@adaptivemcp/approval

Intent, plan, and tool approval experiments for Adaptive MCP.

Status: published to npm as @adaptivemcp/approval. It is the enforcement hook for autonomy boundaries.

approval is the enforcement hook for autonomy boundaries. Given a tool name, it returns an allow/confirm/deny decision based on its learned risk and flakiness.

Usage

import { MemoryStore } from "@adaptivemcp/memory";
import { ApprovalGate } from "@adaptivemcp/approval";

const memory = new MemoryStore();
const gate = new ApprovalGate({ memory });

// Before a tool runs:
const decision = gate.gate("deploy_service");
// -> "allow" | "require_confirmation" | "deny"
// High-risk tools (annotation risk: "high") or flaky tools require confirmation.

What it does

• gate(toolName) → "allow" | "require_confirmation" | "deny".
• isFlaky(toolName): predicate used by the gate and other packages.

Defaults

| Option | Default | | --- | --- | | confirmRiskLevels | ["high"] | | flakyFailureRate | 0.2 | | minInvocations | 10 |

Open questions

What should users approve? Can approval boundaries adapt? How much autonomy should agents receive? See the root AGENTS.md.

License

Released under the MIT License. Copyright (c) 2026 Kemal Elmizan.