Give your agents safe tools

Every published script is also a tool for AI agents. Here’s how to connect an agent, and how to make sure it can’t take a real action without someone checking first.

1. Connect the agent

Add Reminix’s MCP server, https://mcp.reminix.com/mcp, to your agent. In Claude Code:

claude mcp add --transport http reminix https://mcp.reminix.com/mcp

In Claude, add it under Settings → Connectors → Add custom connector. The first time, your browser opens: sign in, choose one workspace, choose what the agent may do, and click Allow. The agent acts as you, with your access, and can be disconnected at any time under Account → Connected apps. See MCP server.

2. What the agent sees

Each published script you may use is a tool with the script’s own name (refund-customer is refund_customer), taking the script’s inputs. The agent can also search the catalog, read runs, and list secrets by name. It can’t save, change or read secrets.

3. Put a person in the loop

For any script that takes a real action, open its Settings and set Which runs need approval to Agents need approval. Through the API:

curl -X PATCH https://api.reminix.com/v1/scripts/refund-customer \
  -H "Authorization: Bearer $REMINIX_TOKEN" -H "X-Workspace: acme" \
  -H "Content-Type: application/json" \
  -d '{ "approvalPolicy": "agents" }'

Now when an agent calls the tool:

1. Reminix records the run as Awaiting approval, and the tool tells the agent it’s waiting, with a link.
2. Workspace owners and admins get a notification, and see exactly what will run, with which inputs.
3. Approved, it runs. Denied, or not decided within a day, it’s cancelled.

People on your team still run it straight away. On Business, you can also require approval for every run.

4. Publishing stays with people

An agent can write and upload a script, and test-run its draft, but its publish is always a request that a person approves. See Publish a script with your agent.

Related: Governance, Approvals.