a core.fsmonitor entry in git config pointing at a script. They sent a read-only prompt, told the model to run no commands, and it ran none. The script fired anyway, no permission prompt.
James: Right, because the harness launches its own git commands, like ls-files and config, with no sandbox. The fix is universal git hardening, setting core.fsmonitor to false and core.hooksPath to dev null on every git launch. They verified it in build twenty twenty-six point oh eight point oh four.
Maya: One note worth flagging, James: accomplish.ai also markets a competing product that runs the whole agent inside a virtual machine. They disclosed responsibly, though, verifying the fix before publishing October sixth.
James: Onto Grok Bot, the app that lets you run a team of AI bots. It now has a primary bot, a single lead that takes your requests and routes each to the right specialist. Coding tasks go to Cursor cloud agents.
Maya: And the demo makes it concrete, James. They asked the lead bot to redo a list view using Opus five point five, and it delegated to a developer bot without being told. That bot used Cursor, and the change went straight to production.
James: Yep, and any bot can become primary, you just ask it to take the role. One builder describes a daily loop: watch Slack for feedback, create a Linear ticket, a Cursor cloud agent reproduces the bug and puts up a fix.
Maya: Exactly, and it tests the pull request with a swarm of agents, then rebases and auto-merges after one hour unless changes are requested. Another user hit a five-hour Claude limit mid-project and handed the work to Cursor. Video published October fifth.
James: This next one is clever, Maya. A developer published an open-source protocol written entirely as .cursorrules instructions. It tells Cursor's agent to save a copy of any file before editing it, keeping the last five versions. It's aimed at early prototyping before you've even initialized Git.
Maya: And the trigger was painful, James. An agent replaced a three-hundred-line logic block with a placeholder comment, and with no Git history the original was gone. So before each edit the agent archives into a local snapshots folder, capped around seventy-five kilobytes per file.
James: Right, and to roll back you restore only the broken file. Good edits elsewhere stay put. No Git, no Docker, no daemon. Though the developer notes it's only as reliable as the agent's adherence to those rules.
Maya: Last one, James. Delta MCP is a free local app being built to sit between Cursor and the MCP servers it connects to, like Linear and Notion. MCP is the standard that lets the agent reach outside services. The developer says it's still pre-release.
James: And the pitch, Maya, is two complaints: multi-step tasks get pricey with several servers attached, and it's hard to know what the agent changed. So instead of separate calls, the agent writes the whole task as one program, and every change is recorded and can be undone.
Maya: Exactly, and that gives you a log of what happened in Linear or Notion plus a way to roll it back. The developer says they're mostly there to listen, collecting feedback in comments and the project's GitHub issues page.
Maya: That's your Cursor HQ daily brief. Check cursorhq.ai for the full stories. Keep building.
James: And I'm James. Let's get into it.
Maya: So Turso, the database company, just put a plugin in the Cursor Marketplace. You install it by typing slash add-plugin turso and signing in. After that, your agent can create, branch, and query databases directly from the editor.
James: And the part that stands out, Maya, is credentials. Turso says there's no API token to copy. You sign in instead, which removes that whole step of pasting a secret key into your agent setup.
Maya: Exactly right, and they posted a demo of an agent giving every user their own database, though the announcement doesn't spell out how it's built. Turso says it also works with Claude Code and Codex, linking a repo called turso-mcp. Announcements dated October eighth.
James: Now this next one is a security write-up. Cursor CLI wraps the shell commands its agent runs in a sandbox that blocks writes to your home folder. Researchers at accomplish.ai say the tool's own git commands skipped that zone.
Maya: And here's the mechanism, James. Their workspace had a core.fsmonitor entry in git config pointing at a script. They sent a read-only prompt, told the model to run no commands, and it ran none. The script fired anyway, no permission prompt.
James: Right, because the harness launches its own git commands, like ls-files and config, with no sandbox. The fix is universal git hardening, setting core.fsmonitor to false and core.hooksPath to dev null on every git launch. They verified it in build twenty twenty-six point oh eight point oh four.
Maya: One note worth flagging, James: accomplish.ai also markets a competing product that runs the whole agent inside a virtual machine. They disclosed responsibly, though, verifying the fix before publishing October sixth.
James: Onto Grok Bot, the app that lets you run a team of AI bots. It now has a primary bot, a single lead that takes your requests and routes each to the right specialist. Coding tasks go to Cursor cloud agents.
Maya: And the demo makes it concrete, James. They asked the lead bot to redo a list view using Opus five point five, and it delegated to a developer bot without being told. That bot used Cursor, and the change went straight to production.
James: Yep, and any bot can become primary, you just ask it to take the role. One builder describes a daily loop: watch Slack for feedback, create a Linear ticket, a Cursor cloud agent reproduces the bug and puts up a fix.
Maya: Exactly, and it tests the pull request with a swarm of agents, then rebases and auto-merges after one hour unless changes are requested. Another user hit a five-hour Claude limit mid-project and handed the work to Cursor. Video published October fifth.
James: This next one is clever, Maya. A developer published an open-source protocol written entirely as .cursorrules instructions. It tells Cursor's agent to save a copy of any file before editing it, keeping the last five versions. It's aimed at early prototyping before you've even initialized Git.
Maya: And the trigger was painful, James. An agent replaced a three-hundred-line logic block with a placeholder comment, and with no Git history the original was gone. So before each edit the agent archives into a local snapshots folder, capped around seventy-five kilobytes per file.
James: Right, and to roll back you restore only the broken file. Good edits elsewhere stay put. No Git, no Docker, no daemon. Though the developer notes it's only as reliable as the agent's adherence to those rules.
Maya: Last one, James. Delta MCP is a free local app being built to sit between Cursor and the MCP servers it connects to, like Linear and Notion. MCP is the standard that lets the agent reach outside services. The developer says it's still pre-release.
James: And the pitch, Maya, is two complaints: multi-step tasks get pricey with several servers attached, and it's hard to know what the agent changed. So instead of separate calls, the agent writes the whole task as one program, and every change is recorded and can be undone.
Maya: Exactly, and that gives you a log of what happened in Linear or Notion plus a way to roll it back. The developer says they're mostly there to listen, collecting feedback in comments and the project's GitHub issues page.
Maya: That's your Cursor HQ daily brief. Check cursorhq.ai for the full stories. Keep building.
