@sanctum-runtime/sdk

Gate AI agent actions before they execute.
Verify tool calls, API requests, and automations with policies, optional human approval, and audit logs — built for teams shipping LLM agents, tool use, and autonomous workflows.

Part of Sanctum Runtime (MIT open core).

📖 Full capabilities: DEVELOPER_GUIDE.md

Install

npm install @sanctum-runtime/sdk
# optional verify-before-execute helper:
npm install @sanctum-runtime/adapter-agent-runtime

Quick start

import { SanctumRuntime } from '@sanctum-runtime/sdk'

const sanctum = new SanctumRuntime({
  baseUrl: process.env.SANCTUM_API_URL!, // your Sanctum API
  agentToken: process.env.SANCTUM_AGENT_TOKEN, // Console > Agents (recommended for one agent)
})

const result = await sanctum.verifyAction({
  actor: 'my-agent',
  action: 'unlock_door',
  context: { heard: 'Open the front door', time: '02:13 AM' },
})
// APPROVED | REQUIRE_VERIFICATION | BLOCKED

agentToken is the signed identity returned when an operator registers an agent in the console. Use apiKey instead for runtime hosts, administrative pipelines, or fleet-level integrations. If a process supplies both, action verification uses the agent token while control-plane requests keep the workspace API key.

What you can do

| API | Purpose | |-----|---------| | verifyAction() | Gate one action (tool call, API, command) | | policy() / registerPolicy() | Unlimited action policies | | exportPoliciesYaml() / importPoliciesYaml() | Policy-as-code | | waitForVerification() | Resume after human approve | | middleware() | Drop-in agent middleware | | getAudit() | Compliance-style audit trail |

Use riskPrompt per action for custom LLM scoring instructions. Works with Ollama, OpenAI-compatible APIs, or heuristics-only.

Easy Connect for model tools

Model providers differ in API shape, but their risky boundary is the same: before executing a named tool/function call. The provider-neutral adapter supports OpenAI, Anthropic/Claude, Google Gemini, xAI Grok, DeepSeek, NVIDIA NIM, Bedrock, local and compatible models:

import { SanctumClient } from '@sanctum-runtime/sdk'
import { wrapModelToolExecutor } from '@sanctum-runtime/adapters/model-tools'

const client = new SanctumClient({
  baseUrl: process.env.SANCTUM_API_URL!,
  agentToken: process.env.SANCTUM_AGENT_TOKEN!,
})

const runTool = wrapModelToolExecutor(executeTool, {
  client,
  provider: 'google-gemini',
  agentId: 'support-agent',
})

await runTool({ name: call.name, arguments: call.arguments })

Sanctum gates connected action/tool execution. It does not passively monitor a consumer ChatGPT, Claude, Gemini or other account that has not connected a tool boundary.

Run the API locally

git clone https://github.com/Matik103/sanctum-runtime.git
cd sanctum-runtime && cp .env.example .env
npm install && npm run dev:runtime && npm run smoke

Keywords

AI agents · LLM security · tool use · function calling · guardrails · human-in-the-loop · policy engine · Ollama · agent middleware · audit log · TypeScript

License

MIT — enterprise fleet/cloud features are separate; see OPEN_CORE.md.