Delta Infra · cloud–device infrastructure
Agent Infra
Delta Infra is the infrastructure layer beneath a device-side agent. It joins compute, data, and tools into one cloud–device surface, so the agent only has to care about the task. Everything below stays transparent.
Core capabilities
Isolated execution
Each sandbox runs isolated from the host and from other tenants, with dangerous commands blocked at the boundary.
Elastic scheduling
Sandboxes are created on demand in seconds, scale out under load, and are reclaimed automatically once idle.
Multi-cloud
Own capacity plus third-party providers, pluggable and not tied to one vendor, with price comparison and failover.
Model routing
One interface over 20+ models, routing on semantics, price, and latency to cut inference cost.
Multi-agent
Runs and manages several agents at once, across frameworks, with full lifecycle handling.
Standard SDK
delta-cli, a REST API, and an MCP server — integration is a single command.
Three clouds
Compute
GPU sandboxes allocate themselves from a declared task; the router picks the lowest-latency, lowest-cost node for inference.
Data
Code travels to the data instead of the reverse. Agent code runs on the data side and only conclusions come back, inside a private space.
Tools
Bayesian optimisation, deep research, and protein zero-shot mutation analysis ship as maintained container images. Nothing to install locally.
Quick start
npx @delta-infra/cli@latest install
delta-cli sandbox create --image opensandbox/pytorch-cuda13:latest --gpu 1
delta-cli sandbox run <sandbox-id> --command "python train.py"
Delta Infra is deliberately unopinionated about what sits on top: Cursor, Codex, OpenCode, and
Claude Code can all install delta-cli and call cloud compute directly.