Tensorlake: Sandboxes, Git and Storage for AI Agents
Tensorlake is a cloud for running AI agents. It gives each agent an isolated sandbox, a small Firecracker virtual machine with its own filesystem, shell, packages and processes, that starts in well under a second, sleeps when idle and wakes when needed. Around those sandboxes, Tensorlake adds two things most rivals don’t: a versioned filesystem that saves an agent’s work automatically and can roll back to any point, and hosted Git built to handle millions of repositories. It also runs durable workflows that survive crashes. The company has raised $8.6 million and counts the car-rental giant Sixt among its customers.
The quick answer: Tensorlake is a newer sandbox platform that competes on keeping agents’ work safe and fast, with storage and Git designed for agents rather than borrowed from developer tools. It suits teams running coding agents, reinforcement learning rollouts, evaluations and agent-driven CI. It started out focused on document processing for AI and has broadened into general agent infrastructure. Everything below comes from Tensorlake’s site, pricing and developer summary, plus press coverage.
Tensorlake at a glance
Checked against Tensorlake’s site, pricing page and developer summary on November 13, 2026.
| Tensorlake | |
|---|---|
| What it is | Sandboxes, versioned storage, hosted Git and durable orchestration for AI agents |
| Isolation | Firecracker microVMs |
| Scale | Up to five million sandboxes per project, per Tensorlake |
| Features | Suspend and resume, snapshots, forking, SSH, Docker image import |
| Access | Python and TypeScript SDKs, CLI, REST API, agent skills for Claude Code and Codex |
| Pricing | Free tier; prepaid credits from $5; Pro $250 a month |
| Company | Founded by Diptanu Gon Choudhury (CEO); $8.6 million raised |
What Tensorlake does
Tensorlake’s developer summary calls it “the sandbox cloud for AI agents: stateful, AI-native microVM sandboxes with sub-second cold starts, snapshot, suspend/resume, and sandbox-native durable orchestration.” Its homepage describes the uses: “deploying agent harnesses, running RL rollouts and evals, and building CI that scales.”
There are two ways to use a sandbox. An agent can live in a persistent sandbox as its own computer, or an app can spin up a short-lived sandbox to isolate each tool call, so AI-generated code never runs on the app’s own servers. Sandboxes “sleep on inactivity and wake instantly when invoked,” so sessions can run for hours or days without paying for idle time. Teams can import any Docker image and have it boot unchanged.
A filesystem and Git built for agents
Tensorlake’s storage is its most distinctive part.
- Versioned filesystem. A durable volume mounts like an ordinary folder. Reads and writes hit fast local disk, and every settled change saves itself to durable storage, so “an agent keeps a versioned working state without ever learning Git.” You can snapshot a permanent checkpoint, roll back to any point in time, and pick the work up on another machine after a crash. Tensorlake suggests uses like keeping a coding agent’s session memory, for example a Claude Code configuration folder, snapshotted after each phase.
- Hosted Git. Works as ordinary Git, but agents can also mount a repository into a sandbox without cloning it, work in a private workspace, and promote the result to a branch. Merges and rebases run on the server, and Tensorlake says it handles hundreds of pushes per second per repository and charges for the paths that changed rather than the size of the repo.
Orchestration and documents
Beyond sandboxes, Tensorlake runs what it calls Applications: durable workflows that let agents “survive system crashes and pick up where they left off,” as SiliconANGLE described when the platform debuted in March 2026. At that launch, Tensorlake emphasized a “batteries-included” platform with request queuing, autoscaling, secure sandboxes and APIs for extracting data from PDFs, handwriting and charts, deployable on AWS, Azure or Google Cloud. Its open-source OpenIngest project still provides document AI workflows on Tensorlake. Since then, its site has put sandboxes, storage and Git at the center.
Speed claims
Tensorlake publishes a file-system benchmark on GitHub. In its test of 100,000 SQLite inserts on sandboxes with 2 vCPUs and about 4 GB of memory, Tensorlake took 2.45 seconds, against 3.00 for Vercel, 3.92 for E2B, 4.66 for Modal and 5.51 for Daytona. It also reports an 84-millisecond cold start. These are Tensorlake’s own runs, though the code is public to reproduce.
Pricing
Tensorlake’s pricing runs on compute units, where one unit is $0.01:
| Plan | Price | Includes |
|---|---|---|
| Free | $0 | 1 sandbox at a time, 1 vCPU, 1 GB RAM, sessions up to 2 hours |
| Credits | $5 to $20 prepaid packs | Up to 100 concurrent sandboxes, 4 vCPU and 16 GB RAM, sessions up to 24 hours |
| Pro | $250 a month, including $250 of usage | 1,000 concurrent sandboxes, up to 16 vCPU and 64 GB RAM, unlimited duration, lower rates |
| Enterprise | Custom | Unlimited concurrency, HIPAA, single sign-on, in-VPC or on-premises deployment |
On credits, active CPU costs $0.07 per core-hour, memory $0.015 per GB-hour and disk $0.0002 per GB-hour, with lower Pro rates. CPU is billed on active use, with a 20% minimum, and network egress is free. Reserved H100 GPUs cost $3.95 per GPU-hour. Tensorlake is SOC 2 Type 2 compliant.
The company
Tensorlake’s founder and CEO is Diptanu Gon Choudhury. “AI agents are becoming pervasive because modern models can reason and take action, but the infrastructure to run them reliably hasn’t kept up,” he told SiliconANGLE in March 2026. The company had raised $8.6 million by then and named Sixt, the car-rental company, plus major U.S. utilities and financial institutions among its customers. Its homepage also shows Hercules and Core Automation. We didn’t find independent reviews.
How Tensorlake compares
Tensorlake competes with E2B and Daytona, the best-known agent sandbox platforms, as well as Modal and Vercel’s sandboxes. Its distinguishing traits are the versioned filesystem, hosted Git built for agents, durable workflows and document APIs in the same platform, and strong published file-system performance. E2B has the larger customer base and an open-source runtime; Daytona offers Windows and macOS desktops and GPUs. For desktop computer use, Orgo and Cua are more focused.
Sandboxes are for agents that write and run their own code. For everyday work in the apps you already use, Carly connects to thousands of apps directly, so from ChatGPT, Claude or a text message it can sort email, book meetings and update your CRM.
Who Tensorlake suits
Tensorlake is a good fit if:
- Your agents produce files or code you can’t afford to lose between runs.
- You run coding agents, RL rollouts, evaluations or agent-driven CI at scale.
- You want Git and storage that agents use natively, without full clones.
Look elsewhere if:
- You need Windows or macOS desktops for computer use.
- You want an open-source sandbox runtime to self-host.
- You want a finished assistant rather than infrastructure.
FAQ
What is Tensorlake?
Tensorlake is a cloud for AI agents. It provides Firecracker microVM sandboxes that start in under a second, plus a versioned filesystem, hosted Git and durable workflows so agents can work safely and keep what they produce.
How much does Tensorlake cost?
Tensorlake has a free tier, prepaid credit packs from $5, and a Pro plan at $250 a month that includes $250 of usage. Active CPU costs $0.07 per core-hour on credits.
Is Tensorlake the same as Tensorlake document AI?
It’s the same company. Tensorlake launched its platform with document processing and durable workflows for AI and has since put sandboxes, storage and Git for agents at the center. Its open-source OpenIngest project provides document AI workflows on Tensorlake.
How does Tensorlake compare to E2B?
Both run AI agents in Firecracker microVM sandboxes. E2B is larger and has an open-source runtime; Tensorlake adds a versioned filesystem, hosted Git and durable workflows, and reports faster file-system performance in its own benchmark.
Who founded Tensorlake?
Diptanu Gon Choudhury founded Tensorlake and is its CEO. The company had raised $8.6 million as of March 2026.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
