# Tensorlake > Tensorlake is the sandbox cloud for AI agents: stateful, AI-native > microVM sandboxes with sub-second cold starts, snapshot, > suspend/resume, and sandbox-native durable orchestration. Securely > run untrusted, LLM-generated code in isolated sandboxes that scale > from one to thousands. This file is a map for LLMs and coding > agents building applications on Tensorlake — it points at the pages > you need to install the SDK, authenticate, launch sandboxes, and > orchestrate durable workflows. Tensorlake's runtime is built on Firecracker microVMs with a custom scheduler for agent workloads: isolated tool/code execution, stateful checkpoints, per-second billing, and up to five million sandboxes per project. The two surfaces you build against are **Sandboxes** and **Applications** (sandbox-native durable orchestration). ## Product - [Home](https://tensorlake.ai/): Product overview, capabilities, benchmarks, and customer signal. - [Agent harnesses](https://tensorlake.ai/#product): Run full AI agents inside isolated, stateful sandboxes with filesystem, shell, and package state. - [Capabilities](https://tensorlake.ai/#capabilities): Sandboxes, agent harnesses, isolated tool execution, RL environments, orchestration. - [Benchmarks](https://tensorlake.ai/#bench): SQLite and filesystem benchmarks vs. E2B, Modal, Vercel, and Daytona. - [Customer signal](https://tensorlake.ai/#customers): Production users and testimonials. ## Pricing and getting started - [Pricing](https://tensorlake.ai/pricing): Plans and usage-based pricing. - [Sign up](https://cloud.tensorlake.ai): Free tier — start building. ## Start here - [Documentation home](https://docs.tensorlake.ai/introduction): Entry point. What Tensorlake is and which surface to pick for your use case. - [Full docs index (llms.txt)](https://docs.tensorlake.ai/llms.txt): Machine-readable index of every page in the Tensorlake documentation — link-per-page for SDK, API reference, guides, and examples. - [Dashboard / sign up](https://cloud.tensorlake.ai/login): Sign in to Tensorlake Cloud to get an API key. Free tier included, no credit card required. - [Authentication](https://docs.tensorlake.ai/platform/authentication): How to create API keys and make authenticated requests to the Tensorlake APIs. - [Agent Skills](https://docs.tensorlake.ai/agent-skills): Skill files that teach coding agents (Claude Code, Codex, OpenCode, Amp) how to ship production workflows on Tensorlake. Drop these into your agent's skill directory. ## Sandboxes — isolated, stateful execution ### Sandbox capability notes for coding agents - Tensorlake Sandboxes expose SDKs and a REST API through each sandbox's proxy URL. - REST process execution supports per-process environment variables via `env` on `POST /api/v1/processes` and `POST /api/v1/processes/run`. - REST streaming stdin is supported for processes started with `stdin_mode: "pipe"`: write bytes with `POST /api/v1/processes/{pid}/stdin`, then send EOF with `POST /api/v1/processes/{pid}/stdin/close`. - `POST /api/v1/processes/run` is the non-interactive one-request path: it streams stdout/stderr over Server-Sent Events and always starts with stdin closed. - Use PTY sessions for interactive terminal semantics over WebSocket. Use TCP tunnels for raw binary WebSocket forwarding to sandbox-local TCP ports. - The raw OpenAPI spec is the canonical machine-readable REST reference: https://docs.tensorlake.ai/api-reference/openapi.yaml - [Sandboxes overview](https://docs.tensorlake.ai/sandboxes/introduction): Isolated, stateful Firecracker-backed execution environments for agents. Sub-second cold starts, suspend/resume, snapshots. - [Sandboxes quickstart](https://docs.tensorlake.ai/sandboxes/quickstart): Install the SDK, authenticate, and run your first sandbox in under five minutes. - [Sandbox lifecycle](https://docs.tensorlake.ai/sandboxes/lifecycle): States, creation, suspend/resume, cleanup. - [Sandbox images](https://docs.tensorlake.ai/sandboxes/images): Define reusable named sandbox images in Python, TypeScript, or Dockerfiles. Valid base images include `tensorlake/ubuntu-minimal` and `tensorlake/ubuntu-systemd`. - [Snapshots](https://docs.tensorlake.ai/sandboxes/snapshots): Save and restore filesystem, memory, and running processes. - [Execute commands](https://docs.tensorlake.ai/sandboxes/commands): Run commands with output capture, streaming, and error handling. - [Process management](https://docs.tensorlake.ai/sandboxes/processes): Start, monitor, and manage background processes. - [PTY sessions](https://docs.tensorlake.ai/sandboxes/pty-sessions): Interactive shell sessions over WebSocket with live terminal I/O. - [File operations](https://docs.tensorlake.ai/sandboxes/file-operations): Copy, read, write, and manage files inside a sandbox. - [Environment variables](https://docs.tensorlake.ai/sandboxes/environment-variables): Per-command and per-PTY env vars in CLI, Python, and TypeScript. - [Networking](https://docs.tensorlake.ai/sandboxes/networking): Route internet traffic into sandbox applications and control outbound access. - [AI code execution](https://docs.tensorlake.ai/sandboxes/ai-code-execution): Run LLM-generated code in isolated containers with resource limits; integrate sandboxes as agent tools. - [Computer use](https://docs.tensorlake.ai/sandboxes/computer-use): Launch `tensorlake/ubuntu-vnc` sandboxes and drive their desktop from Python or JavaScript. - [Skills in sandboxes](https://docs.tensorlake.ai/sandboxes/skills-in-sandboxes): Pre-load Tensorlake skill files inside sandbox images so coding agents auto-discover them at startup. ## Applications — durable orchestration for agents - [Applications introduction](https://docs.tensorlake.ai/applications/introduction): Add serverless orchestration to any agent. - [Applications quickstart](https://docs.tensorlake.ai/applications/quickstart): Ship a first durable agent endpoint. - [Programming agents (overview)](https://docs.tensorlake.ai/applications/overview): Core concepts and common patterns for running agents on Tensorlake. - [SDK reference (functions, apps, decorators)](https://docs.tensorlake.ai/applications/concepts): Functions, applications, decorators, request context, and lifecycle. - [Architecture](https://docs.tensorlake.ai/applications/architecture): How the Application Runtime runs your code under the hood. - [Building workflows](https://docs.tensorlake.ai/applications/building-workflows): Multi-step data workflows with parallel execution. - [Durable execution](https://docs.tensorlake.ai/applications/durability): How crashes and restarts preserve progress. - [Crash recovery](https://docs.tensorlake.ai/applications/crash-recovery): How agents survive failures and resume without losing work. - [Futures](https://docs.tensorlake.ai/applications/futures): Run multiple function calls in parallel. - [Map-Reduce](https://docs.tensorlake.ai/applications/map-reduce): Fan-out / fan-in patterns. - [Parallel sub-agents](https://docs.tensorlake.ai/applications/parallel-sub-agents): Fan out work to specialist agents running in parallel. - [Async functions](https://docs.tensorlake.ai/applications/async-functions): Python `async`/`await` inside Tensorlake functions. - [Sandboxes from Applications](https://docs.tensorlake.ai/applications/sandboxes): Two patterns for running agents with isolated code execution from an Application. - [Retries & rate limits](https://docs.tensorlake.ai/applications/retries): Handle LLM rate limits, transient failures, and validation with durable retries. - [Timeouts](https://docs.tensorlake.ai/applications/timeouts): How function timeouts work and how progress updates reset them. - [Error handling](https://docs.tensorlake.ai/applications/error-handling). - [Secrets](https://docs.tensorlake.ai/applications/secrets): Provide secrets to Tensorlake functions. - [Container images](https://docs.tensorlake.ai/applications/images): Custom images for Applications. - [Cron scheduler](https://docs.tensorlake.ai/applications/cron-scheduler): Schedule recurring invocations of your Orchestration endpoints. - [Scale-out & queuing](https://docs.tensorlake.ai/applications/scale-out-queuing): Automatic scaling per function. - [Autoscaling](https://docs.tensorlake.ai/applications/scaling-agents): Autoscaling guide for orchestration endpoints. - [Observability](https://docs.tensorlake.ai/applications/observability): Built-in tracing and execution timelines. - [Logging](https://docs.tensorlake.ai/applications/guides/logging). - [Streaming progress updates](https://docs.tensorlake.ai/applications/guides/streaming-progress): Stream real-time progress from functions. - [Troubleshooting](https://docs.tensorlake.ai/applications/production/troubleshooting): Common issues and how to debug them. ## API reference (v2) - [API introduction](https://docs.tensorlake.ai/api-reference/v2/introduction): Conventions, authentication, and errors. - [Sandboxes — create](https://docs.tensorlake.ai/api-reference/v2/sandboxes/create), [get](https://docs.tensorlake.ai/api-reference/v2/sandboxes/get), [list](https://docs.tensorlake.ai/api-reference/v2/sandboxes/list), [update](https://docs.tensorlake.ai/api-reference/v2/sandboxes/update), [suspend](https://docs.tensorlake.ai/api-reference/v2/sandboxes/suspend), [resume](https://docs.tensorlake.ai/api-reference/v2/sandboxes/resume), [snapshot](https://docs.tensorlake.ai/api-reference/v2/sandboxes/snapshot), [restore](https://docs.tensorlake.ai/api-reference/v2/sandboxes/restore), [delete](https://docs.tensorlake.ai/api-reference/v2/sandboxes/delete). - [Processes — start](https://docs.tensorlake.ai/api-reference/v2/processes/start), [run](https://docs.tensorlake.ai/api-reference/v2/processes/run), [get](https://docs.tensorlake.ai/api-reference/v2/processes/get), [list](https://docs.tensorlake.ai/api-reference/v2/processes/list), [kill](https://docs.tensorlake.ai/api-reference/v2/processes/kill), [signal](https://docs.tensorlake.ai/api-reference/v2/processes/signal), [stdout](https://docs.tensorlake.ai/api-reference/v2/processes/stdout), [stderr](https://docs.tensorlake.ai/api-reference/v2/processes/stderr), [output](https://docs.tensorlake.ai/api-reference/v2/processes/output), [follow-output (SSE)](https://docs.tensorlake.ai/api-reference/v2/processes/follow-output), [stdin](https://docs.tensorlake.ai/api-reference/v2/processes/stdin), [close-stdin](https://docs.tensorlake.ai/api-reference/v2/processes/close-stdin). - [PTY — create](https://docs.tensorlake.ai/api-reference/v2/pty/create), [get](https://docs.tensorlake.ai/api-reference/v2/pty/get), [list](https://docs.tensorlake.ai/api-reference/v2/pty/list), [resize](https://docs.tensorlake.ai/api-reference/v2/pty/resize), [kill](https://docs.tensorlake.ai/api-reference/v2/pty/kill), [WebSocket attach](https://docs.tensorlake.ai/api-reference/v2/pty/websocket). - [Sandbox files — read](https://docs.tensorlake.ai/api-reference/v2/sandbox-files/read), [write](https://docs.tensorlake.ai/api-reference/v2/sandbox-files/write), [list](https://docs.tensorlake.ai/api-reference/v2/sandbox-files/list), [delete](https://docs.tensorlake.ai/api-reference/v2/sandbox-files/delete). - [Runtime metadata — health](https://docs.tensorlake.ai/api-reference/v2/runtime/health), [info](https://docs.tensorlake.ai/api-reference/v2/runtime/info). - [TCP tunnel WebSocket](https://docs.tensorlake.ai/api-reference/v2/tunnels/tcp): Authenticated binary WebSocket tunnel to sandbox-local TCP ports. - [SSH lifecycle — status](https://docs.tensorlake.ai/api-reference/v2/ssh/status), [enable](https://docs.tensorlake.ai/api-reference/v2/ssh/enable), [disable](https://docs.tensorlake.ai/api-reference/v2/ssh/disable): In-sandbox SSH daemon used by Tensorlake's SSH proxy. - [OpenAPI spec](https://docs.tensorlake.ai/api-reference/openapi.yaml): Machine-readable OpenAPI definition. ## Platform - [Access control](https://docs.tensorlake.ai/platform/access-control): Organizations, projects, roles, users. - [Billing](https://docs.tensorlake.ai/platform/billing): Per-second sandbox billing, plan limits, usage. - [EU data residency](https://docs.tensorlake.ai/platform/eu-data-residency): Regional endpoints and residency controls. - [Security policies](https://docs.tensorlake.ai/platform/security): Kernel boundaries, tenancy, compliance posture. - [Single Sign-On (SSO)](https://docs.tensorlake.ai/platform/sso): OIDC and SAML 2.0. - [Webhooks overview](https://docs.tensorlake.ai/platform/webhooks/overview), [configuration](https://docs.tensorlake.ai/platform/webhooks/configuration), [signature verification](https://docs.tensorlake.ai/platform/webhooks/signature-verification), [testing](https://docs.tensorlake.ai/platform/webhooks/testing). - [Webhook payload — workflows](https://docs.tensorlake.ai/platform/webhooks/payloads/workflows). ## Examples & cookbooks - [Examples overview](https://docs.tensorlake.ai/examples/overview). - [Agent with tool calling](https://docs.tensorlake.ai/examples/agentic-applications/agent-with-tools): Claude agent orchestrating workflows via tool calls. - [Code interpreter agent](https://docs.tensorlake.ai/examples/agentic-applications/code-interpreter): Secure code-execution agent on Tensorlake + OpenAI. - [Deep research agent](https://docs.tensorlake.ai/examples/agentic-applications/deep-research): Multi-agent deep research pipeline. - [Personal finance manager](https://docs.tensorlake.ai/examples/agentic-applications/personal-finance-manager). - [Weather agent](https://docs.tensorlake.ai/examples/agentic-applications/weather-agent). - [Web scraper → MongoDB Atlas](https://docs.tensorlake.ai/examples/agentic-applications/web-scraper). - [Workflow tutorial — Hello World](https://docs.tensorlake.ai/examples/tutorials/workflow-tutorial). ## FAQs - [Workflows FAQ](https://docs.tensorlake.ai/faqs/workflows-faq). - [Sandbox and platform FAQ](https://tensorlake.ai/faq): Plain-language answers on sandboxes, security, pricing, and how Tensorlake compares to E2B, Modal, and Daytona. ## Open source - [Indexify](https://docs.tensorlake.ai/opensource/indexify): Open-source compute engine for agentic data applications. - [Indexify — deployment](https://docs.tensorlake.ai/opensource/deployment), [configuration](https://docs.tensorlake.ai/opensource/configuration), [monitoring](https://docs.tensorlake.ai/opensource/monitoring). - [GitHub — tensorlakeai](https://github.com/tensorlakeai): SDK source, benchmarks, and open-source projects. - [Slack community](https://join.slack.com/t/tensorlakecloud/shared_invite/zt-32fq4nmib-gO0OM5RIar3zLOBm~ZGqKg): Ask questions and share what you're building. ## Blog - [Build Your Own CI infrastructure](https://tensorlake.ai/blog/build-your-own-ci-infrastructure): You get a GitHub runner, and you get a GitHub runner... - [Introducing Tensorlake BYOC: Serverless Sandboxes in Your Private Cloud](https://tensorlake.ai/blog/introducing-tensorlake-byoc): Run Tensorlake Sandboxes on compute you own across AWS, GCP, Azure, CoreWeave, Nebius, and bare metal. Same APIs, images, and SDKs as Tensorlake Cloud, with compute, storage, and workload traffic inside your own network. - [Run Devin Outposts on Tensorlake Sandboxes.](https://tensorlake.ai/blog/devin-outposts-on-tensorlake): Cognition keeps the agent loop; execution runs inside a Tensorlake sandbox you control. We built the integration, and you can clone and run it today. - [Firecracker disk snapshots in O(changed bytes), not O(disk size)](https://tensorlake.ai/blog/firecracker-disk-snapshots-o-changed-bytes): We added copy-on-write image layering inside Firecracker's block device, and the disk got faster, not slower. Snapshots pause the VM for tens of milliseconds and cost only what changed, so a running Postgres barely notices them, at any disk size. - [Zero-copy TLS ingress with kTLS and splice(2) for sandboxes](https://tensorlake.ai/blog/near-zero-overhead-sandbox-networking) - [Accelerating Harbor with Tensorlake.](https://tensorlake.ai/blog/accelerating-harbor-with-tensorlake): Tensorlake runs the real task Docker images for Harbor on microVM sandboxes. On Terminal-Bench 2.1 the oracle passes all 89 tasks in 36m 44s — every image published, cold starts in seconds. - [Tensorlake Sandboxes as the runtime for Claude Managed Agents.](https://tensorlake.ai/blog/claude-managed-agents): Run Claude Managed Agents on Tensorlake Sandboxes: you keep the control of self-hosting, and drop the always-on box. The short version — the server goes away. - [Claude Opus 4.7 vs. Kimi K2.6: Real-World Coding Test](https://tensorlake.ai/blog/claude-opus-4-7-vs-kimi-k2-6-real-world-coding-test): A practical comparison between the premium closed frontier model and the cheaper open-weight agentic coder, on the same two-part coding task. - [Filter and rank: how we schedule sandboxes across every cloud](https://tensorlake.ai/blog/multi-cloud-scheduling): One primitive — filter, then rank — runs through statefulness, multi-cloud, BYOC, and the data path. A tour of the scheduler behind Tensorlake's sandbox fleet. - [Browser Harness: direct Chrome DevTools Protocol access and a self-healing harness for browser agents](https://tensorlake.ai/blog/browser-harness-cdp-self-healing): Browser-use strips their harness to 592 lines and gives the LLM a raw WebSocket to Chrome's DevTools Protocol. When the agent hits a gap, it writes a helper function — and saves it for future runs. - [How scaffold design affects coding agent benchmark scores: lessons from Droid](https://tensorlake.ai/blog/droid-scaffold-design-benchmark): Contextual instruction injection, per-model tool schemas, and planning/execution splits — the harness decisions behind Droid's 77.3% on Terminal-Bench. - [Hermes: the coding agent that gets better the more you use it](https://tensorlake.ai/blog/hermes-coding-agent-learns-from-use): Most coding agents are stateless. Hermes uses a closed learning loop where the agent creates skills from experience, improves them during use, and builds a persistent model of who you are across sessions. - [Pi: a coding agent with efficient system prompting](https://tensorlake.ai/blog/pi-coding-agent-efficient-system-prompting): Pi keeps its entire system prompt — including all tool definitions — under 1,000 tokens, a 10× reduction over tools like Claude Code or Cline. - [ForgeCode: top open source coding agent in Terminal-Bench@2.0](https://tensorlake.ai/blog/forgecode-terminal-bench): ForgeCode reaches 81.8% on Terminal-Bench 2.0 with both Claude Opus 4.6 and GPT-5.4 — a look at what the harness is doing that the model isn't. - [Building Sandboxes for Computer Use](https://tensorlake.ai/blog/building-sandboxes-for-computer-use): Computer Use is a tiny loop. The hard part is building the boring, reproducible desktop around it. - [Starting hundreds of sandboxes in parallel, and the design that makes it possible.](https://tensorlake.ai/blog/starting-1000-sandboxes-in-parallel): We replaced reconciliation loops with durable command outbox to scale our sandbox scheduler to start 1000s of sandboxes every second - [Tensorlake is now an official Harbor environment runtime](https://tensorlake.ai/blog/tensorlake-harbor-environment-runtime): Harbor defines and evaluates terminal tasks. Tensorlake provides the MicroVM execution layer. Together they are a full evaluation stack for CLI agents. - [Suspend vs. snapshot: pause a sandbox, or save it for reuse?](https://tensorlake.ai/blog/suspend-vs-snapshot): One is a pause button, the other is a save file. Same state, different question — and the answer shapes your cost model, your fan-out pattern, and which failures you can recover from. - [How we got to 5,000,000 sandboxes per project](https://tensorlake.ai/blog/5-million-sandboxes-one-api): A walk through the scheduler rewrite, the snapshot format, and the two syscalls that freed us from the noisy-neighbor tax. - [Sub-second cold starts for stateful microVMs](https://tensorlake.ai/blog/sub-second-cold-starts): Snapshot-restore beats boot every time. How we cut P50 wake-up to 80ms. - [Autoresearch on steroids with sandboxes](https://tensorlake.ai/blog/autoresearch-on-steroids-with-sandboxes): An LLM agent can propose incremental training-script improvements, but safely executing untrusted code requires isolated sandboxes with resource limits — that's where Tensorlake comes in. Here's the overnight hill-climb, end to end. - [RL environments are generally available](https://tensorlake.ai/blog/rl-environments-ga): Ship your reward model alongside the same sandbox that runs inference — no separate training harness. - [Durable agent loops without a queue](https://tensorlake.ai/blog/durable-agent-loops): A checkpoint per step, at 2.4ms. Why we stopped writing to Redis and started writing to disk. - [Harbor x TensorLake: Infrastructure for Agentic Evals](https://tensorlake.ai/blog/harbor-x-tensorlake): TensorLake is now integrated as a first-class environment provider in Harbor, enabling scalable agent evaluation with secure ephemeral MicroVMs. - [Filesystem benchmarks: sandboxed vs. bare metal](https://tensorlake.ai/blog/sandbox-fs-benchmarks): Seven workloads, four runtimes. Most sandboxes leave 40% of your disk throughput on the table. - [BYOC: run Tensorlake in your own AWS account](https://tensorlake.ai/blog/byoc-private-link): Single-tenant control plane, PrivateLink, and data residency — without giving up the managed experience. - [We raised our Series A](https://tensorlake.ai/blog/series-a-announcement): 32M to build the sandbox layer every agent needs. A note on why, whats next, and who were hiring. - [Five patterns for tool execution in production agents](https://tensorlake.ai/blog/tool-execution-patterns): Ephemeral vs. persistent, per-tenant vs. shared, and when to pay the cold-start tax on purpose. - [Paged KV caches across sandbox boundaries](https://tensorlake.ai/blog/paged-kv-isolation): A small experiment in sharing attention caches between sibling sandboxes without breaking isolation. - [Claude Opus 4.6 vs GPT 5.3 Codex](https://tensorlake.ai/blog/claude-opus-4-6-vs-gpt-5-3-codex): GPT 5.3 Codex excels at fast execution and agentic coding, while Claude Opus 4.6 stands out in reasoning-heavy and long-context tasks. - [SOC 2 Type II, and a primer on our security model](https://tensorlake.ai/blog/soc2-and-security-primer): What we attest to, what we dont, and the exact kernel boundaries youre trusting. - [OpenCode: The Best Claude Code Alternative](https://tensorlake.ai/blog/opencode-the-best-claude-code-alternative): Most developers now use AI agents for coding tasks, but Claude Code's managed approach and constraints around model selection, pricing, and customization have spurred interest in open-source alternatives like OpenCode, which offers greater flexibility and control. - [The End of Database-Backed Workflow Engines: Building GraphRAG on Object Storage](https://tensorlake.ai/blog/end-of-database-backed-workflow-engines-graphrag-object-storage): GraphRAG ingestion needs fan-out parallelism and durable retries that database-backed workflow engines can't give it; object storage can. - [Python 3.12 is now the default image](https://tensorlake.ai/blog/python-312-default-image): Plus a new base image with uv pre-cached for 6x faster dependency installs. - [Building HackerNews Podcast Generator with Gemini 3, Elevenlabs](https://tensorlake.ai/blog/building-hackernews-podcast-generator): This article shows how to build a simple podcast generator that turns Hacker News posts into short audio summaries using a single Tensorlake Application. - [OpenAI GPT-5.2-Codex (high) vs. Claude Opus 4.5 vs. Gemini 3 Pro: Real-World Coding comparison](https://tensorlake.ai/blog/gpt5-2-codex-high-vs-opus-4-5-vs-gemini-3-pro): Three leading AI models were tested on real coding tasks in an 8K+ star repository to compare code quality, reliability, and practical performance. - [Claude Cowork: Architecture, Capabilities, and Usage Overview](https://tensorlake.ai/blog/claude-cowork-architecture-overview): Claude Cowork is an agentic task execution mode allowing Claude to plan and complete multi-step work on your behalf. - [Building an Outage Agent That Handles Repeated Production Incidents](https://tensorlake.ai/blog/building-outage-agent): An outage agent built on Tensorlake Applications triages alerts, investigates with Groq and Exa, and only escalates when a human is actually needed. - [The Next Enterprise Platform Isn't Data-Driven, It's Context-Driven](https://tensorlake.ai/blog/context-driven-enterprise-platform): Context graphs capture the reasoning behind decisions, not just their outputs, making AI-driven workflows auditable without replacing existing systems of record. - [Gemini 3 vs GPT-5.2: Detailed Coding Comparison](https://tensorlake.ai/blog/gemini3-vs-gpt5-coding): Gemini 3 Pro is fast and multimodal but produces simpler output; GPT-5.2 trades speed for more reliable, production-ready code. - [I Tried Letting Antigravity Build An Agent For Me. Here's What Actually Happened](https://tensorlake.ai/blog/antigravity-builds-agent): We let Antigravity build a guest checkout flow and abandoned-cart recovery emails end to end, and tracked where it held up and where it didn't. - [OpenAI GPT-5.2 Codex vs. Gemini 3 Pro vs Opus 4.5: Coding comparison](https://tensorlake.ai/blog/gpt5-2-gemini3-opus4-5-coding): GPT-5.2 Codex, Gemini 3 Pro, and Claude Opus 4.5 build a Minecraft clone, a Figma dashboard, and solve a hard LeetCode problem, with very different results. - [TOON vs JSON: A Token-Optimized Data Format for Reducing LLM Costs](https://tensorlake.ai/blog/toon-vs-json): TOON strips the repetitive syntax out of JSON for LLM prompts, cutting token usage while improving structured extraction accuracy. ## Community builds Developer builds and benchmarks on Tensorlake, written by the community and published on Towards AI / Medium. Curated index with editorial notes: https://tensorlake.ai/blog/community - [I Forked One AI Agent Into 12 Workers — and Every Clone Woke Up Already Knowing Everything](https://pub.towardsai.net/i-forked-one-ai-agent-into-12-workers-and-every-clone-woke-up-already-knowing-everything-4ea807811cc8): Cold-starting a worker costs 11 to 14 seconds of setup tax. Snapshot the parent once and each of 12 forks lands in about 2.4 seconds, with the isolation checked by deleting files inside one of them. - [Getting Started with Tensorlake Sandboxes: Build, Run, and Manage Your First Isolated AI Workload with Python](https://pub.towardsai.net/getting-started-with-tensorlake-sandboxes-build-run-and-manage-your-first-isolated-ai-workload-41715d797305): Seven steps from an empty sandbox to one that keeps its files, its packages, and its state across a snapshot and a suspend/resume cycle. - [Git Without the Clone: Durable, Versioned Workspaces for AI Agents](https://pub.towardsai.net/git-without-the-clone-durable-versioned-workspaces-for-ai-agents-b280241fe5ca): Mount a repo instead of cloning it: FUSE-backed workspaces with 30-second autosave, private snapshots, and server-side merges — unpublished work survives a sandbox crash. - [Crabbox Adds Tensorlake as a Delegated-Run Provider](https://crabbox.sh/providers/tensorlake.html): OpenClaw's Crabbox — warm a box, sync the diff, run the suite — drives Firecracker microVMs through the tensorlake CLI, so one flag puts your test run in a sandbox. - [OpenCode Is Powerful. That's Exactly the Problem.](https://pub.towardsai.net/opencode-is-powerful-thats-exactly-the-problem-06eebf266279): Routing OpenCode's bash, write, and edit tools into a disposable microVM with one plugin — the agent keeps its autonomy, your laptop stops being the blast radius. - [Building Stateful AI Agents That Survive Session Kills](https://pub.towardsai.net/building-stateful-ai-agents-that-survive-session-kills-e1877e3c78f0): Agent amnesia, solved: 84ms cold starts, snapshots as memory that carries prior experience, and forked sandboxes to evaluate candidate solutions in parallel. - [Your Agent Benchmark Is Lying to You](https://medium.com/data-science-collective/your-agent-benchmark-is-lying-to-you-65a552fea151): An 87% reported pass rate fell to 33% once verified — 53% of passes were fraudulent. Fork-per-task microVM isolation plus hidden verifiers make evals trustworthy. - [Why Your AI Experiments Keep Starting From Scratch (And How Tensorlake Fixes It)](https://pub.towardsai.net/why-your-ai-experiments-keep-starting-from-scratch-and-how-tensorlake-fixes-it-2e6091edf143): Warm up the environment once, memory-snapshot it, then fork N workers that resume mid-execution — instead of paying the same 40s setup tax on every run. - [I Ran 5 AI Agents in Parallel on Tensorlake. The Isolation Held. Here Is How I Built It.](https://pub.towardsai.net/i-ran-5-ai-agents-in-parallel-on-tensorlake-the-isolation-held-here-is-how-i-built-it-769384e046cf): Five agents in parallel, one deliberately crashed mid-run — the others finished clean with no defensive code, just per-agent microVMs. Registered images cut boot from 19.49s to 4.51s. - [Building Long-Running Claude-Managed Agents: Why State Matters More Than Compute](https://medium.com/towards-artificial-intelligence/building-long-running-claude-managed-agents-why-state-matters-more-than-compute-b5a35c03d680): An orchestrator/worker architecture on sandboxes that suspend when idle — checkpoint state and stop paying for compute instead of keeping boxes alive. - [Demo: A Sandbox Agent Built on Tensorlake](https://x.com/skjtwts/status/2067315371112386635): An AI coding assistant that writes, runs, and exports code inside a Tensorlake sandbox — demo on X, code on GitHub. - [Snapshot Once, Rollout a Thousand Times: A Practical RL Setup for Coding Agents](https://medium.com/data-science-collective/snapshot-once-rollout-a-thousand-times-a-practical-rl-setup-for-coding-agents-0f880a450610): Snapshot the environment once and fork every rollout from it: 8,000 RL rollouts drop from ~20 hours of naive environment rebuilds to ~4.4 hours. - [Tensorlake Is Now a Sandbox Provider in Deputies](https://x.com/sidpalas/status/2066897078022742428): Deputies, an open-source control plane for background agents, added Tensorlake as a sandbox provider — after it met every criterion from the project's own "what makes a good sandbox" post. - [Why Most Multi-Agent AI Systems Waste 90% of Their Time (And How to Fix It)](https://pub.towardsai.net/why-most-multi-agent-ai-systems-waste-90-of-their-time-and-how-to-fix-it-c0ce81f0e323): The bottleneck isn't concurrency — it's setup overhead. At 50 agents that's 4,500s spent preparing to work vs 50s with memory snapshots. - [Building a Stateful Code Interpreter with Tensorlake Sandboxes](https://medium.com/towards-artificial-intelligence/building-a-stateful-code-interpreter-with-tensorlake-sandboxes-df2f6d623a47): A build-along tutorial: from a toy code interpreter to one that remembers across sessions, suspends, snapshots, and forks — all on the free tier. - [Your AI Agent Is One rm -rf Away From Disaster. Here Is What I Found After 5 Sandbox Experiments.](https://pub.towardsai.net/your-ai-agent-is-one-rm-rf-away-from-disaster-here-is-what-i-found-after-5-sandbox-experiments-2d9d03ccca02): Five isolation experiments running hostile, LLM-generated code. All four isolation boundaries held, and boots averaged 1.121s with only 148ms of variance. - [8 AI Agent Sandboxes, Ranked — and Why the Fastest One Will Wreck Your Agent in Production](https://pub.towardsai.net/8-ai-agent-sandboxes-ranked-and-why-the-fastest-one-will-wreck-your-agent-in-production-c559f3967bc4): Eight sandboxes benchmarked head-to-head. Tensorlake places fourth on raw cold-start speed but first on consistency: a 0.09s median-to-P99 spread vs 0.37s+ elsewhere. - [I Built a Stateful Research Agent Inside a Sandbox. Here's What the Numbers Actually Looked Like.](https://pub.towardsai.net/i-built-a-stateful-research-agent-inside-a-sandbox-heres-what-the-numbers-actually-looked-like-2004aa378683): A research agent that keeps its state in a Tensorlake sandbox: it survives restarts and resumes mid-task, with real latency numbers throughout. ## Company - [Careers](https://tensorlake.ai/careers): Open roles. - [GitHub](https://github.com/tensorlakeai): Source code, examples, benchmarks. ## Optional - [Terms](https://tensorlake.ai/terms) - [Privacy](https://tensorlake.ai/privacy) - [DPA](https://tensorlake.ai/dpa) ## TODO Gaps that would help LLMs and coding agents build against Tensorlake faster. Update this file as docs.tensorlake.ai fills them in. - TODO: standalone **Python SDK reference** (classes, methods, typed signatures) for the `tensorlake` Python package — `applications/concepts` documents Applications-specific APIs, but there is no per-class reference for the Sandboxes SDK. - TODO: standalone **TypeScript SDK reference** for `@tensorlake/sdk` — mirror of the Python reference with matching methods. - TODO: standalone **CLI reference** (`tensorlake --help`) — current docs reference the CLI inline but not as a single command index. - TODO: **Error / exception reference** — a single page listing Tensorlake error types, HTTP status codes, and recommended retry behavior (the retry guide covers strategy, not error taxonomy). - TODO: **Rate limits & quotas** reference — per-plan concurrency, API-call, and storage limits in one place. - TODO: **Migration guides** from E2B, Modal, and Daytona — API-to-API mapping for teams moving off comparable sandboxes. - TODO: **Changelog / release notes** page — today updates land in the Engineering Log; a machine-readable release feed would help agents target specific SDK versions. - TODO: **Status endpoint** referenced here directly — https://status.tensorlake.ai/ exists as a dashboard; publishing a JSON feed would let agents check uptime programmatically.