tensorlake

 ![Tensorlake — sandbox-native cloud for AI agents](https://github.com/user-attachments/assets/ac9adfc2-53cb-427e-ad6a-91394cdee961) 

Build agents with sandboxes and serverless orchestration runtime

Tensorlake is a compute infrastructure platform for building agentic applications with sandboxes.

The Sandbox API creates MicroVM sandboxes which you can use to run agents, or use them as an isolated environment for running tools or LLM generated code.

In addition to stateful VMs, you can also add long running orchestration capabilities to Agents using a serverless function runtime with fan-out capabilities.

Sandboxes

Tensorlake Sandboxes provide instant, stateful execution environments for AI agents — spin up millions of sandboxes with near-SSD filesystem performance.

Key capabilities

• Fastest Filesystem I/O — Block-based storage achieving near-SSD speeds inside virtual machines. In SQLite benchmarks (2 vCPUs, 4 GB RAM), Tensorlake completes in 2.45s vs Vercel 3.00s (1.2×), E2B 3.92s (1.6×), Modal 4.66s (1.9×), and Daytona 5.51s (2.2×).
• Fast startup — Sandboxes created in under a second via Lattice, a dynamic cluster scheduler.
• Snapshots & cloning — Snapshot at any point to create durable memory and filesystem checkpoints; clone running sandboxes instantaneously across machines.
• Auto suspend/resume — Sandboxes suspend when idle and resume in under a second without losing any memory or filesystem state.
• Live migration — Sandboxes automatically move between machines during updates with only a brief pause of a few seconds.
• Scale — Supports up to 5 million sandboxes in a single project.

Python SDK Installation

pip install tensorlake

CLI Installation

The tl CLI is distributed as a standalone binary, not through PyPI or npm. Install it with the install script:

curl -fsSL https://tensorlake.ai/install | sh

Setup

Sign up at cloud.tensorlake.ai and get your API key.

export TENSORLAKE_API_KEY="your-api-key"
tl login

Work in More Than One Project (CLI)

tl login saves a context named default. A context is one organization, one project, and one token. Each token comes from its own browser login and works for that one project only. Log in once for each project you use, then switch between the saved contexts with no new login:

tl login --context staging                        # browser login; pick the staging project there
tl context list                                   # the current context is marked with *
tl context use staging                            # switch; no network
tl --context default sbx ls                       # run one command in another context
TENSORLAKE_CONTEXT=staging tl sbx ls
tl context delete staging                         # forget its token and remove it
tl logout                                         # forget the token of the current context
tl logout --all                                   # forget every saved token

With --context or TENSORLAKE_CONTEXT set, tl login saves into that context, so TENSORLAKE_CONTEXT=staging tl login creates or refreshes staging. tl login, tl logout --all, and tl context run even when the named context does not exist.

tl profile is an alias for tl context. tl whoami shows the active context and where it came from. Contexts live in ~/.config/tensorlake/contexts.toml (no secrets). Tokens go to the OS keychain: the macOS Keychain, the Windows Credential Manager, or the Secret Service on Linux. Where there is no keychain, for example in CI or in a container, the token goes to credentials.toml, which only you can read. Set TENSORLAKE_TOKEN_STORAGE=file to keep tokens in that file, for example over SSH to a Mac. tl context list shows where each token is.

Create Your First Sandbox (CLI)

Create a sandbox, run a command, and clean up:

# Create a sandbox (one blocking request; a timeout leaves it queued, never cancels it)
tl sbx create

# Queue a sandbox and return at once, then collect it later
tl sbx create --queue
tl sbx wait  --timeout 1800

# Run a command inside it
tl sbx exec  -- sh -lc "printf 'Hello from the sandbox!\n'"

# Copy a file into the sandbox
tl sbx cp ./my_script.py :/tmp/my_script.py

# Open an interactive terminal
tl sbx ssh 

# Block all outbound internet access on a running sandbox
tl sbx update  --no-internet

# Allow outbound traffic only to selected destinations
tl sbx update  --network-allow api.example.com

# Remove the network policy and restore unrestricted outbound access
tl sbx update  --clear-network

# Terminate when done
tl sbx terminate 

Omit --image to use Tensorlake's default managed environment. To select a custom environment, pass the name of a registered Sandbox Image; arbitrary Docker image references are not supported.

Create a Sandbox Programmatically

from tensorlake.sandbox import SandboxClient

client = SandboxClient.for_cloud(api_key="your-api-key")

# Create a sandbox and connect to it
with client.create_and_connect() as sandbox:
    # Run a command
    result = sandbox.run("sh", ["-lc", "printf 'Hello from the sandbox!\\n'"])
    print(result.stdout)  # "Hello from the sandbox!"

    # Write and read files
    sandbox.write_file("/tmp/data.txt", b"some data")
    content = sandbox.read_file("/tmp/data.txt")

    # Start a long-running process
    proc = sandbox.start_process("sleep", ["300"])
    print(proc.pid)

# Sandbox is automatically terminated when the context manager exits

create_and_connect() (and Sandbox.create()) send the create with wait=False, so the server acknowledges the sandbox at once, and then poll it until it runs. If the wait runs out, the sandbox is not deleted: a SandboxPending is raised with its sandbox_id, and the sandbox keeps its place in the queue until capacity arrives. Collect it later with Sandbox.connect(sandbox_id), or request sandboxes ahead of capacity with Sandbox.create(..., wait=False), which returns a PendingSandbox handle, and call pending.ready() on each one when you need it. Pass cancel_on_timeout=True for the previous delete-then-raise behaviour, or max_pending_secs to let the server give up after a bound. See wait-free create and readiness by polling.

from tensorlake.sandbox import Sandbox, SandboxPending

pending = Sandbox.create(
    name="job-17", image="tensorlake/ubuntu-minimal", max_pending_secs=45 * 60, wait=False
)
try:
    sandbox = pending.ready(timeout=30)
except SandboxPending as still:
    print(still.pending_reason)  # still queued; call ready() again

Running CPU-only CAS sandboxes support live CPU, memory, and root-disk updates through tl sbx update and every Sandbox SDK. Running GPU CAS sandboxes support root-disk growth; CPU and memory changes are rejected. Non-CAS sandboxes do not support live resize. Resource arguments match create, and updates wait for confirmed completion by default. See sandbox resource resize.

Registry credentials for image builds and imports

Supply Docker config.json contents per operation when importing private images or building Dockerfiles with private base images. Each call keeps its own configuration, so concurrent calls can use different credentials without changing DOCKER_CONFIG.

import { readFile } from "node:fs/promises";
import { createSandboxImage, importSandboxImage } from "@tensorlakeai/tensorlake";

const dockerConfigJson = await readFile("/credentials/config.json", "utf8");
await importSandboxImage("registry.example.com/team/image:latest", { dockerConfigJson });
await createSandboxImage("./Dockerfile", { dockerConfigJson });
// Image.build({ dockerConfigJson }) accepts the same option.

from pathlib import Path
from tensorlake import import_sandbox_image
from tensorlake.image.sandbox_builder import build_sandbox_image

config_json = Path("/credentials/config.json").read_text(encoding="utf-8")
import_sandbox_image("registry.example.com/team/image:latest", docker_config_json=config_json)
build_sandbox_image("./Dockerfile", docker_config_json=config_json)
# image.build(docker_config_json=config_json) accepts the same option.

Omitting the option preserves local Docker-config discovery. Configuration is captured and credentials are resolved at the start of the build/import, before provisioning. Supplied contents replace discovery entirely; pass "{}" to use no registry credentials. Empty strings and malformed configurations are errors. Credential helpers (credHelpers and credsStore) run locally using the captured configuration; only resolved registry credentials are forwarded to the builder. Both CAS and non-CAS image builds use this behavior.

The CLI captures Docker configuration immediately after parsing a build/import or deploy command, before authentication and application loading. It looks for $DOCKER_CONFIG/config.json, then ~/.docker/config.json, and passes the captured contents to every image operation in that invocation. A missing config supplies an explicit empty document. Credential changes during a running invocation do not cause the CLI to reread the config file.

Python deployers that advertise pipe support receive this snapshot through a dedicated anonymous pipe before loading application code. Credentials are never written to temporary files, command-line arguments, or environment variables, and stdin remains available to the application. Older protocol-v1 deployers continue using their existing Docker-config discovery behavior.

Rust callers must now supply the additional Option argument: build_sandbox_image(options, docker_config_json, emit) and import_sandbox_image(options, docker_config_json, emit). Use None for discovery or Some(contents) for explicit configuration. Python and TypeScript callers that omit the option continue to work unchanged.

Snapshots

Save the state of a sandbox and restore it later:

# Snapshot a running sandbox
snapshot = client.snapshot_and_wait(sandbox_id)

# Later, create a new sandbox from the snapshot
with client.create_and_connect(snapshot_id=snapshot.snapshot_id) as sandbox:
    # Picks up right where you left off
    result = sandbox.run("ls", ["/tmp"])
    print(result.stdout)

Sandbox Pools

Pre-warm containers for fast startup:

from tensorlake.sandbox import FileSystemMount, NetworkConfig

# Create a pool with warm containers and no internet access
pool = client.create_pool(
    image="tensorlake/ubuntu-minimal",
    warm_containers=3,
    network=NetworkConfig(allow_internet_access=False),
)

# Claim a sandbox instantly from the pool. Claim-specific file systems are
# mounted before the sandbox is reported as running.
resp = client.claim(
    pool.pool_id,
    file_systems=[
        FileSystemMount(
            file_system_id="file_system_abc",
            mount_path="/mnt/skills",
        )
    ],
)
sandbox = client.connect(resp.sandbox_id)

# Named sandboxes can be reconnected later by name
named = client.create(name="stable-name")
sandbox = client.connect("stable-name")

Pool root disks default to the registered image's size. Pass disk_mb to grow a filesystem-only image; a pool disk cannot be smaller than its image.

GPU CAS pools allocate GPUs per container. Use a CAS image and pass gpus=1 with gpu_model="A10" (or gpu=GpuRequest(count=1, model=GpuModel.A10)) to create_pool and update_pool. The asynchronous Python API accepts the same arguments; TypeScript uses gpus, gpuModel, or gpu in its pool options. Passing only gpu_model / gpuModel allocates one GPU per container for both standalone sandbox creation and pool APIs, matching tl sbx create --gpu. Claimed sandboxes inherit the pool's GPU allocation. Include that allocation on each pool update; omitting it configures a CPU-only CAS pool. Changing the allocation recycles unclaimed warm containers while claimed sandboxes keep their original allocation. Set max_containers=1 and warm_containers=1 for a single-GPU pool; it can replenish warm capacity after the claim releases its GPU.

Pool responses expose the allocation as pool.resources.gpu_configs in Python and Rust, and pool.resources.gpuConfigs in TypeScript. Python response allocations use GpuAllocation, whose model preserves the server's exact identifier, such as H100-PCIe-80GB. Typed requests use GpuRequest with a supported GpuModel. When updating a pool, pass the reported count and the corresponding supported request model.

Inspect a claimed sandbox's GPU allocation with tl sbx describe or the GPUs column in tl sbx ls. Both show count and model, such as 2 x H100; sandboxes without a GPU allocation show -. Archived sandboxes show the same information with tl sbx ls --archived.

Set the pool network policy when you create the pool, replace it later with a pool update, or pass CLEAR_NETWORK_POLICY (Python) / null (TypeScript) to remove it. On a change the service recycles the pool's unclaimed warm containers onto the new policy, while containers already claimed by sandboxes keep the policy they booted with.

Sandbox startup errors include the server's reason and actionable diagnostic. The Python and TypeScript RemoteAPIError exposes structured failure fields; the CLI shows the same details in create/copy errors and tl sbx describe. See sandbox failure diagnostics for error handling examples.

Cloud Volumes

FilesystemClient manages durable, versioned file trees without mounting them. SDK writes hash files locally, upload missing 64 MiB parts directly to checksum-bound object-store URLs, and then atomically publish metadata. Reads resolve an authenticated immutable plan and fetch the selected records directly from signed object-store URLs. File payloads normally do not pass through the Tensorlake API service.

import { FilesystemClient } from "@tensorlakeai/tensorlake";

const client = new FilesystemClient({
  apiKey: "your-api-key",
});
const fs = await client.create("agent-artifacts");

// In-memory data; all changes become visible atomically.
await fs.writeFiles({
  "run/config.json": JSON.stringify({ model: "gpt-5" }),
  "run/input.txt": "hello",
});

// Multi-GiB files use bounded memory and stream directly to object storage.
await fs.writeFileFromPath("models/weights.bin", "./weights.bin");

// These reuse immutable content references and transfer no payload bytes.
await fs.copyFile("run/input.txt", "run/input-copy.txt");
await fs.moveFile("run/config.json", "archive/config.json");

// One request returns only the selected bytes plus full-file identity/size.
const read = await fs.readFileWithMetadata("models/weights.bin", {
  range: { offset: 0, length: 1024 * 1024 },
});

// Snapshot retention and forks are metadata-only operations.
const snapshot = await fs.snapshot("ready for evaluation");
const fork = await client.fork("agent-artifacts-eval", fs.name, snapshot.id);
await fork.writeFile("results/score.txt", "0.98");

await fs.deleteSnapshot(snapshot.id);

Choose storage placement when creating a filesystem in an enabled region. The API endpoint stays the same; uploads and read plans use the selected region's object store:

from tensorlake.filesystem import FilesystemClient

client = FilesystemClient(api_key="your-api-key")
fs = client.create("frankfurt-artifacts", region="eu-central-1")

const fs = await client.create("frankfurt-artifacts", { region: "eu-central-1" });

Rust callers use artifact_storage.create_filesystem(project, name, Some("eu-central-1")). Omitting region keeps the legacy location. Placement is fixed for the filesystem's storage network: forks inherit it, and changes to a project's region do not move existing files. An unsupported region is rejected; the SDK never silently substitutes another region. The compatibility create_file_system(..., region="eu-central-1") and createFileSystem(name, description, { region: "eu-central-1" }) helpers accept the same choice.

writeFile() and writeFiles() accept bytes already in memory. Prefer writeFileFromPath() or writeFilesFromPaths() for large local files so neither JavaScript nor Rust retains the complete payload. A successful write is durable before it returns, but only the live head is retained automatically; snapshot() pins the current head permanently in one client/server round trip. readFileWithMetadata() returns immutable content identity and total size with the bytes; its optional range downloads only overlapping immutable records. The SDK bounds every response, verifies stored and logical content checksums, and falls back to the authenticated service only when the object store cannot honor the signed range. Deleting a snapshot releases that retention root, while bytes still reachable from a live head, another snapshot, a fork, or a mount remain durable.

Orchestrate

Create orchestration APIs on a distributed runtime with automatic scaling, fan-out capabilities and built-in tracking. The orchestration APIs can be invoked using HTTP requests or using the Python SDK.

Quickstart

Decorate your entrypoint with @application() and functions with @function(). Each function runs in its own isolated sandbox.

Example: City guide using OpenAI Agents with web search and code execution:

from agents import Agent, Runner
from agents.tool import WebSearchTool, function_tool
from tensorlake.applications import application, function, Image

# Define the image with necessary dependencies
FUNCTION_CONTAINER_IMAGE = Image(base_image="python:3.11-slim", name="city_guide_image").run(
    "pip install openai openai-agents"
)

@function_tool
@function(
    description="Gets the weather for a city using an OpenAI Agent with web search",
    secrets=["OPENAI_API_KEY"],
    image=FUNCTION_CONTAINER_IMAGE,
)
def get_weather_tool(city: str) -> str:
    """Uses an OpenAI Agent with WebSearchTool to find current weather."""
    agent = Agent(
        name="Weather Reporter",
        instructions="Use web search to find current weather in Fahrenheit for the city.",
        tools=[WebSearchTool()],  # Agent can search the web
    )
    result = Runner.run_sync(agent, f"City: {city}")
    return result.final_output.strip()

@application(tags={"type": "example", "use_case": "city_guide"})
@function(
    description="Creates a guide with temperature conversion using function_tool",
    secrets=["OPENAI_API_KEY"],
    image=FUNCTION_CONTAINER_IMAGE,
)
def city_guide_app(city: str) -> str:
    """Uses an OpenAI Agent with function_tool to run Python code for conversion."""

    @function_tool
    def convert_to_celsius_tool(python_code: str) -> float:
        """Converts Fahrenheit to Celsius - runs as Python code via Agent."""
        return float(eval(python_code))

    agent = Agent(
        name="Guide Creator",
        instructions="Using the appropriate tools, get the weather for the purposes of the guide. If the city uses Celsius, call convert_to_celsius_tool to convert the temperature, passing in the code needed to convert the temperature to Celsius. Create a friendly guide that references the temperature of the city in Celsius if the city typically uses Celsius, otherwise reference the temperature in Fahrenheit. Only reference Celsius or Fahrenheit, not both.",
        tools=[get_weather_tool, convert_to_celsius_tool],  # Agent can execute this Python function
    )
    result = Runner.run_sync(agent, f"City: {city}")
    return result.final_output.strip()

Deploy to Tensorlake

1. Set your API keys:

export TENSORLAKE_API_KEY="your-api-key"
tl secrets set OPENAI_API_KEY "your-openai-key"

2. Deploy:

tl deploy examples/readme_example/city_guide.py

tl deploy builds a version-scoped runtime image for every function and stores its registered sandbox-template name in the application manifest. The image contains the dependencies declared by Image, the language runner, and the native function-agent core. Application source remains a separate ZIP that is uploaded with the manifest; it is not copied into the runtime image. Reusing an application version with different code or configuration is rejected; an identical redeploy is idempotent. Image and application requests normally route through TENSORLAKE_API_URL so ingress can authenticate them and attach trusted project identity. Application deployment, invocation, output, listing, and inspection all use that same origin. Trusted split local installations can set TENSORLAKE_IMAGE_SERVICE_URL and TENSORLAKE_FUNCTION_SERVICE_URL independently.

Sandbox create, list, snapshot and pool requests go to the sandbox API, which is derived from TENSORLAKE_API_URL (api.tensorlake.ai becomes sandbox.tensorlake.ai). Set TENSORLAKE_SANDBOX_API_URL to an absolute http(s) origin (no query string, fragment or credentials) to send them elsewhere, for example to a dedicated BYOC proxy such as https://byoc-.tensorlake.ai. The CLI, the Python SDK and the TypeScript SDK all honor it; tl whoami --output json shows the resolved value as sandboxApi. TENSORLAKE_SANDBOX_PROXY_URL is different: it only affects per-sandbox connections when the server did not return a sandbox URL.

Call via HTTP

# Invoke the application
curl https://api.tensorlake.ai/applications/city_guide_app \
  -H "Authorization: Bearer $TENSORLAKE_API_KEY" \
  --json '"San Francisco"'
# Returns: {"request_id": "beae8736ece31ef9"}

# Get the result
curl https://api.tensorlake.ai/applications/city_guide_app/requests/{request_id}/output \
  -H "Authorization: Bearer $TENSORLAKE_API_KEY"

# Stream results with SSE
curl https://api.tensorlake.ai/applications/city_guide_app \
  -H "Authorization: Bearer $TENSORLAKE_API_KEY" \
  -H "Accept: text/event-stream" \
  --json '"San Francisco"'

FAQ

What is Tensorlake? Tensorlake is the sandbox-native cloud for AI agents — a compute platform for securely running untrusted, LLM-generated code in isolated sandboxes and orchestrating agentic applications at scale.

How do I run untrusted or LLM-generated code safely? Each Tensorlake sandbox provides an isolated environment for running untrusted or LLM-generated code. Create a CPU-only CAS or GPU CAS sandbox with the Python or TypeScript SDK, or the CLI, in a few lines.

How is Tensorlake different from E2B, Modal, or Daytona? Tensorlake is built for heavy filesystem I/O, fast startup, and large-scale fan-out. In SQLite benchmarks (2 vCPUs, 4 GB RAM) it completes in 2.45s versus E2B (3.92s), Modal (4.66s), and Daytona (5.51s), and it supports snapshots, auto suspend/resume, live migration, and up to 5 million sandboxes per project.

Can I checkpoint and resume an AI agent? Yes. Snapshot a running sandbox at any point to capture both memory and filesystem state, then create a new sandbox from that snapshot to pick up exactly where you left off. Sandboxes also auto-suspend when idle and resume in under a second without losing state.

How fast do sandboxes start? Sandboxes are created in under a second via Lattice, a dynamic cluster scheduler. For even faster starts, use sandbox pools to keep warm containers ready to claim instantly.

How do I run code interpreter / tool execution for an LLM agent? Spin up a sandbox as an isolated execution environment for an agent's tools or generated code, run commands or processes inside it, read and write files, and terminate it when done — all from the Python or TypeScript SDK, or the CLI.

What languages and interfaces are supported? Tensorlake provides a Python SDK, a TypeScript SDK, and a standalone CLI (tl), plus an HTTP API for invoking orchestration applications.

How do I get started? Sign up at cloud.tensorlake.ai, run pip install tensorlake for the Python SDK, install the CLI with curl -fsSL https://tensorlake.ai/install | sh, set your TENSORLAKE_API_KEY, and create your first sandbox. See the documentation for full guides.

Learn More

• Sandbox Documentation
• Orchestrate Documentation
• tl git mount repository and subtree workflow

GPU model names

Sandbox GPU requests also accept RTX-PRO-6000 and L40 in the Python, TypeScript, and Rust SDKs and the CLI --gpu option. Python exposes GpuModel.RTX_PRO_6000 and GpuModel.L40. RTX-PRO-6000 selects the 96 GB Blackwell Server Edition, not the workstation editions. These names request exact models; L40 does not select L40S. The server and dataplane must support the requested model and have matching capacity.