How much does Tensorlake cost?

Free plan with one 1 vCPU sandbox and 2 hour sessions. Usage Credits from $5 prepaid packs at $0.01 per compute unit, with optional auto refill. Pro is $250 a month including 25,000 compute units, 1,000 concurrent sandboxes and lower rates, with overage at $0.01 per unit. Enterprise is quoted with HIPAA, SSO, RBAC, in VPC or on premises deployment and a one hour P1 SLA. Reserved H100 GPUs are $3.95 per GPU hour; egress is free.

Does Tensorlake support MCP and tool calling?

Full under the Agentic Index verification standard. Tensorlake publishes public evidence of Model Context Protocol (MCP) support, tool calling, and API or SDK extensibility for connecting agents to external tools. Evidence basis: Every part of Tensorlake is programmable. Python and TypeScript SDKs, including an async Python client, cover sandbox creation, commands, files, snapshots, pools and resizing; the tl CLI manages sandboxes, filesystems and projects; a REST A. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Does Tensorlake have SOC 2, SSO and role based access control?

Full under the Agentic Index verification standard. Tensorlake publishes public evidence of enterprise security and identity controls, covering a compliance posture such as SOC 2 or ISO 27001, single sign on (SSO or SAML), and role based access control (RBAC). Evidence basis: Organizations hold projects, and access runs through admin and member roles that apply to dashboard users and API keys alike; each API key lives inside one project with member permissions and never reaches organization settings. Enterprise. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Can Tensorlake be self hosted or run in a private VPC?

Full under the Agentic Index verification standard. Tensorlake publishes public evidence of self hosted or private VPC deployment, along with data residency options for controlling where agent data is processed and stored. Evidence basis: Sandboxes can run in the customer's own AWS account: with bring your own cloud, executors run on the customer's Amazon EC2 so workloads and data stay there, and Enterprise extends this to deployment inside the customer's VPC or on premises. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Back to vendors

Tensorlake

Q: What are the best alternatives to Tensorlake?

Based on documented capability coverage across the Agentic Index 14 point taxonomy, the closest alternatives to Tensorlake among agent infrastructure platforms are DBOS, Hatchet, Temporal, Daytona. Each is graded on the same public evidence, and no vendor pays for placement.

Also known as: Tensorlake, Inc., tensorlake.ai, Tensorlake Sandboxes, Tensorlake Cloud

Visit site

Agent infrastructureindependentVerified 2026-11-09

Entry priceFree to start. Prepaid credit packs from $5 at $0.01 per compute unit, or Pro at $250 a month including $250 of usage.Full pricing detail

Tensorlake sells infrastructure for running AI agents: Firecracker microVM sandboxes that give an agent its own computer or isolate a single tool call and sleep and wake with their state, a versioned filesystem that mounts like a disk with snapshots and time travel, hosted Git built for agents, and a serverless orchestration runtime with durable execution and cron schedules. Python and TypeScript SDKs, a CLI and REST APIs drive it, with a free tier, prepaid credits from $5 and Pro at $250 a month.

Tensorlake, Inc. builds composable infrastructure for teams that deploy agent harnesses, run reinforcement learning rollouts and evaluations, and build CI at scale, under terms governed by California law. Its sandboxes serve as the execution layer behind other agent products: guides run Claude Managed Agents, Cursor Cloud Agents, Devin Outposts, OpenCode and the Crabbox test runner on Tensorlake machines, and the Harbor evaluation framework uses them for rollouts.

Each sandbox is a Firecracker microVM with its own filesystem, shell, packages and processes. A persistent sandbox acts as an agent's computer for sessions of hours or days, while an ephemeral one isolates a single tool call from the harness. Sandboxes suspend when idle and resume with their filesystem, memory and running processes, can be snapshotted and cloned, resized while running, kept warm in pools, and can run Docker, nested virtual machines or GPU workloads with CUDA, vLLM and SGLang. Each running sandbox gets a public URL for the ports a team exposes, and its outbound internet access can be restricted.

The ubuntu-vnc image boots an XFCE desktop with Firefox that an agent drives with screenshots, mouse and keyboard, and Chrome can be driven over CDP. Cloud Volumes are versioned POSIX directories that autosave settled writes to durable storage, take permanent snapshots and restore to an earlier point, with read only mounts for handing skills and tools to a fleet. Hosted Git scales to millions of repositories and mounts a repository into a sandbox without a clone. Orchestration turns Python functions into serverless applications with durable execution that skips completed steps on retry, futures, map reduce, parallel specialist subagents, cron schedules and built in tracing.

The Free plan runs one sandbox at a time with 1 vCPU, 1 GB of RAM and 10 GB of disk for sessions up to 2 hours. Usage Credits start with a $5 prepaid pack at $0.01 per compute unit and allow 100 concurrent sandboxes and 24 hour sessions. Pro costs $250 a month with 25,000 compute units included, 1,000 concurrent sandboxes, up to 16 vCPU and 64 GB of RAM, unlimited session length and rates about 40 percent lower, with overage at $0.01 per unit.

Enterprise is quoted and adds unlimited concurrency, HIPAA, single sign on with SAML, role based access, a DPA, deployment in the customer's VPC or on premises and a one hour P1 SLA. Active CPU costs $0.07 per core hour on credits and $0.042 on Pro, reserved H100 GPUs cost $3.95 per GPU hour, and network egress is free. Organizations use admin and member roles with project scoped API keys, Tensorlake holds SOC 2 Type 2, and sandboxes can run inside the customer's own AWS account.

Tensorlake fits teams building coding agents, agent harness platforms and reinforcement learning or evaluation pipelines that need thousands of isolated machines that keep their state, and agents that need a desktop to operate. It fits less well for buyers looking for a finished agent with approval workflows, a knowledge base for retrieval, or scoring of agent quality, since it supplies the compute, storage and orchestration rather than those judgments.

Vendor details

Canonical URL

https://www.tensorlake.ai

Subcategory

Firecracker microVM sandboxes, versioned filesystems, hosted Git and serverless orchestration for agents

Funding status

Privately held. Tensorlake, Inc.; no funding rounds are named on the pages read.

Company status

independent

Use cases & customers

Primary use cases

giving each agent its own persistent computer that sleeps when idle and resumes with its stateisolating LLM generated code and tool calls in short lived microVMskeeping agent work in versioned filesystems and hosted Git with snapshots and rollbackrunning reinforcement learning rollouts, Harbor evaluations and CI jobs on parallel sandboxesorchestrating agent workflows with durable execution, parallel subagents and cron schedules

Target customers

teams building coding agents and agent harness platforms that need an isolated computer per agent or per tool callAI labs and research teams running reinforcement learning rollouts and evaluations on thousands of reproducible machinesplatform teams that want CI and GitHub Actions runners on fast microVMsbuilders of computer use agents that need a remote desktop or Chrome to drive

Deployment options

Tensorlake CloudBring your own cloud on AWS, with executors on the customer's EC2In VPC or on premises on Enterprise

Integrations

Python and TypeScript SDKs, the tl CLI and REST APIs for sandboxes, filesystems and Git; guides for Claude Managed Agents, Cursor Cloud Agents, Devin Outposts, OpenCode, Crabbox and Harbor; GitHub Actions runners; signed webhooks through Svix; agent skills for Claude Code and Codex.

In practice

A coding agent company gives each user session its own Tensorlake sandbox with the repository mounted from hosted Git. The sandbox sleeps when the user goes quiet and wakes with its processes intact the next morning, and every settled change autosaves to a versioned volume so a crashed session restarts from where it stopped.

A research team runs Harbor evaluations and reinforcement learning rollouts across hundreds of sandboxes started from one snapshot, so every rollout begins from an identical machine, and a reserved H100 sandbox serves the policy model with vLLM.

An operations team deploys a nightly research agent as a Tensorlake application on a cron schedule. It fans out to parallel specialist subagents, each step's output is persisted so a failure retries without redoing finished work, and the execution timeline shows which step took the longest.

Sources & related URLs

Related / legacy domains

https://docs.tensorlake.ai https://www.tensorlake.ai/pricing https://trust.tensorlake.ai https://github.com/tensorlakeai/tensorlake https://cloud.tensorlake.ai

Research sources

https://www.tensorlake.ai https://www.tensorlake.ai/pricing https://www.tensorlake.ai/terms https://www.tensorlake.ai/dpa https://docs.tensorlake.ai/llms.txt https://docs.tensorlake.ai/platform/access-control https://docs.tensorlake.ai/platform/authentication https://docs.tensorlake.ai/sandboxes/computer-use https://docs.tensorlake.ai/sandboxes/networking https://docs.tensorlake.ai/applications/parallel-sub-agents https://docs.tensorlake.ai/applications/cron-scheduler

Agentic Index coverage score

8.5 / 14 capabilities · 61%

Integrations & Tool Calling	Partial
An agent in a Tensorlake sandbox reaches outside systems through whatever runs inside the machine: Git pushes to hosted or outside repositories, the packages and clients the team installs, and outbound internet access the project can restrict. Secrets and environment variables carry the credentials. Tensorlake itself offers no connector catalog, MCP gateway or action framework for business apps, so which systems an agent can act in, and how it authenticates, sits in the customer's own harness code. Sourcedocs.tensorlake.ai/sandboxes/networkingread 2026-11-09
Workflow Orchestration	Full
Tensorlake Orchestration turns Python functions into serverless applications that run multistep agent workflows. Futures run calls in parallel, map reduce spreads work across many inputs, and a parallel subagents pattern fans a task out to specialist agents and combines their results. Durable execution persists each function's output, so a retry or replay skips finished steps and a crashed agent resumes without losing work, and agents can drive sandboxes from inside the workflow or run inside one. Sourcedocs.tensorlake.ai/applications/parallel-sub-agentsread 2026-11-09
Knowledge Grounding & RAG	Not documented
Tensorlake stores files for agents rather than retrieving knowledge for them. Read only mounts hand versioned manuals, skills, configs and binary tools to a whole fleet, pinned to a snapshot or following the latest version, and hosted Git holds code and documents agents produce. None of it is indexed for search or embedded for retrieval, so grounding an agent in a company's documents means the team builds and runs its own retrieval inside the sandbox. Sourcedocs.tensorlake.ai/filesystems/read-only-mountsread 2026-11-09
Human Oversight & Guardrails	Not documented
There is no approval step, consent checkpoint or escalation that holds an agent's action until a person signs off. A project can restrict a sandbox's outbound internet access, choose which ports are exposed and whether they need authentication, and cap CPU, memory, disk, concurrency and session length, so the team sets the boundaries that any code in the box runs inside. A person can also open an SSH or PTY session into a running sandbox to watch the work or step in by hand. Sourcedocs.tensorlake.ai/sandboxes/networkingread 2026-11-09
Security, Identity & Governance	Full
Organizations hold projects, and access runs through admin and member roles that apply to dashboard users and API keys alike; each API key lives inside one project with member permissions and never reaches organization settings. Enterprise adds single sign on with SAML, role based access and a DPA, and HIPAA coverage. Tensorlake holds SOC 2 Type 2 across every plan, keeps storage encrypted, and runs each sandbox as its own Firecracker microVM isolated from the harness and from other tenants. Sourcedocs.tensorlake.ai/platform/access-controlread 2026-11-09
Observability & Auditability	Full
Agent workflows deployed on Orchestration come with built in tracing, execution timelines and monitoring of each function call, alongside logs the application emits. Inside sandboxes, stdout and stderr from every process can be searched and filtered in the console, and usage and cost are reported per sandbox. On hosted Git, commit history carries landing attribution, so each change and each line traces back to the workspace and commit that produced it. Sourcedocs.tensorlake.ai/applications/observabilityread 2026-11-09
Memory & State Persistence	Partial
State stays with the machine. A sandbox suspends when idle and resumes with its filesystem, memory and running processes, and snapshots save and clone that whole state. Cloud Volumes autosave every settled write to durable storage, keep permanent snapshots and restore to an earlier point, and Orchestration persists each step's output so a run picks up where it failed. No feature gives an agent a memory store with a stated scope and lifetime that it reads as context across separate sessions. Sourcedocs.tensorlake.ai/sandboxes/snapshotsread 2026-11-09
Deployment & Data Residency	Full
Sandboxes can run in the customer's own AWS account: with bring your own cloud, executors run on the customer's Amazon EC2 so workloads and data stay there, and Enterprise extends this to deployment inside the customer's VPC or on premises. Tensorlake Cloud serves everyone else from its hosted service, and Cloud Volumes mount on any sandbox, cloud VM or Kubernetes pod on Linux or macOS, so files can live where the team's other infrastructure runs. Sourcedocs.tensorlake.ai/platform/self-hosted-compute/awsread 2026-11-09
Prebuilt Agents / Templates / Packs	Partial
Starting points come as parts a team assembles: managed images such as ubuntu-minimal, the ubuntu-vnc desktop and ubuntu-systemd for Docker, warm sandbox pools, a Tensorlake skill for Claude Code and Codex, and step by step guides for running Claude Managed Agents, Cursor Cloud Agents, Devin Outposts, OpenCode and Crabbox on its sandboxes. Example apps cover a code interpreter, an agentic swarm, reinforcement learning fine tuning and an overnight research loop, all code to adapt rather than agents to switch on. Sourcedocs.tensorlake.ai/sandboxes/tensorlake-imagesread 2026-11-09
Triggers & Channel Coverage	Full
Deployed agent applications start on their own: the cron scheduler invokes an orchestration endpoint on a recurring schedule, managed through the API or the Applications UI, and a public endpoint lets outside callers invoke an application without a Tensorlake API key. Workflows scale out and queue as calls arrive. GitHub Actions jobs can run on Tensorlake runners, and signed webhooks report sandbox and application lifecycle events outward to the customer's own systems. Sourcedocs.tensorlake.ai/applications/cron-schedulerread 2026-11-09
Model Flexibility & Routing	Not documented
No model runs inside Tensorlake itself. Sandboxes execute whatever agent the team deploys, from Claude Managed Agents and Cursor Cloud Agents to OpenCode or an OpenAI based code interpreter, and that agent calls its provider with the team's own keys. GPU sandboxes on reserved H100s at $3.95 per GPU hour give a team compute to serve its own open models with vLLM or SGLang, but the serving stack is the team's code rather than a Tensorlake model service. Orchestration, Cloud Volumes and hosted Git carry no inference step either, so the platform has no model of its own for a customer to pick or route. Sourcedocs.tensorlake.ai/sandboxes/gpuread 2026-11-09
APIs / SDKs / MCP Extensibility	Full
Every part of Tensorlake is programmable. Python and TypeScript SDKs, including an async Python client, cover sandbox creation, commands, files, snapshots, pools and resizing; the tl CLI manages sandboxes, filesystems and projects; a REST API accepts the same project API key as a bearer token; Cloud Volumes have their own HTTP API for files, directories and snapshots; and repository SDKs create repositories, push worktrees and merge branches. Sourcedocs.tensorlake.ai/sandboxes/sdk-referenceread 2026-11-09
Testing, Debugging & Optimization	Not documented
Tensorlake supplies the machines evaluations run on rather than the evaluation itself. Harbor evaluations and reinforcement learning rollouts start hundreds of identical sandboxes from one snapshot, and CI and test suites run in isolated microVMs, but the harness, the test cases and the scoring belong to Harbor or to the team's own code. Nothing in Tensorlake scores agent output or compares one version of an agent against another. Sourcedocs.tensorlake.ai/sandboxes/harborread 2026-11-09
Browser / Computer-use	Full
Agents get a real desktop to operate. The managed ubuntu-vnc image boots XFCE, TigerVNC and Firefox, and the SDK connects through the authenticated sandbox proxy so an agent can capture screenshots and send mouse and keyboard input from Python or JavaScript without exposing the VNC port. For browser work, Chrome runs inside the same sandbox and can be driven over the Chrome DevTools Protocol. Sourcedocs.tensorlake.ai/sandboxes/computer-useread 2026-11-09

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Free to start. Prepaid credit packs from $5 at $0.01 per compute unit, or Pro at $250 a month including $250 of usage.

Compute units drawn by active CPU, allocated RAM and disk while a sandbox runs and snapshot storage while it is suspended; Pro adds a monthly platform fee that includes 25,000 units

Free tier

Included quota

Free allows one sandbox at a time and 2 hour sessions; Usage Credits allow 100 concurrent sandboxes and 24 hour sessions; Pro includes 25,000 compute units, 1,000 concurrent sandboxes and unlimited session length.

What is public

Plan limits, credit pack prices, per resource rates for credits and Pro, the reserved H100 rate and an interactive cost calculator.

Billing mechanics

Usage Credits deduct from a prepaid balance; Pro charges $250 each monthly cycle and draws its included units first before billing overage.

Cost watchouts

Running sandboxes bill at least 20 percent of allocated CPU even when idle, suspended snapshots keep billing for storage, and model calls from agents in the sandbox go on the team's own provider accounts.

Variable cost rationale

Every sandbox bills for CPU, memory, disk and snapshot storage by the hour, with idle CPU billed at a 20 percent minimum while a sandbox runs, so spend tracks how many agents run, how large their machines are and how long they stay up.

Additional watchouts

Large fleets of long lived sandboxes can outgrow the Pro allowance quickly, and GPU capacity is billed on a separate invoice.

Overage / add-ons

Usage Credits stop at the prepaid balance with no postpaid overage unless auto refill is on; Pro bills usage past its 25,000 included units at $0.01 per unit at Pro rates.

Sales call required

Mixed (some tiers require a call)

Free / trial

The Free plan runs one sandbox at a time with 1 vCPU, 1 GB of RAM and 10 GB of disk, for sessions up to 2 hours, with no time limit on the plan.

Lowest paid plan

Usage Credits from a $5 prepaid pack of 500 compute units; Pro at $250 a month with 25,000 compute units included

Commercial notes

Tensorlake prices one set of Firecracker microVMs and APIs across every tier, with tiers differing in concurrency, machine size, session length, rates and support.

Key ambiguities

Hosting regions for Tensorlake Cloud are not stated on the pricing page, and Enterprise rates are custom.

Cancellation / refund

Credit packs are prepaid with optional auto refill and no postpaid overage; Pro bills on a monthly cycle.

Support SLA / resale

Free has docs and a community Slack, Pro has 24x7 Slack and email support with a 24 hour P1 SLA, and Enterprise has a 1 hour P1 SLA and a resident solutions architect.

Missing data

Enterprise rates, Cloud hosting regions and contract terms.

Agentic Index verified 2026-11-09

Official pricing page Docs

Related vendors

Alternatives to Tensorlake

The closest documented capability profiles to Tensorlake among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.

DBOS8.5 / 14Adds documented Human Oversight & Guardrails and Testing, Debugging & Optimization
Hatchet8.5 / 14Adds documented Human Oversight & Guardrails and Testing, Debugging & Optimization
Temporal8.5 / 14Adds documented Human Oversight & Guardrails and Testing, Debugging & Optimization
Daytona8.0 / 14Adds documented Model Flexibility & Routing
Modal8.0 / 14Adds documented Model Flexibility & Routing
Northflank9.0 / 14Adds documented Human Oversight & Guardrails and Model Flexibility & Routing

Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded

Compare Tensorlake with other vendors →Compare all agent infrastructure platforms →See where Tensorlake ranks →

Data confidence: medium