Tensorlake Documentation

Run shell commands inside sandboxes with full stdout/stderr capture, real-time streaming, and configurable timeouts — then keep long-running services alive with background processes, signals, stdin, and supervised restarts. Commands and processes share one API. Use run / exec when you want to run something and wait for the result; use start_process when you want to start something that keeps running and manage its lifecycle.

Use Sandbox Process Logs when you want to browse, search, and filter retained stdout/stderr across processes in the Console.

Sandbox-specific operations use sandbox-specific ingress, commonly seen as https://<sandbox-id-or-name>.sandbox.tensorlake.ai, derived from the sandbox’s ingress_endpoint.For named sandboxes, you can use the sandbox name in place of the ID — both in the proxy hostname and in CLI/API commands. For example, https://my-env.sandbox.tensorlake.ai/api/v1/processes and tl sbx exec my-env python main.py work the same as their ID-based equivalents. The proxy resolves the name to the underlying sandbox automatically.The command and process APIs documented here run on the management URL on port 9501, which always requires authentication. Unauthenticated proxy access applies only to exposed user ports.

Basic Execution

CLI
Python
TypeScript
HTTP

# Run in an existing sandbox — use the sandbox ID or name
tl sbx exec my-env python -c 'print("Hello from sandbox!")'

# Or create, run, and tear down in one step
tl sbx run python -c 'print("Hello from sandbox!")'

from tensorlake.sandbox import Sandbox

sandbox = Sandbox.create()
result = sandbox.run("python", ["-c", "print('Hello from sandbox!')"])
print(result.stdout)     # Hello from sandbox!
print(result.exit_code)  # 0

const result = await sandbox.run("python", {
  args: ["-c", "print('Hello from sandbox!')"],
});

console.log(result.stdout);   // Hello from sandbox!
console.log(result.exitCode); // 0

# Start a Python process inside the sandbox
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "command": "python",
    "args": ["-c", "print(\"Hello from sandbox!\")"]
  }'

Response:

{
  "pid": 294,
  "status": "running",
  "exit_code": null,
  "signal": null,
  "stdin_writable": false,
  "command": "python",
  "args": ["-c", "print(\"Hello from sandbox!\")"],
  "started_at": 1773950042728,
  "ended_at": null
}

CLI Options

CLI
Python
TypeScript
HTTP

# Timeout is in seconds
tl sbx exec <sandbox-id> --timeout 10 python -c 'print("hi")'

# Run from a specific working directory
tl sbx exec <sandbox-id> --workdir /workspace python main.py

# Inject environment variables into a single command
tl sbx exec <sandbox-id> --env MODE=prod --env DEBUG=0 /bin/sh -lc 'printf "%s %s\n" "$MODE" "$DEBUG"'

# Keep the sandbox after a one-shot run so you can inspect it afterwards
tl sbx run --keep /bin/sh -lc 'echo KEEP_TEST && sleep 1'

A verified --env run printed prod 0. A verified --keep run ended with Sandbox <id> kept alive., and tl sbx ls --all then showed that sandbox as running.

# Run with a timeout, working directory, and per-command environment
result = sandbox.run(
    "python",
    ["main.py"],
    env={"MODE": "prod", "DEBUG": "0"},
    working_dir="/workspace",
    timeout=10,
)

const result = await sandbox.run("python", {
  args: ["main.py"],
  env: { MODE: "prod", DEBUG: "0" },
  workingDir: "/workspace",
  timeout: 10,
});

console.log(result.exitCode);

# Start a process with custom environment variables and working directory
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "command": "python",
    "args": ["main.py"],
    "env": {"MODE": "prod", "DEBUG": "0"},
    "working_dir": "/workspace"
  }'

Shell Commands

CLI
Python
TypeScript
HTTP

# Count Python files in /workspace with a pipe
tl sbx exec <sandbox-id> bash -c "ls -la /workspace | grep '.py' | wc -l"

# Redirect stdout and stderr to files
tl sbx exec <sandbox-id> bash -c "python script.py > output.txt 2> errors.txt"

# Chain setup and execution in one shell command
tl sbx exec <sandbox-id> bash -c "cd /workspace && pip install -r requirements.txt && python main.py"

# Pipes
result = sandbox.run("bash", ["-c", "ls -la /workspace | grep '.py' | wc -l"])
print(result.stdout)

# Redirects
sandbox.run("bash", ["-c", "python script.py > output.txt 2> errors.txt"])

# Command chaining
sandbox.run("bash", ["-c", "cd /workspace && pip install -r requirements.txt && python main.py"])

const count = await sandbox.run("bash", {
  args: ["-lc", "ls -la /workspace | grep '.py' | wc -l"],
});
console.log(count.stdout);

await sandbox.run("bash", {
  args: ["-lc", "python script.py > output.txt 2> errors.txt"],
});

await sandbox.run("bash", {
  args: [
    "-lc",
    "cd /workspace && pip install -r requirements.txt && python main.py",
  ],
});

# Use bash when you need pipes, redirects, or command chaining
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "command": "bash",
    "args": ["-c", "ls -la /workspace | grep .py | wc -l"]
  }'

Get Process Output

Fetch the buffered stdout, stderr, or combined output of a running or finished process.

Python
TypeScript
HTTP
CLI

sandbox = Sandbox.create()
result = sandbox.run("python", ["-c", "print('hello')"])
print(result.stdout)
print(result.stderr)

import { ProcessStatus } from "tensorlake";

const proc = await sandbox.startProcess("python", {
  args: [
    "-c",
    "import sys; print('hello'); print('oops', file=sys.stderr)",
  ],
});

let info = await sandbox.getProcess(proc.pid);
while (info.status === ProcessStatus.RUNNING) {
  await new Promise((resolve) => setTimeout(resolve, 100));
  info = await sandbox.getProcess(proc.pid);
}

console.log((await sandbox.getStdout(proc.pid)).lines);
console.log((await sandbox.getStderr(proc.pid)).lines);
console.log((await sandbox.getOutput(proc.pid)).lines);

# Get stdout
curl https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/stdout \
  -H "Authorization: Bearer $TL_API_KEY"

# Get stderr
curl https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/stderr \
  -H "Authorization: Bearer $TL_API_KEY"

# Get combined output
curl https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/output \
  -H "Authorization: Bearer $TL_API_KEY"

Combined output response:

{
  "pid": 297,
  "lines": ["hello", "oops"],
  "line_count": 2
}

Error Handling

CLI
Python
TypeScript
HTTP

# The CLI prints stderr and returns a non-zero exit code on failure
tl sbx exec <sandbox-id> python -c "import nonexistent_module"

sandbox = Sandbox.create()
result = sandbox.run("python", ["-c", "import nonexistent_module"])

if result.exit_code != 0:
    print(f"Command failed with exit code {result.exit_code}")
    print(f"stderr: {result.stderr}")

const result = await sandbox.run("python", {
  args: ["-c", "import nonexistent_module"],
});

if (result.exitCode !== 0) {
  console.log(`Command failed with exit code ${result.exitCode}`);
  console.log(`stderr: ${result.stderr}`);
}

# Check the exited process status
curl https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid> \
  -H "Authorization: Bearer $TL_API_KEY"

Process status response:

{
  "pid": 305,
  "status": "exited",
  "exit_code": 1,
  "signal": null,
  "stdin_writable": false,
  "command": "python",
  "args": ["-c", "import nonexistent_module"],
  "started_at": 1773950228855,
  "ended_at": 1773950228866
}

stderr response:

{
  "pid": 305,
  "lines": [
    "Traceback (most recent call last):",
    "  File \"<string>\", line 1, in <module>",
    "ModuleNotFoundError: No module named 'nonexistent_module'"
  ],
  "line_count": 3
}

Start a Background Process

Use start_process for work that should keep running after the call returns — servers, watchers, and other long-lived jobs.

Python
TypeScript
CLI
HTTP

from tensorlake.sandbox import Sandbox

sandbox = Sandbox.create()
# Start a background process
proc = sandbox.start_process("python", ["-m", "http.server", "8080"])
print(f"PID: {proc.pid}")

const proc = await sandbox.startProcess("python", {
  args: ["-m", "http.server", "8080"],
});

console.log(`PID: ${proc.pid}`);

# Run a command in the background using shell syntax
tl sbx exec <sandbox-id> bash -c "python -m http.server 8080 &"

curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "command": "python",
    "args": ["-m", "http.server", "8080"]
  }'

Response (201 Created):

{
  "pid": 42,
  "status": "running",
  "command": "python",
  "args": ["-m", "http.server", "8080"],
  "stdin_writable": false,
  "started_at": 1710000000000
}

Request options:

{
  "command": "python",
  "args": ["-m", "http.server", "8080"],
  "env": {"PORT": "8080"},
  "working_dir": "/workspace",
  "stdin_mode": "pipe",
  "stdout_mode": "capture",
  "stderr_mode": "capture"
}

List Processes

Python
TypeScript
HTTP

procs = sandbox.list_processes()
for p in procs:
    print(f"PID {p.pid}: {p.status}")

const processes = await sandbox.listProcesses();
for (const proc of processes) {
  console.log(`PID ${proc.pid}: ${proc.status}`);
}

curl https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY"

Response:

{
  "processes": [
    {
      "pid": 42,
      "status": "running",
      "command": "python",
      "args": ["-m", "http.server", "8080"],
      "stdin_writable": false,
      "started_at": 1710000000000
    }
  ]
}

Stream Output

Stream stdout/stderr in real time for long-running commands using Server-Sent Events:

CLI
Python
TypeScript
HTTP

tl sbx exec streams combined output to your terminal while the process runs.

# Start a long-running process
sandbox = Sandbox.create()
proc = sandbox.start_process("python", ["-c", """
import time
for i in range(5):
    print(f"Step {i+1}/5")
    time.sleep(1)
"""])

# Stream output as it arrives
for event in sandbox.follow_output(proc.pid):
    print(event.line, end="")

const proc = await sandbox.startProcess("python", {
  args: [
    "-c",
    "import time\nfor i in range(5):\n print(f'Step {i+1}/5')\n time.sleep(1)",
  ],
});

for await (const event of sandbox.followOutput(proc.pid)) {
  process.stdout.write(event.line);
}

# Follow stdout via SSE
curl -N https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/stdout/follow \
  -H "Authorization: Bearer $TL_API_KEY"

# Follow combined output via SSE
curl -N https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/output/follow \
  -H "Authorization: Bearer $TL_API_KEY"

SSE stream:

event: output
data: {"line":"Step 1/2","timestamp":1773950220162,"stream":"stdout"}

event: output
data: {"line":"Step 2/2","timestamp":1773950220162,"stream":"stdout"}

event: eof
data: {}

Send Signals

Send POSIX signals to running processes:

Python
TypeScript
HTTP

import signal

proc = sandbox.start_process("python", ["-m", "http.server", "8080"])

# Gracefully stop the process
sandbox.send_signal(proc.pid, signal.SIGTERM)

await sandbox.sendSignal(proc.pid, 15);

# Send SIGTERM (15)
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/signal \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"signal": 15}'

# Send SIGKILL (9)
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/signal \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"signal": 9}'

Kill a Process

Python
TypeScript
HTTP

import signal

sandbox.send_signal(proc.pid, signal.SIGKILL)

await sandbox.killProcess(proc.pid);

curl -X DELETE https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid> \
  -H "Authorization: Bearer $TL_API_KEY"

Write to Stdin

Send input to a running process started with stdin in pipe mode:

import { StdinMode } from "tensorlake";

const proc = await sandbox.startProcess("python", {
  args: ["-i"],
  stdinMode: StdinMode.PIPE,
});

await sandbox.writeStdin(
  proc.pid,
  new TextEncoder().encode("print('hello')\n"),
);

await sandbox.closeStdin(proc.pid);

# Start a process with stdin pipe
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command": "python", "args": ["-i"], "stdin_mode": "pipe"}'

# Write to stdin
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/stdin \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/octet-stream" \
  --data-binary "print('hello')\n"

# Close stdin
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/stdin/close \
  -H "Authorization: Bearer $TL_API_KEY"

Managed Processes

Use a managed process for long-running services that should restart after a crash or failed health check. Managed processes use the same process API as normal background commands, but opt into supervision when you provide a name, a restart policy, or a health check.

Managed process flags start a background process. In the CLI, these flags require --detach. For blocking one-shot commands that stream output and return an exit code, use plain tl sbx exec or sandbox.run(...).

CLI
Python
TypeScript
HTTP

# Start and supervise a web server. The CLI prints the PID.
tl sbx exec <sandbox-id> --detach \
  --name dev-server \
  --restart always \
  --health-http 8080 \
  python -m http.server 8080

# Inspect the managed process
tl sbx ps <sandbox-id> <pid> --json

# Manually restart it through the supervisor
tl sbx restart <sandbox-id> <pid>

# Stop the process and remove it from supervision
tl sbx kill <sandbox-id> <pid>

from tensorlake.sandbox import (
    ProcessHealthCheck,
    ProcessHealthCheckType,
    RestartPolicy,
    RestartPolicyConfig,
    Sandbox,
)


sandbox = Sandbox.create()
proc = sandbox.start_process(
    "python",
    args=["-m", "http.server", "8080"],
    user="root",
    name="dev-server",
    restart=RestartPolicyConfig(
        policy=RestartPolicy.ALWAYS,
        max_restarts=10,
        initial_backoff_ms=500,
        max_backoff_ms=30_000,
    ),
    health_check=ProcessHealthCheck(
        type=ProcessHealthCheckType.HTTP,
        port=8080,
        path="/",
        interval_ms=1_000,
        failure_threshold=3,
    ),
)

print(proc.pid)
print(proc.managed.status)
print(proc.managed.health_status)

current = sandbox.get_process(proc.pid)
restarted = sandbox.restart_process(proc.pid)

import { Sandbox } from "tensorlake";

const sandbox = await Sandbox.create();
const proc = await sandbox.startProcess("python", {
  args: ["-m", "http.server", "8080"],
  user: "root",
  name: "dev-server",
  restart: {
    policy: "always",
    maxRestarts: 10,
    initialBackoffMs: 500,
    maxBackoffMs: 30_000,
  },
  healthCheck: {
    type: "http",
    port: 8080,
    path: "/",
    intervalMs: 1_000,
    failureThreshold: 3,
  },
});

console.log(proc.pid);
console.log(proc.managed?.status);
console.log(proc.managed?.healthStatus);

const current = await sandbox.getProcess(proc.pid);
const restarted = await sandbox.restartProcess(proc.pid);

curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "command": "python",
    "args": ["-m", "http.server", "8080"],
    "user": "root",
    "name": "dev-server",
    "restart": {
      "policy": "always",
      "max_restarts": 10,
      "initial_backoff_ms": 500,
      "max_backoff_ms": 30000
    },
    "health_check": {
      "type": "http",
      "port": 8080,
      "path": "/",
      "interval_ms": 1000,
      "failure_threshold": 3
    }
  }'

Managed response fields:

{
  "pid": 42,
  "status": "running",
  "command": "python",
  "args": ["-m", "http.server", "8080"],
  "managed": {
    "id": "dev-server",
    "name": "dev-server",
    "status": "running",
    "restart_count": 0,
    "restart": {"policy": "always"},
    "health_status": "starting",
    "consecutive_health_failures": 0
  }
}

# Manual managed restart
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/processes/<pid>/restart \
  -H "Authorization: Bearer $TL_API_KEY"

Restart policies are never, on_failure, and always. Health checks can be HTTP checks against a local sandbox port and optional path, or TCP checks against a local port. Process user selection accepts a username such as root, a UID string, a uid:gid string, or an object such as {"uid": 1000, "gid": 1000}. The default user is tl-user.

Interactive Shell

CLI
Python
TypeScript
HTTP

# Open an interactive shell in the sandbox
tl sbx ssh <sandbox-id>

# Use a custom shell
tl sbx ssh <sandbox-id> --shell /bin/sh

pty = sandbox.create_pty(
    command="/bin/bash",
    rows=24,
    cols=80,
)

pty.send_input("pwd\nexit\n")
print(pty.wait())

const pty = await sandbox.createPty({
  command: "/bin/bash",
  rows: 24,
  cols: 80,
});

await pty.sendInput("pwd\nexit\n");
console.log(await pty.wait());

# 1. Create a PTY session
curl -X POST https://<sandbox-id>.sandbox.tensorlake.ai/api/v1/pty \
  -H "Authorization: Bearer $TL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command": "/bin/bash", "rows": 24, "cols": 80}'

Response:

{
  "session_id": "LYtJOrxE9Kz3bphPUDzuX",
  "token": "<pty-session-token>"
}

# 2. Connect via WebSocket
wscat -c "wss://<sandbox-id>.sandbox.tensorlake.ai/api/v1/pty/<session-id>/ws?token=<token>"

tl sbx ssh requires an interactive terminal and automatically resumes a suspended sandbox before opening the PTY session. For the full programmatic PTY flow, including the READY handshake, binary WebSocket opcodes, and clean shutdown, see PTY Sessions.

Learn More

PTY Sessions

Interactive shells over WebSocket.

File Operations

Read, write, and copy files.

Lifecycle

Sandbox states, resources, and timeouts.

Networking

Control internet access and outbound destinations.

Overview

Guides

Use Cases

Commands & Processes

Basic Execution

CLI Options

Shell Commands

Get Process Output

Error Handling

Start a Background Process

List Processes

Stream Output

Send Signals

Kill a Process

Write to Stdin

Managed Processes

Interactive Shell

Learn More

PTY Sessions

File Operations

Lifecycle

Networking

​Basic Execution

​CLI Options

​Shell Commands

​Get Process Output

​Error Handling

​Start a Background Process

​List Processes

​Stream Output

​Send Signals

​Kill a Process

​Write to Stdin

​Managed Processes

​Interactive Shell

​Learn More

PTY Sessions

File Operations

Lifecycle

Networking

Basic Execution

CLI Options

Shell Commands

Get Process Output

Error Handling

Start a Background Process

List Processes

Stream Output

Send Signals

Kill a Process

Write to Stdin

Managed Processes

Interactive Shell

Learn More