Questions about Submilli
Clear answers about runtime permissions, integrations, customer scope, and code execution.
What is Submilli?
Submilli runs agent-written TypeScript programs under explicit permissions. A Blueprint defines the available Packages, Capabilities, resources, and limits. The runtime checks gated operations against that policy before they proceed.
Is it a sandbox?
Yes. Submilli runs agent-generated TypeScript in WebAssembly isolation and lets you specify which operations it may perform, on which resources, and with which arguments. You define permissions in a Blueprint, and the runtime checks covered operations against them before they proceed. Trusted Packages supply the operation names and fields used by those checks, such as a refund's customer and amount.
Can I use my existing APIs and MCP tools?
Yes. Packages expose service operations as typed functions with explicit Capability checks. MCP tools configured in a Blueprint are also exposed as typed functions, with permission checks applied before calls are sent.
Can permissions use context supplied by my application?
Yes. Your application binds trusted values when it opens a session, such as a customer or tenant ID, an allowed resource, or a transaction limit. Permission rules compare an operation’s resource and arguments against those values before it proceeds. For example, a rule can restrict access to the current customer’s records or cap an amount at the limit supplied by your app. The generated code cannot change these bindings. Your application remains responsible for authenticating users and supplying the right values.
How does Submilli keep startup fast and server overhead low?
Submilli runs WebAssembly instances inside the host process. You can embed the runtime in your application or use the server through HTTP or MCP. Each run gets an isolated instance without starting a separate VM or container. Fast instance creation and a small per-run memory footprint let more concurrent runs share a server. Total latency and capacity still depend on compilation, workload, and configured limits.
Can code mode reduce cost, improve task success, or make agents faster?
Yes. Code mode lets an agent write a program that calls tools, processes intermediate responses, and returns a concise result without a model round trip for every step. Anthropic’s example reduces tool-definition context from 150,000 to 2,000 tokens, a 98.7% reduction. In CodeAct’s M³ToolEval benchmark, GPT-4-1106-preview achieved 74.4% task success with Python code actions versus 53.7% with text actions, a gain of 20.7 percentage points. Submilli provides the runtime to execute agent-written code with isolation, permissions, typed APIs, and actionable errors. These are results from other systems, not Submilli benchmarks. Results depend on the model, workload, and integration.
Anthropic: code execution with MCPCloudflare: Code ModeCodeAct benchmark resultsLearn why code execution works
What does the permission system not guarantee?
Permissions govern operations covered by runtime and trusted Package checks. Package authors must implement the relevant Capability checks. Permissions do not prove that a program’s result is correct, replace application authentication, or determine which personal fields a permitted service should return.
Is Submilli open source, and where do I start?
The runtime, CLI, and server are open source under Apache 2.0. Start with the quickstart to install the tools, register a Blueprint, and run a program. The public repository contains the implementation, examples, and documentation.