anycloud.sh provides a command-line interface that lets developers and AI agents run containerized jobs, services, and VMs across existing AWS, Azure, GCP, Lambda, and Vast cloud accounts. The platform includes agent-session-scoped spend controls—throttles and budgets—that keep autonomous coding and research agents from exceeding cost limits. Ångstrom AI used anycloud to run over 100,000 GPU hours for crystal structure prediction research while keeping agent-driven experiment loops within budget.
- Artificial Intelligence
- AI Agents
- Developer Tools
- Software Only
Funding
Funding not disclosed
Founders
Product
Problem
AI research teams and developers increasingly rely on autonomous agents to launch and manage GPU workloads, but giving agents unrestricted cloud access risks runaway spend from incorrect batches or unbounded job fan-out. Existing cloud tools lack session-level guardrails, forcing teams to choose between manual oversight and agent autonomy, which slows iteration and increases financial risk.
Solution
anycloud.sh provides a unified CLI for deploying containerized jobs, long-running services, and interactive VMs across AWS, Azure, GCP, Lambda, and Vast cloud accounts. The platform is designed for agent-driven workflows: Claude Code, Codex, Cursor, and Aider sessions are automatically detected and tagged, and each session can be bounded by spend throttles and budgets that pause new workload dispatch when caps are reached while allowing running jobs to finish. Researchers and developers use the same CLI for interactive and agent-driven operations, with structured JSON output for status, cost, and results to support programmatic decision loops. Cloud credentials stay with the user's existing accounts, and the scheduler handles capacity provisioning, spot instance usage, and workload lifecycle management.
Target Audience
Primary users are AI research teams, machine learning engineers, and developer organizations that run GPU-heavy workloads and want to delegate job execution and experiment management to autonomous agents without exceeding budget controls.
Features
- Agent session scoping that detects Claude Code, Codex, Cursor, and Aider sessions and applies per-session spend throttles and budgets independently
- Throttle controls that cap estimated live burn rate (dollars per hour) and pre-charge candidate VMs to prevent queue bursts from exceeding limits
- Budget controls that enforce daily, weekly, or monthly UTC-period spend caps using settled plus estimated costs, with automatic queue blocking and reset at period boundaries
- Job, Service, and VM primitives with support for GPU types including H100 and L40S, spot capacity, checkpoint buckets, and public HTTPS URLs for services
- Python SDK (sync and async) for submitting jobs, monitoring status, and managing deployments with explicit credential and cloud configuration controls
- CLI commands for SSH access, container execution, cost inspection, and workload termination across on-demand and spot instances