Scira is now a sandboxed research agent

Most AI search tools answer questions. Scira now works through them.

Today I'm shipping a sandboxed research environment. Every signed-in user gets bash, file I/O, a code interpreter, and a filesystem that survives across messages. The model can download data, run scripts, write files, and read them back on the next turn. It doesn't start over every time.

The sandbox#

The environment is an isolated VM. It starts when a chat starts and stays alive for the length of the session.

Every tool in a chat shares the same workspace. A file downloaded in one step is available to the code interpreter in the next. A script that produces a CSV can feed a chart without any copying. Work accumulates across messages.

Common scientific packages come pre-installed. Anything else installs on demand and stays for the session.

Here's a real example. The prompt was: "Research the revival of extinct wildlife, pull the metrics, build charts, and explain why it's happening now."

Scira researching extinct wildlife revival and generating charts

The model searched web and academic sources, then called the code interpreter twice to build the charts. Each step is labelled in the timeline. You can expand any of them to see exactly what ran.

Skills#

The sandbox is the engine. Skills are what make it useful.

A skill is a playbook. It tells the model how to approach a specific kind of task: what to search for, how to structure the work, when to write code. Not a prompt tweak. A methodology.

Scira Skill Directory

Eleven ship today. Six are on by default.

Deep Research breaks a question into sub-topics, searches across web, Reddit, X, and academic sources, then cross-validates before writing a final answer.

Extreme Search fans out as wide as possible. Good for contested topics, niche questions, or anything where the first-page results are wrong.

Python Analysis guides multi-step data work. Download a file, run a script, save the output, come back to it next message.

Web Extract pulls structured data from specific URLs when a regular search won't cut it.

Artifact Builder sequences the work of building rich outputs: UIs, documents, spreadsheets, code projects.

Deslop rewrites AI-sounding text to sound human. Paste in anything and it removes the inflated language, the filler, the sycophantic openers.

The other five (Social Pulse, Academic Review, Market Intel, Media Scout, and Local Guide) are domain-specific playbooks for sentiment analysis, literature reviews, financial research, media discovery, and location intelligence. They start disabled. Toggle them on from the Skills page.

What's next#

Skills are text files. The current set is built around the tools Scira already has, but the pattern works for any workflow worth encoding. I want to make them shareable eventually.

The sandbox persists within a session. Next up is durability across sessions and direct workspace access for Pro users.

Try it at scira.ai.