Reconcile numbers across files
Hand your employee two or three exports and ask for the real total, the real average, or the real percentage change, computed and checked, instead of estimated in a reply.
Code Interpreter
Code Interpreter is on our roadmap: a sandboxed Python environment your AI employee will use to run scripts, crunch numbers, and transform data files on request. Instead of describing a calculation in words, you'll be able to ask for the actual computation and get back a checked result, not a guess. It is not available yet.
Most day-to-day requests to an AI employee are language work: draft this, summarize that, find the answer. Some requests are math and data work instead: total a column across three spreadsheets, clean a messy export, work out a real percentage change. Code Interpreter is built for that second kind of request. When it ships, your employee will be able to write a short Python script, run it in an isolated sandbox, and hand back the actual output: a number, a cleaned file, or a chart, rather than a language model's best guess at arithmetic.
The design goal is to keep this invisible to you as a separate tool. You will not open a code editor or manage a runtime. You will just ask, 'how much did we spend on ads last quarter across these three files' or 'plot signups by week', and the employee will decide code is the right way to answer, write it, execute it, and file the result in Drive alongside your other work products. The sandbox runs isolated per request, so nothing it executes touches your account, your other employees' work, or any system outside that one job.
The reason this is worth waiting for instead of settling for a spreadsheet formula or a generic chatbot is verifiability. A model asked to do math in plain text is pattern-matching toward a plausible-looking answer; a model that writes and runs actual code is checking its own work against real execution, and can retry when the first attempt errors. That distinction matters most exactly where mistakes are expensive: financial totals, data cleanup before a decision, or any calculation you'd otherwise have to redo by hand to trust.
You can already see this one coming. Code Interpreter appears in your employee's Tools tab today, marked as not yet available, alongside the tools that do work. Your employees know which tools are still on the way, so asking one to run a calculation on a file right now gets you a straight answer about what it cannot do yet, not a confident attempt at the wrong number.
It is worth being clear about where the line will sit. Code Interpreter is planned as a utility, the kind of thing an employee reaches for mid task, not a workspace you visit. There will be no editor to open, no environment to keep alive, and no runtime for you to manage. If it ever feels like a product you have to learn, the design has gone wrong.
Everything it produces is planned to land where your employee's other output already lands. A number comes back in the conversation you asked in. A cleaned file, a reshaped export, or a chart is filed to Drive beside your employees' other work. There is no separate console holding results you have to collect.
What It Will Handle
The clearest use is anything that mixes numbers and files: reconciling totals across multiple exports, computing a real statistic instead of an eyeballed one, reshaping a messy CSV into something usable, or turning a column of numbers into a quick chart. These are tasks people currently do in a spreadsheet, by hand, because asking a chat-only assistant for the number risks getting a confident but wrong one.
It's scoped to computation and data processing, not general software development. It's not a coding assistant for building your product, and it won't replace a database connection or an API integration; those are separate tools. Think calculator and data-wrangler for your employee, not a development environment.
Where It Fits In Your Workflow
Because the sandbox is temporary and scoped to a single request, Code Interpreter is meant to sit underneath a normal conversation rather than become a new place you go to work. You keep talking to your employee the way you already do; when a request needs real computation, that's handled automatically and the file or number lands back in the same thread, filed to Drive if it produced one.
That also means it inherits the same approval and visibility model as the rest of the employee's tools: nothing runs outside the request that triggered it, and the output shows up where you'd already look for the employee's work, not in a separate console you have to remember to check.
What To Do About Numbers Today
Nothing on this page is usable yet, so it is worth saying plainly what does work in the meantime. Your employees can already read and summarise spreadsheets and exported files, pull figures out of a connected app, and walk you through what a set of numbers appears to say. What they should not be asked to do today is arrive at a total or a percentage you then act on unchecked.
That distinction is the whole reason this capability is being built. Explaining is language work. Arriving at an exact figure is arithmetic, and arithmetic wants to be executed rather than reasoned toward. Until Code Interpreter closes that gap, treat a number that came back inside a paragraph as something to verify, not something to file.
How It Works
A sandboxed Python run per request, not a shared environment
When you ask for something that needs real computation, the employee will recognize it as a code task rather than a writing task, generate a short Python script for that specific request, and execute it inside an isolated sandbox created just for that run. There's no persistent server per employee and no shared state between requests, each execution starts clean and is torn down when it finishes.
If the script errors, such as a malformed input file or an unexpected data shape, the employee will read the error and rewrite the code rather than handing you a stack trace. Once the run succeeds, the result comes back in the conversation, and any file it produces, a cleaned CSV, a chart image, a computed report, gets saved to Drive the same way any other employee output does, so it stays where your other files already live.
Until that ships, nothing here is switched on. Code Interpreter is listed in the Tools tab as coming soon and is not attached to any employee, so there is nothing to enable and no way to spend anything on it. When it arrives, it is designed to appear in the same tools list you already use rather than asking you to set up something new.
Use Cases
Clean up a messy data export
Ask for a raw CSV to be de-duplicated, reformatted, or filtered down to the rows that matter, and get back a usable file in Drive instead of doing it by hand.
Turn numbers into a quick chart
Describe the comparison you want to see, like signups by week or spend by channel, and get a chart back without opening a spreadsheet tool yourself.
Check a number before you act on it
Ask for the figure behind a decision, a runway, a margin, a conversion rate, and get one that was computed rather than reasoned toward. The point is being able to use the answer without redoing it yourself first.
Comparison
| Before | After |
|---|---|
| Asking an assistant to total something across three files gets you a confident answer you then have to check by hand anyway. | The design has the employee write and run real code, so the number comes back from an actual execution rather than an estimate. |
| Reshaping a messy export means opening a spreadsheet and working through it row by row yourself. | You will describe the shape you want and get a usable file filed to Drive with the rest of your employee's work. |
| Real computation means a separate tool, a separate tab, and a separate place to go looking for the answer. | It is designed to sit underneath your normal conversation, with the result landing in the thread you asked in. |
| A calculation that hits a bad row gives you an error message to interpret. | The plan is for the employee to read the error, rewrite the code, and try again before it comes back to you. |
FAQ
Is Code Interpreter available now?
No. It's on the roadmap and marked coming soon; it isn't wired into any employee yet. This page describes what it will do once it ships, not a capability you can use today.
What kind of tasks will it be for?
Computation and data processing: totaling numbers across files, cleaning up a messy export, computing a real statistic, or turning data into a chart. It's aimed at requests where the answer needs to be calculated, not written.
Will the code it runs be able to reach my other files or accounts?
The design runs each request in an isolated sandbox scoped to that one job, so it won't have standing access to your other connected tools or accounts. It reads what you give it for that request and produces a result or a file, nothing more.
Will I need to know Python to use it?
No. You describe what you want in plain language and the employee decides when writing and running code is the right way to answer; you never see or edit the code yourself unless you ask to.
Can I see Code Interpreter in my workspace before it ships?
Yes. Code Interpreter is listed in your employee's Tools tab today, marked as not yet available, alongside the tools that already work. It is not attached to any employee and cannot be switched on, so there is nothing to configure and no way to spend anything on it before it arrives.
Will Code Interpreter replace my spreadsheet?
No. It is designed for the moments a spreadsheet is the slow way round: reconciling figures across several exports, cleaning up a messy file, or producing a chart on request. Your spreadsheet stays exactly where it is for the work you actually want to do by hand.
What happens if the code my employee writes fails to run?
The design has the employee read the error, rewrite the code, and try again rather than handing you a technical failure message. A malformed file or an unexpected column shape is treated as something to work around, and you see the finished result rather than the attempts behind it.
Where Code Interpreter fits
Code Interpreter is part of Things they can do out of the box.
Every AI employee ships with built-in abilities that work out of the box. Search the live web for real-time information, scrape and crawl any website for structured data, and generate images directly in conversation. No plugins to install, no APIs to configure.
- Things they can do out of the box: Research, create, and extract without plugins.
Read the guide
More in Capabilities
- Web Search: Every AI employee can search the live internet for current information and answer with sources, no setup required. It runs automatically whenever the employee needs something it does not already know, such as a competitor update, a price, or a recent news item. Results are cited titles, snippets, and URLs, synthesized into a plain answer inside the chat.
- Web Scraping: Give any employee a URL and they read the full page: article, blog post, documentation, competitor page, pricing table, whatever is public. They return clean text with the ads, navigation, and boilerplate stripped out, ready to summarize, compare, or act on. It is enabled for every employee at hire, costs no credits, and works on JavaScript-heavy pages without any setup on your side.
- Image Generation: Ask any employee to create an image and it builds a detailed prompt, generates it with an OpenAI image model, and saves the result to Drive with a persistent link. Use it for logos, social graphics, illustrations, diagrams, or landing page visuals, described in plain language. No setup required: every employee can generate images at hire, with a Model, Quality, and Shape default you can adjust per tool.
- Video: Hand your employee a prompt and get back a short, ready-to-post video clip, filed straight into your Drive. Built for social cuts, ad concepts, and quick product demos. Editing your own footage is coming next.
- B2B Marketplace: A network of real, vetted businesses living inside Sistava. Ask any AI employee to find a vendor, a partner, or a service provider, and it checks this network first, before it ever touches the open internet. Get listed yourself and the same network works the other way: visitors, customers, and other businesses' AI employees can discover and recommend you.
- Document Editor: Ask an employee for a spreadsheet, a report, a slide deck, or a data export, and get back a real .xlsx, .docx, .pptx, .csv, or .pdf file, not a wall of text pasted into chat. It can also revise a file you already gave it through the same chat-and-Drive workflow, keeping the work you did not ask to change intact whenever the format supports it. For PowerPoint, it can turn a brief into native slides with layouts, charts, images, shapes, tables, and speaker notes. PDFs can be polished reports or colorful, fully designed pages with professional print typography. Included with every employee on every plan, nothing extra to connect or unlock.
- Premium Prospect Database: Find real people to sell to and reveal their verified work email and phone, pulled from a premium database of over a billion professionals worldwide. Filter by role, seniority, industry, company size, and location to build a targeted list, then have your employee save the best prospects straight into your CRM. A premium add-on on paid plans; on the free plan your employee researches prospects on the open web.
- Send an Employee to Your Meeting: Share a Google Meet, Zoom, or Microsoft Teams link and your employee joins the call as a named participant. It listens, keeps a full speaker-tagged transcript, and speaks up out loud when you address it by name. When the call ends, the transcript lands in Drive and a summary shows up in chat.
- Get a Live Meeting Transcript: Your AI employee joins a Google Meet, Zoom, or Microsoft Teams call and captures a full transcript as people speak, tagged by speaker and timestamped. Ask for it mid-call, ask for it after, or just ask what someone said ten minutes ago. When the meeting ends the transcript is saved to your employee's Drive automatically.
- Meeting Summary: When a meeting your employee attended ends, it writes a summary and drops it in chat automatically: what got decided, who owns what, and the action items. No need to ask for a recap, it is already waiting when you check in.
- Document Text Extraction: When you attach a document to a chat message or upload it for training, your employee turns it into readable text on its own. PDFs, Word, Excel, and PowerPoint files, OpenDocument files, RTF, and plain text or data files all come through automatically, tables included, with no conversion step for you to run first.
- Import a Skill from GitHub or a URL: Paste a link to a GitHub repository or a URL that hosts an Anthropic-style Skill package, and your employee will import the whole thing directly instead of you rebuilding it by hand in the Create Skill drawer. This is on our roadmap and is not active yet: the option appears in the Skills toolbar and in the Create Skill drawer today, but importing does not run until we ship it.