Deterministic WebAssembly Sandbox for AI Agent Tools
Stop spinning up 2-second Docker containers for untrusted LLM tool calls. RunVero executes WebAssembly components with 0.49ms warm starts, 1.06 MiB RAM, deterministic instruction fuel metering, and zero-ambient authority.
$curl -sSL https://runvero.dev/install.sh | sh
runvero — execution-telemetry
$ runvero run target/wasm32-wasip2/release/mock_tool.wasm --input '{"query":"cluster_analysis"}'
Execution succeeded in 0.49ms
├── Memory Allocated: 1.06 MiB (Peak: 1.06 MiB / Limit: 16.00 MiB)
├── Fuel Consumed: 14,821 units (Remaining: 9,985,179 units)
└── Security Policy: Zero-Trust WASI 0.2 Sandbox Verified
├── Fuel Consumed: 14,821 units (Remaining: 9,985,179 units)
└── Security Policy: Zero-Trust WASI 0.2 Sandbox Verified
Output: {"clusters": 4, "anomalies": 0, "status": "deterministic_converged"}
from runvero import RunVeroSandbox
# Configure air-gapped sandbox with strict fuel and memory limits
sandbox = RunVeroSandbox(
component_path="tools/analyzer.wasm",
fuel_limit=10_000_000,
memory_limit_mb=16,
)
# Execute untrusted tool call from LLM agent
res = sandbox.execute({"code": llm_generated_query})
print(f"Completed in {res.latency_ms:.2f}ms | Memory: {res.memory_mb:.2f}MB")
# Configure air-gapped sandbox with strict fuel and memory limits
sandbox = RunVeroSandbox(
component_path="tools/analyzer.wasm",
fuel_limit=10_000_000,
memory_limit_mb=16,
)
# Execute untrusted tool call from LLM agent
res = sandbox.execute({"code": llm_generated_query})
print(f"Completed in {res.latency_ms:.2f}ms | Memory: {res.memory_mb:.2f}MB")
package runvero:agent;
interface execution {
record tool-request {
tool-name: string,
params-json: string,
context-id: string,
}
record tool-response {
success: bool,
output-json: string,
error: option<string>,
fuel-consumed: u64,
}
execute: func(req: tool-request) -> tool-response;
}
interface execution {
record tool-request {
tool-name: string,
params-json: string,
context-id: string,
}
record tool-response {
success: bool,
output-json: string,
error: option<string>,
fuel-consumed: u64,
}
execute: func(req: tool-request) -> tool-response;
}
// Bolt SIMD Zero-Exposure Secret Replacement
1. LLM Context injects placeholder: "{{RUNVERO_SECRET_0}}"
2. Sandbox guest executes without knowing raw secret key.
3. Outbound request replaces token at line speed via SIMD Aho-Corasick:
Latency: < 12 µs (Zero allocation bulk memcpy)
Result: Zero raw secrets exposed to LLM context or logs.
1. LLM Context injects placeholder: "{{RUNVERO_SECRET_0}}"
2. Sandbox guest executes without knowing raw secret key.
3. Outbound request replaces token at line speed via SIMD Aho-Corasick:
Latency: < 12 µs (Zero allocation bulk memcpy)
Result: Zero raw secrets exposed to LLM context or logs.