We gave just-bash persistent storage

Agents love bash. They quickly generate complex shell commands, and are extremely powerful at manipulating data and processing text, all without tool calls. But there's a problem: they need to be sandboxed, and most sandboxes are overkill for simple agents that do things like jq, sed, and awk. Sure, there are safer ways to give agents access to real command line tools like virtual machines, microvms, and just-bash, but all of them have significant drawbacks ranging from needing to have root to create virtual machines to just-bash's memory-only persistence.

We think that just-bash is the right shape of a solution for something like agents built on Tigris, so we gave just-bash persistent storage backed by Tigris. Give your agents our @tigrisdata/agent-shell package, and they can work in a virtual bash environment on an in-memory filesystem, and when they're ready to write their changes, they flush their changes to write to a Tigris bucket.

Want to learn more? Keep on reading.

Security doesn't have to be shell-ish

Giving agents a regular old shell with a real filesystem is about as secure as it sounds. Simon Willison has been writing about this for months. Virtual shells like just-bash are an experiment in how to give agents a shell with proper isolation, without the weight of a sandbox. During a run, data only mutates within the shell, and agents only have access to the buckets you give them.

That's why we built agent-shell on top of just-bash to give it the persistence it craves without the overhead of a sandbox around the entire world. agent-shell is a virtual bash-like environment, based on just-bash, that mounts a Tigris bucket as if it was a normal Unix filesystem. As far as your agents care it's a Unix system, they know this! This lets you ls, sha256sum, and more on your data as if it was on your laptop instead of existing as objects in the multicloud. Why should you have to download your data to operate on it?

More importantly, when you combine this with forks and snapshots, even commands as scary as rm -rf / become entirely safe. Add in some @tigrisdata/agent-kit as seasoning and your agents get a shell to an isolated bucket fork:

import { createWorkspace, teardownWorkspace } from "@tigrisdata/agent-kit";
import { TigrisShell } from "@tigrisdata/agent-shell";

const { data: workspace } = await createWorkspace("agent-workspace-abc", {
  ttl: { days: 1 }, // auto-expire objects after 1 day
  enableSnapshots: true, // allow checkpointing later
  credentials: { role: "Editor" }, // optional scoped access key
});

const shell = new TigrisShell({
  accessKeyId: workspace.credentials?.accessKeyId,
  secretAccessKey: workspace.credentials?.secretAccessKey,
  bucket: workspace.bucket, // auto-mounted at /workspace
});

// pass this as an agent tool
await shell.exec("ls -la");

Et voila! You're off to the agentic races. You can see how this would develop from here: your agents have access to your data as if it was local, even though it isn't. You also don't have to worry about prompt injection tricking the agent into running exploit code because you've completely bypassed the normal shell. Agents can do anything they want, but malicious actions have no permanent effect.

But wait, there's more!

This is cool enough on its own, but we didn't stop there. We added our own custom commands to just-bash that let you fork and snapshot buckets as well as presign URLs for your agentic workflows. Here's a sneak peek at the buckety goodness:

$ presign hello.txt
https://my-awesome-bucket.t3.tigrisfiles.io/hello.txt?...

$ presign --put hello.txt

$ fork xe-screenshots xe-screenshots-botw-only

$ snapshot xe-screenshots --name "Before adding Funky Kong dataset"

$ snapshot xe-screenshots --list

This lets you generate download links for your objects, upload links for clients to directly upload objects, fork your buckets, generate snapshots for your buckets, and list all the snapshots for your buckets. Snapshots let your agents make mistakes and roll back with the restore function in agent-kit.

Yo dawg, heard you like buckets

You can even mount many buckets in the same environment:

import { TigrisShell } from "@tigrisdata/agent-shell";

const shell = new TigrisShell({
  accessKeyId: process.env.TIGRIS_STORAGE_ACCESS_KEY_ID,
  secretAccessKey: process.env.TIGRIS_STORAGE_SECRET_ACCESS_KEY,
  bucket: "agent-workspace", // auto-mounted at /workspace
});

shell.mount("shared-datasets", "/datasets");

If you want to see all your buckets at once, you can run agent-shell locally with npx:

npx @tigrisdata/agent-shell

Or check out https://tigris.sh:

Bash ahead fearlessly

When you combine bucket forks, agent-kit, and agent-shell, your agents can run whatever destructive nonsense they want and the worst case is rolling back a snapshot. We think that's key to making agents safe so that you can offload context to them and then never get woken up about it again.

Try it on your next agent and see how much sandbox you can ditch.

Shell-yeah!

agent-shell lets you isolate workloads with a lightweight bash-like environment so they can't step on other workloads. Give it a try!

NPM 📦