Keyturn

Your AI proposes. You turn the key.

Run IT for many organizations in natural language — with a human hand on every change.

A local command center that lets your AI assistant (Claude, Cursor, Codex, Devin, any MCP client) investigate and operate your RMM fleet — while every endpoint action waits for your approval, signed with a passkey.

Try the demo · Setup · Features · Tools · Connect an AI client · Security model

Demo tenant with fictional data · the chat panel stands in for any MCP client (Claude, Cursor, Codex, …)

Why

MSPs and IT teams already have an agent on every machine. What they don't have is a safe way to let an AI use it. Keyturn is that missing layer:

• Talk to your fleet. "Which DHCP scopes are almost full?" "Why is the reception PC offline?" "Inventory the GPOs on the domain and flag anything unlinked." The assistant answers from collected evidence — not guesses.

• Nothing runs without you. Every endpoint action is an immutable, hashed plan that waits for a human. With a YubiKey, Windows Hello, or Bitwarden passkey enrolled, an AI holding every token on the box still cannot approve anything.

• Everything is remembered. Findings, questions, answers, decisions, and receipts are durable. The next session — or the next AI — picks up where the last one left off.

• One pane per client, or all of them. A global organization scope follows you across every page.

• Local by design. No cloud service, no telemetry. Your RMM credentials never leave your machine.

Related MCP server: enterprise-agent-lab

Try it in 60 seconds

No RMM account needed. The demo runs the full UI against a fictional MSP with three clients and ~85 devices, sandboxed and read-only:

git clone https://github.com/Solve4x-ai/Keyturn.git
cd Keyturn
npm install
npm run demo

Your browser opens at http://localhost:39399. Press Ctrl+C to stop; npm run demo -- --reset regenerates the data.

Screenshots

Overview — fleet health, a server constellation with DC / DNS / DHCP roles taken only from collected evidence, what's waiting on you, and operation success at a glance.

How it works

flowchart LR
    AI["AI client<br/>(Claude · Cursor · Codex · Devin)"] -- "MCP · stdio" --> MCP["Connector layer<br/>reporting · command profiles"]
    MCP -- "create_plan" --> Store[("Local evidence store<br/>SQLite · per tenant")]
    Human["You"] -- "review + passkey" --> UI["Keyturn dashboard<br/>localhost · bearer + WebAuthn"]
    UI -- "approve · dispatch" --> Store
    Store -- "approved plan only" --> RMM["RMM API<br/>(NinjaOne connector)"]
    RMM --> Agents["RMM agent<br/>on every endpoint"]
    Agents -- "receipts + evidence" --> Store

1. Collect — read-only diagnostic runbooks (AD health, DNS, DHCP, GPO, …) run through an approved PowerShell runner. Results land as timestamped observations with explicit coverage, so "not found" is only claimed where a complete enumeration ran.

2. Understand — deterministic rules raise findings; the AI adds context, questions, and recommendations in the Review Center.

3. Propose — endpoint work becomes a plan: exact script, targets, parameters, and a hash that binds them.

4. Approve — you approve in the local UI. Approvals expire, are single-use, bind to the plan hash, and require a passkey once one is enrolled.

5. Execute & verify — dispatch is idempotent, canary batches gate the rest of the fleet, and "accepted" never means "done" until a receipt proves it.

Safety model

Details: docs/security-model.md.

Use it with your NinjaOne tenant

Requirements: Windows, Node.js 22.13+, and two NinjaOne API applications (an API Services app for read-only reporting and a Native app for commands). The UI and MCP server are cross-platform Node, but the endpoint runbooks and setup scripts are PowerShell-first.

git clone https://github.com/Solve4x-ai/Keyturn.git
cd Keyturn
npm install
npm run build
Copy-Item config\reporting.env.example config\reporting.env
Copy-Item config\command.env.example   config\command.env
Copy-Item config\policy.example.json   config\policy.json

Then follow SETUP.md — about 15 minutes — to create the NinjaOne apps, authorize the command profile, start the dashboard, enroll a passkey, and connect your AI client from Settings → MCP clients.

The tracked policy example disables every write. Enable only what you need, per organization.

Development

npm run build     # TypeScript → dist/
npm test          # mocked tests — never touch a live tenant
npm run verify    # build + tests + PowerShell runner tests

The UI is vanilla ES modules and CSS (oklch design tokens, View Transitions, container queries) — no framework and no build step. See CONTRIBUTING.md.

Roadmap

Additional RMM connectors (the core is connector-agnostic) · SNMPv3 network-edge agent for switches, firewalls, APs, and printers · reboot / service / patch actions as approved runbooks · scheduled management reports · notifications · OS-account separation for the command server. Details in FEATURES.md.

Support

• Questions and ideas: Discussions, or support@solve4x.ai

• Bugs: GitHub issues

• Security: see SECURITY.md — please report privately

License

MIT. Originally derived from NinjaOneMCP. The NinjaOne connector is built on the NinjaOne public API; this project is not affiliated with or endorsed by NinjaOne.