What's new in this release:
- Cursor, Claude Code, and custom agents sign in with browser OAuth
- Admins set Read; Read & Write; or per-tool access for each integration
- Guardrails block writes that break policy
- Every tool call logged by user, tool, and status
Every CIO I talk to asks the same thing about agents: what stops it from doing something dumb in our files? In 12 months, "which MCP tools is this agent allowed to call?" will be a standard line in enterprise security reviews.
Our answer in @Box MCP: admins choose which tools each agent gets, Guardrails block policy-breaking writes, and every tool call is logged.
Connecting an AI agent to company files takes minutes. Controlling what it does once it’s in there is the hard part.
The Box MCP server puts a governed layer between every agent and your content:
• Each client, from AI assistants to coding agents and custom frameworks, runs under the user’s own Box identity and permissions
• MCP Guardrails set what an agent may do, such as no deletes and no new files in externally shared folders
• Classification policies decide which content an agent can reach
• Every request lands in the audit trail
More: blog.box.com/box-mcp-server-…
The same shift is coming for knowledge work! Agents need to work where the content lives, beside the humans, under the same rules: see only what you can have permissions for, log everything, lose access when determined by admin etc.
That's the problem we're solving at @Box humans and agents collaborating on the same content, often in one single interface headlessly.
software factories feel like a temporary problem. the models will learn to build them for you
the more interesting question to me is how to build a workspace where agents can collaborate with you
Excited about this launch with @claudeai! We are fully invested in a headless experience for @Box. The side panels are connective tissue across documents for AI-powered knowledge work - minimizing context switching across apps.
This brings your documents to your AI and gets work done faster.
Also new: with interactive connectors, you can now open the full version of a tool inside Claude. Work in Adobe, Atlassian, and Box beside your conversation. Make changes yourself, or ask Claude to make them with you.
Rolling out on web and desktop.
Today, the calendar makes a right triangle, for the last time this century.
This Saturday is October 24, 2026.
Write it as 10/24/26 or 24/10/26, and the same beautiful relationship appears:
10² + 24² = 26²
100 + 576 = 676
But here’s what makes today special.
There are infinitely many Pythagorean triples. Only a handful fit our calendar.
Using the rule
month² + day² = (last two digits of the year)²,
the month cannot exceed 12, and the day cannot exceed 31. Those small limits leave exactly 12 qualifying dates in this century.
Today is the last.
The next will be March 4, 2105:
3² + 4² = 5²
That is a wait of more than 78 years.
An infinite family of triangles. Just twelve appearances on a century’s calendar.
Happy Pythagorean Triple Day.
Enterprise AI gets useful when understanding a document leads to action. Deciding which action requires fast decisions. Our work with Box Extract and OpenAI’s Decisions API explores that fast-decision layer - and turning doc understanding into an automated workflow.
We saw ~5× faster route selection, with the same accuracy. Lower-confidence cases went to reviewers.
Great to see this work featured in OpenAI’s launch. Thanks to the @Box and @OpenAIDevs teams!
Since the beta, we’ve added new capabilities to the Decisions API so you can now use:
• Native GPT-Live integration so voice apps can react as people speak.
• Prompt caching to help lower costs when reusing context.
• Native JSON inputs so you can pass structured app data directly.
• Images in individual question instructions so you can compare multiple images against a shared reference.
• Up to 400 questions per request so you can make twice as many decisions in a single call.
openai.com/index/decisions-a…
Enterprise AI gets useful when understanding a document leads to action. Deciding which action requires fast decisions. Our work with Box Extract and OpenAI’s Decisions API explores that fast-decision layer - and turning doc understanding into an automated workflow.
We saw ~5× faster route selection, with the same accuracy. Lower-confidence cases went to reviewers.
Great to see this work featured in OpenAI’s launch. Thanks to the Box and OpenAI Developers teams!
Since the beta, we’ve added new capabilities to the Decisions API so you can now use:
• Native GPT-Live integration so voice apps can react as people speak.
• Prompt caching to help lower costs when reusing context.
• Native JSON inputs so you can pass structured app data directly.
• Images in individual question instructions so you can compare multiple images against a shared reference.
• Up to 400 questions per request so you can make twice as many decisions in a single call.
openai.com/index/decisions-a…
🚨 OH MY WORLD ! 🤯🤯🤯
THIS IS HUUUUUUUUGEEEE,🔥🔥
THIS IS MAAASSSSSSIVEEE,🔥🔥
18-YEAR OLD,🇮🇳 INDIA'S ANAHAT SINGH HAS DOWNED TOP-SEED LEGENDARY NOUR ELSHERBINI 🇪🇬
WR 15 Anahat DOMINATED WR 3 Nour Elsherbini as she beat her 3-0 to enter QFs of Chicago Open (PSA Gold Level Event)
11-8, 11-8, 11-7
WE ALL BOW DOWN TO YOU, THE QUEEEEN 👑👑👑👑
WHAT A PLAYYAAAAA !
Guilty! And my fav category is agentic workflows for my kids :)
I have a math worksheet app for my kid. Every night at 11pm Claude starts a fresh sandbox and installs the repo from Github. It gets her progress file and photos of today's sheets from Box, grades them, and creates 2 new PDF worksheets which get stored back Box. Then the sandbox is deleted.
Box is the only layer that persists. My wife and I can both upload photos to the Box app on our phones, we can access even when we are traveling (long trips to India)!
No infra to run or maintain (no server, no db), collab across multiple parents!
I asked @Muse to connect to @Box. No built in connector but it guided me flawlessly - see screenshot.
Next - an overlap of my personal and professional life. We now have a "family brain". My wife and I each have a Muse. The agents use shared Box folders for kids’ activities, school lunch menus and travel plans. Now our calendars get auto-populated with school events, days we dont need to pack lunch, reminders for upcoming trips with packing lists etc..
Wild times - looking forward to discovering other use cases for this :)
Delightful @Muse experience - sold my car earlier this week and the buyer handled transferring the title.
In discussing with Muse, it mentioned that I need to fill the release of liability form with the DMV. Then went ahead and filled up the form with some inputs from me and submitted it - all while I was in the park with kids!!
At @Box as well as at our customers, teams are rolling out agents - for external customers and internal users. Serious agents need a top quality harness with access to a computer, obviously a model and a set of tools and context.
The Agents API launch by @OpenAIDevs is a huge deal IMO! It makes this part of agent building something you can rent - an AWS moment for agent building.
The computer (sandbox) gives an agent a place to run code, work across files, keep intermediate state, and coordinate longer-running tasks. The filesystem becomes a much more important part of the agent stack. For enterprise agents, a huge amount of the context and work lives in documents. Thats why we launched Box Mount earlier this week.
In this demo, you see the agent gets a real working environment (Agents API + OpenAI Sandbox), while Box Mount brings enterprise content directly into that sandbox as files the agent can reason across, and produce new work from.
As agents with a sandbox become more common, Box Mount makes enterprise content a native part of the agent’s workspace - with all the associated permissions and governance in place.
Demo 👇
Go from idea to a working agent faster with the Agents API.
Build and run cloud agents with the Codex harness, fully managed by OpenAI.
We handle orchestration, long-running sessions, and context management. You focus on what makes your agent unique.
Available in public beta.
Enterprises shouldn’t have to move content into a every new AI agent or recreate permissions and governance. The agent should be able to reach the enterprise context it needs, where it already lives, with the same access controls already in place.
As @ChatGPT, and other agents become the place where work gets done, the role of systems like Box changes: headless context layer that can power many different AI experiences.
Being the infrastructure agents are obligated to call is a nice position to be in.
OpenAI and Box are bringing your enterprise content directly into @ChatGPT.
As the file system for AI, Box provides headless access to enterprise context in ChatGPT while keeping the content managed, governed and protected by your existing Box permissions. Demo ↓
As the move towards headless mode for SaaS continues, being the infrastructure agents are obligated to call is a fairly lucrative position to be in.
Similar trends working at Snowflake, MongoDB and Cloudflare as well.
We've had 6.3M MCP tool calls run in August which is up 5x since March. Agents querying product data is our fastest growing usage category of Amplitude.
Inspired by @zenorocha