Bring agent controls
into your organization.
Create your account, install your Node and connect a real agent workflow. Your local administrator manages permissions, approvals and audit records in your installation.
Use Linux x86-64 or Mac with Apple Silicon, with Docker Engine 28.1 or later and Docker Compose running. The download includes the compiled node and PostgreSQL images.
Start in Agent Control Center
Create your account using an email address you can access. Then name your first node so your team can recognize it.
Create account →Download for your machine
Choose the installer in the Center and extract it into a new folder. On Mac, start Docker and double-click Install.command. On Linux, start Docker and run ./install.sh from the extracted folder.
The installer selects an available port, starts the node and displays its console URL and permanent Node ID.
Link it to your account
Copy the Node ID into the Center and select Link node. If the node can reach the Center, its status updates after it checks in. You can also register a node that operates without an internet connection.
Create your local administrator
Open the console URL printed by the installer. Use the installation setup key from runtime.env to create your first administrator. Choose a password of 15–128 characters.
Your local administrator manages policies, tools and approvals for this installation. Your Center account manages node registration and connectivity.
Connect your MCP servers
Add the MCP servers you operate to the upstreams list in config/acl.local.json. A stdio server needs a command and arguments; a Streamable HTTP server needs its URL and any required headers. Keep secrets local. Restart the node after changing upstream configuration. Check Overview for discovery errors and the MCP endpoint.
{
"id": "my-tools",
"prefix": "work",
"transport": "stdio",
"command": "node",
"args": ["/absolute/path/to/your-mcp-server.mjs"]
}
Register an agent identity
Open MCP clients. Enter a unique Client ID, Agent ID, and Tenant, then select Create client and credential. Save the credential in your agent’s secret configuration: it is shown once. Use a separate identity for each agent. The screenshot shows an existing fictional demo client; your fresh installation starts without clients.

Define what the agent may do
Open Policies. Enter a rule ID, choose Allow, enter the registered client ID, and use the exact tool name from Tools. For the billing demo, use billing__create_draft and choose Human approval in Review mode. Select Save rule. Unmatched calls are denied; a matching Deny takes precedence. Policy changes invalidate earlier pending approvals.

Authorize the exact tool definition
Open Tools and select Review definition for each intended tool. Read its description and input schema, then select Authorize this definition. A policy alone does not authorize a tool definition. New or changed definitions need another review before calls can proceed.

Connect your agent to the Local Node
In an MCP client that supports Streamable HTTP, configure the endpoint below and send the client credential in the Authorization header. Use the address shown in your own Overview. Screenshots use a separate demo instance on port 8501; the default installation uses 8500. Route governed tool calls through ACL and restrict direct access to upstream servers.
Endpoint: http://127.0.0.1:8500/mcp
Authorization: Bearer <your-client-credential>Review a request and let the agent resume
For a rule requiring human approval, the agent includes a stable _meta.acl.requestId on its tool call. ACL returns a pending approval ID without executing the action. In Approvals, open the request, view its full parameters, then approve or reject it. Approval allows the agent to call acl__approval_resume with only the approvalId. The original parameters are retained. Use acl__approval_status to check progress. An unknown outcome must be investigated before attempting a new action.
{
"name": "acl__approval_resume",
"arguments": { "approvalId": "<approval-id>" }
}For the included billing demo, the original MCP tool call can look like this. Replace the fictional account and request IDs with values for your own evaluation.
{
"name": "billing__create_draft",
"arguments": {
"accountId": "demo-account",
"amountCents": 1000,
"currency": "USD",
"description": "Installation walkthrough"
},
"_meta": { "acl": { "requestId": "walkthrough-001" } }
}The screenshot shows a completed demo request. Pending requests present the decision controls after you review the full parameters.

Add operators and review the audit
In Operators, an administrator can add colleagues as Administrator, Approver, or Auditor. New operators change their temporary password on first sign-in. Approvers can review requests; auditors have read-only, redacted access. Use Audit to follow decisions and outcomes. Preserve runtime.env and the database in secure backups, including ACL_ENCRYPTION_KEY: encrypted records depend on that key.
Troubleshooting
The console does not open
Start Docker and run the installer again. Use the console URL printed when installation completes. The installer selects an available port automatically.
A tool call is denied
Check the client credential, an Allow rule for that client and tool, and the tool’s exact definition authorization. A Deny rule takes precedence. Changed definitions need a new review.
I cannot sign in
The setup key only creates the first administrator. After setup, use the account username and password. An installation owner with local database and encryption-key access can recover an existing operator using the command below; the password is entered interactively.
docker compose --env-file runtime.env run --rm local-node recover-operator --username <existing-username>An approval has an unknown outcome
Inspect the upstream system and audit trail before creating a new request. ACL does not automatically resend an uncertain action.
What can I share with support?
Share the step, error message, and a redacted screenshot. Do not send runtime.env, passwords, setup keys, client credentials, or full tool parameters.