Computer Zero

Hardware-backed evidence lets you verify where your agents run—not simply trust the server operator. Supported NVIDIA GPUs and Intel CPUs provide attestation for confidential computing environments.

• Verify the environment. Intel TDX and NVIDIA GPU attestation produce signed evidence about the protected environment, including relevant software measurements and security configuration. A verifier checks that evidence against approved reference values and policy.
• Authorize access after verification. A fresh challenge helps prevent replaying an old attestation. Workload keys can then be released through a secure channel only when the required checks pass.
• Protect data during processing. CPU and GPU isolation work together to reduce exposure to the host operating system and infrastructure administrators, with an encrypted channel between the protected environments.

Attestation establishes specific hardware and software properties. It does not prove that an application is bug-free or that its outputs are correct.

Your personal keys remain under your control. Modern elliptic-curve cryptography supports secure key agreement, while authenticated symmetric encryption protects file contents and detects tampering.

• Separate storage from access. Holding encrypted files does not, by itself, give a storage provider the ability to read them. Decryption requires the appropriate keys and authorization.
• Use your device’s protection. On supported devices, biometric verification can gate access to locally protected credentials or key-unlocking operations. Fingerprint or face verification happens on the device; biometric data is not sent to the website through WebAuthn.
• Delegate without handing over your identity. An authorized worker needs access to the data it processes, but that does not require sharing your personal master key. Scoped workload keys keep delegation separate from ownership.

Biometrics authorize access; they are not the encryption key. Key backup and recovery remain important, and external services can read information you explicitly send to them.

Open models give you more control over how intelligence is deployed. Where weights and licenses permit, a model can run on independent infrastructure rather than being available only through its creator’s API.

• Choose models around the work. Reasoning, coding, document analysis, and multimodal tasks have different requirements. Compare capability, latency, cost, and privacy instead of relying on one model for everything.
• Keep versions reproducible. A retained model checkpoint can be tested and deployed consistently, reducing dependence on a provider’s changing API catalog. Open availability enables continuity; it does not automatically guarantee quality or uptime.
• Reduce provider lock-in. Published weights and usable licenses allow more hosting options, independent evaluation, and adaptation. The exact license determines what you can redistribute, modify, or use commercially.

Model selection

• Moonshot AI · Kimi K3
• DeepSeek · V4.1 Flash
• Z.ai · GLM 5.3 Flash
• Qwen · Qwen3.8 Omni Flash
• Google · Gemma 4 31B Turbo
• Mistral AI · Mistral Large 4
• NVIDIA · Nemotron 3.5 Bolt

Open-source models, open-weight releases, and hosted API variants are not interchangeable. Availability and licensing must be checked for each exact release; the selection above is not a benchmark ranking.

Automation graphs connect AI reasoning, executable code, and external tools in one workflow. Each step has a role: interpret information, make a decision, perform a calculation, or carry out an approved action.

• Combine flexibility with precision. Use AI for research, extraction, planning, and drafting. Use code for calculations, validation, data transformations, and repeatable operations.
• Run on a schedule or an event. Server-side workers can operate around the clock without keeping your browser open. Connect steps through conditions, dependencies, and approval points instead of manually repeating the process.
• Keep execution observable. Follow progress, inspect results, and handle failures with controlled retries and limits. Autonomous execution stays within the permissions you grant.

Example: Every morning, collect project updates, calculate changes with code, draft a summary with AI, and request approval before sending it to your team.

Your agent swarm builds useful context over time, together with you. Preferences, ongoing projects, decisions, and feedback can persist beyond individual conversations, so each task does not begin from a blank slate.

• Remember what matters. Preserve relevant facts and working preferences rather than treating every interaction as equally important. Retrieve the context needed for the current task instead of loading an entire history.
• Consolidate with your guidance. Combine related information, resolve conflicting details, and update outdated assumptions. Your corrections help distinguish confirmed preferences from uncertain inferences.
• Personalize without losing control. Inspect, correct, or remove saved information and decide which agents may use it. Shared memory does not mean unrestricted access for every agent.

Example: Your agents remember a project’s terminology, your preferred reporting format, and earlier decisions—reducing repeated explanations and making follow-up work more relevant.

Memory helps agents understand your work; skills help them carry it out more effectively. A skill packages instructions, tools, and execution steps into a reusable method for a recurring task.

• Recognize recurring needs. With your permission, agents use task history, feedback, and stated goals to identify repeated work and propose relevant workflows.
• Improve through evidence. Successful outcomes become reusable patterns. Corrections, failures, and tests help refine instructions, tool choices, and validation steps rather than repeating the same mistakes.
• Keep changes reviewable. Evaluate proposed improvements before adopting them. A more capable skill does not automatically receive broader permissions. Workflow refinement is separate from retraining the underlying AI model.

Example: After several reporting tasks, your agent proposes a reusable workflow with your preferred sources, calculations, and format, then improves it using your feedback.

Share useful workflows, import community-built skills, and extend your workspace without starting from scratch. Apps bring reusable interfaces, tools, and automation into your own permission-controlled environment.

• Build on community work. Inspect an app’s purpose, source, dependencies, and requested access before importing it. Adapt reusable workflows to your own projects rather than recreating common solutions.
• Review security before granting access. Apps and skills can be submitted for in-depth security review through the RedTeam subnet. Review records should identify the tested version, scope, findings, and remediation status; a badge alone is not enough.
• Publish on your terms. Keep a workflow private or publish it for others to use, with usage-based earnings where enabled. Shared workflows must exclude your private files, credentials, and personal memory.

Software security review is different from hardware attestation. A reviewed release is not a guarantee of zero vulnerabilities, and material updates need renewed review.

Connect agents to third-party services without giving them unrestricted access to the outside world. API connectivity brings your existing tools into workflows, subject to each service’s authentication, interface, and limits.

• Choose how approval works. Allow ongoing access within a defined scope, require approval for each use, or disable a connection entirely. Your agents do not decide their own access policy.
• Control what can leave or change. Define which accounts, data, and actions are available. Separate reading information from actions such as sending messages, updating records, or making purchases.
• Keep a traceable record. Review which agent contacted which service, when it happened, the action requested, its authorization, and the outcome. Logs should protect credentials and avoid unnecessary copies of sensitive payloads.

Revoking access stops future authorized calls; it cannot retract data already sent or undo completed external actions. Some services require a dedicated connector or additional setup.