@skillsvault-io/sdk

Govern agent skills from any framework — LangChain, LangGraph, OpenAI Agents SDK, Vercel AI SDK, Microsoft Copilot / Semantic Kernel, or plain code. These have no skills directory and no harness hook, so they connect through this SDK instead.

Runtime-agnostic: built on Web fetch + WebCrypto, so the same code runs on Node 20+, edge / Vercel Functions, browsers, Deno and Bun. Zero dependencies.

Model

sync() the signed policy bundle once, then gate(skill) locally before each skill/tool use (no model, no network on the hot path). governed() wraps a tool so a banned skill throws before it runs and every use is audited.

import { createSkillsvaultClient } from "@skillsvault-io/sdk";

const skillsvault = createSkillsvaultClient({
  baseUrl: "https://skillsvault.io",
  org: "acme",
  key: process.env.SKILLSVAULT_AGENT_KEY!, // minted in Settings → API keys
  harness: "ai-sdk",                 // free-form framework id
  machine: "prod-agent-7",
});

await skillsvault.sync();                          // verify the Ed25519-signed bundle
const skills = await skillsvault.listSkills();     // detect org-approved skills
const skill = await skillsvault.getSkill("acme/deploy"); // load SKILL.md + files at runtime
const decision = skillsvault.gate("acme/deploy");  // { decision: "allow" | "warn" | "halt", reason, ... }

Framework adapters

governed(skill, fn) returns a wrapped fn — drop it in wherever a framework expects the tool's execute function.

Vercel AI SDK

import { tool } from "ai";
import { z } from "zod";

const deploy = tool({
  description: "Deploy the app",
  inputSchema: z.object({ env: z.string() }),
  execute: skillsvault.governed("acme/deploy", async ({ env }) => deployTo(env)),
});
// HALT → the tool throws; the model sees the org's reason and stops.

LangChain

import { DynamicStructuredTool } from "@langchain/core/tools";

const deploy = new DynamicStructuredTool({
  name: "deploy",
  description: "Deploy the app",
  schema: z.object({ env: z.string() }),
  func: skillsvault.governed("acme/deploy", async ({ env }) => deployTo(env)),
});

OpenAI Agents SDK

import { tool } from "@openai/agents";

const deploy = tool({
  name: "deploy",
  description: "Deploy the app",
  parameters: z.object({ env: z.string() }),
  execute: skillsvault.governed("acme/deploy", async ({ env }) => deployTo(env)),
});

LangGraph / plain code

const run = skillsvault.governed("acme/deploy", doDeploy);
await run(args); // throws if banned/halted; audits every call

Why a connector (vs the CLI)

| Client | Integration | |---|---| | Coding harness (Claude Code, Codex, Cursor, …) | skillsvault CLI: pull installs skills into its dir + a hook gates at runtime | | Agent framework / SDK (the above) | this SDK: sync → gate → governed tools, in your agent code |

Same control plane, same signed bundle, same audit. Enforcement is cooperative either way.

Publishing (next)

This proof ships as ESM + hand-authored .d.ts. A published package would add a tsup/tshy dual ESM+CJS build and per-framework entry points (@skillsvault-io/sdk/ai, /langchain, /openai).