Choose what works for you.

Email us

Questions, feedback, or ideas.

hello@submilli.ai
Email us

Book a call

Talk through your use case with the builders.

Book a call

Safely run the code your agents write

Your permissions enforced on each call, with an audit trail.

Why Submilli

Why Submilli

Every program gets its own sandbox

Each program runs in its own WebAssembly sandbox, with no network or file system access by default, and no startup time.

Sandboxes share one server process.

See how a program runs

Only granted tools

  • stripe.refundGranted
  • NetworkBlocked
  • File systemBlocked

Default: nothing

Permissions down to each argument

A tool allowlist only says which tools. A Blueprint goes down to each argument, and the runtime enforces it outside the model.

Stored credentials stay hidden from the model. Your agent's code can use them.

Watch how Blueprints work

Every call checked

Your permissions, enforced

  1. Read details✓ Allowed
  2. Send message? Ask a human
  3. Delete record× Denied

Runs that pick up where they stopped

A run continues after a server crash or restart. Finished steps are not repeated, with no separate workflow engine or workflow code.

An interrupted step can run again, so external effects must be safe to repeat. A stuck run needs an explicit resume.

See how runs recover

Finished steps never re-run

Run progress2 kept · 1 resumed
Server restart
Read details
kept
Prepare message
kept
Send message
resumed

Built for agents that write code

The agent writes one program instead of calling tools one at a time. The model sees the result.

Submilli helps it find Packages and fix errors in that code.

Learn why code execution works

Run your agents on Submilli.