aca.reference/claude-under-control

How to deploy

A drop-in governing proxy you run yourself — your infrastructure, your Claude key, your policy. It is open source; nothing routes through a third party. No full platform required.

1 · Deploy it (one click — you are prompted for your key):

Deploy to Render  Docker / other

Or run it directly:

git clone https://github.com/ai-ankqush/aca-mcp-gateway
cd aca-mcp-gateway
ANTHROPIC_API_KEY=your-key GATEWAY_TOKEN=choose-a-secret node server.mjs   # :8080

2 · Point your apps at it — base URL = your gateway, key = your gateway token (the gateway holds the ANTHROPIC_API_KEY):

new Anthropic({ baseURL: GATEWAY_URL, apiKey: GATEWAY_TOKEN })
# requests go to  GATEWAY_URL/v1/messages

This model's quickstart: models/claude · Full guide, scaling to tens of thousands of employees, and per-team policy: docs/DEPLOY.md · Source: ai-ankqush/aca-mcp-gateway

What it is

Deploy Claude under control — a runnable ACA gateway plus a one-click governing proxy. Bring your own Anthropic key; point your Claude apps at the gateway and every request and tool action is governed and logged. A Policy Enforcement Point applies ACA's SEE / DECIDE / DO to every action (scope, sensitive-data boundary, human approval, action-class limits, kill switch), writes a tamper-evident evidence log, and ships a passing assurance suite (prompt-injection, data-access-boundary, tool/action-limit, approval-bypass, evidence-reconstruction). Vendor-neutral: one repo, swappable model adapters. Stateless per request, so it scales to tens of thousands of employees behind a load balancer. Zero dependencies, Apache-2.0.

Declared permissions

Scope	Value
reads	mcp-tools
writes	evidence
network	api.anthropic.com
dataScopes	internal

Shown before install. Open-track add-ons always run sandboxed with a permission prompt.

Trust

Publisher aca.reference · License Apache-2.0 · Content hash 9f0a54497ec120a01d7d6b39…

Version	Integrity	Signature	Published
v0.1.0	9f0a54497ec120a0…	unsigned	2026-09-03