ORCHESTRATORS
HELM Orchestrator Integrations
Route tool execution through the HELM execution kernel in your orchestrator framework. Orchestrators decide sequence. HELM decides whether a proposed side effect is authorized under policy, scope, and proof requirements.
OpenAI Agents SDK
Python
pip install helm
from helm_openai_agents import HelmToolExecutor
executor = HelmToolExecutor(helm_url="http://localhost:8080")
result = executor.execute("search_web", {"query": "HELM"})
print(result.receipt.verdict) # ALLOW or DENY
# Optional organization-scoped metadata
result = executor.execute(
"search_web",
{"query": "HELM"},
metadata={
"organization_id": "northstar-research",
"scope_id": "lab.discovery.search",
"principal_id": "research_lead",
},
)
# Export evidence
executor.export_evidence_pack("evidence.tar")
Proxy mode (zero code changes):
helm proxy --upstream https://api.openai.com/v1
export OPENAI_BASE_URL=http://localhost:9090/v1
python your_app.py
TypeScript / JavaScript
npm install @mindburn/helm-openai-agents
import { HelmToolProxy } from "@mindburn/helm-openai-agents";
const proxy = new HelmToolProxy({ baseUrl: "http://localhost:8080" });
const governedTools = proxy.wrapTools(myTools);
Responses WebSocket mode is not shipped in the OSS proxy runtime yet. Use the HTTP proxy surface for current OSS deployments.
See sdk/ts/openai-agents/ for full docs.
Microsoft Agent Framework (Python)
pip install helm
from helm_ms_agent import HelmAgentToolWrapper
wrapper = HelmAgentToolWrapper(helm_url="http://localhost:8080")
@wrapper.wrap_tool
def execute_query(sql: str = "") -> str:
return run_sql(sql)
See sdk/python/microsoft_agents/ for full docs.
.NET (Minimal Example)
// Source example only. No public NuGet package is currently published.
using Mindburn.Helm;
var helm = new HelmGovernance("http://localhost:8080");
var result = await helm.EvaluateAsync("deploy", new { env = "prod" });
if (result.Verdict == "DENY") throw new Exception(result.ReasonCode);
See examples/ms_agent_framework/dotnet/.
LangChain
from helm_langchain import HelmToolExecutor
executor = HelmToolExecutor(helm_url="http://localhost:8080")
# Wrap any LangChain tool
governed_tool = executor.wrap(search_tool)
See sdk/python/langchain/ for full docs.
Mastra
// Install from npm for the published OSS adapter.
// npm install @mindburn/helm-mastra
import { HelmMastraAdapter } from "@mindburn/helm-mastra";
const adapter = new HelmMastraAdapter({
helmUrl: "http://localhost:8080",
sandboxProvider: "opensandbox", // or 'e2b', 'daytona'
});
See sdk/ts/mastra/ for full docs.
For a dry-run policy simulation flow, see ../POLICY_SIMULATION.md.