Core Components
@enclave-vm/ast
Static analysis and validation layer:- Pre-scanner - Checks size, nesting and literals before parsing
- AST validation - Blocks dangerous constructs
- Code transformation - Wraps code for safe execution
@enclave-vm/core
Runtime execution layer:- Sandboxed context - A realm that holds only the allowed globals, behind a membrane
- Resource limits - Timeout, iteration and tool-call limits, and memory accounting. In process the timeout and the memory limit are the enclave’s own accounting; the worker pool adds the engine’s heap limit and a hard stop
- Tool routing - Controlled external interactions
Tools
Your application’s capabilities exposed to scripts:- Your handlers - Validate inputs and decide what a script may do; this is your code
- Call limits -
maxToolCallsand the operation rate limit of the double VM bound how often tools are called - Audit logging - Every tool call passes through your
toolHandler, where you can log it
Defense-in-Depth Layers
Enclave uses a 6-layer security model:
The layers are built to hold independently of each other: the runtime does not rely on the validator having refused a script. What the sandbox does not deliver is listed in the Security Model.
Deployment Architectures
Single Process (Embedded)
Simplest deployment - everything runs in one process:Worker Pool (Worker Threads)
Code runs in the same sandbox, inside worker threads. A worker adds a heap limit that the engine enforces and a stop that does not depend on the script. A worker thread shares the process with your application: it is not process isolation.Client and Broker (EnclaveJS)
For applications with a browser client and streamed events:- Client -
@enclave-vm/clientand the React hooks of@enclave-vm/react: send a script, follow the events - Broker -
@enclave-vm/broker: HTTP routes, tool registry, session management, and the sandbox, in one process
identify hook so that a session belongs to the caller that created it.
The hook separates sessions, not data. Every caller that may create a session can run a script that calls any registered tool, and a tool handler receives the arguments and a context (sessionId, callId, secrets, signal), not the caller. A tool that reads data restricted per caller has to apply that restriction itself, as users:list does above with the owner of its session, or the data store has to enforce it. A session created in process has no owner unless it is created with owner.
A third tier, in which the broker hands execution to a separate runtime worker, is not implemented: the broker has no option for it, and @enclave-vm/runtime does not execute scripts yet. To scale out, run several brokers (a session lives in the broker that created it, so a reconnect has to reach the same one).
Best for: Real-time UIs with tool calls on the server.
Data Flow
Request Flow
Tool Call Flow
When to Use Each Architecture
Related
- Security Model - 6-layer defense explained
- AgentScript - Language subset definition
- enclave-vm - Core library documentation
- EnclaveJS - Streaming runtime