Skip to main content
In 5 minutes, connect an existing JavaScript or TypeScript agent, run one real tool-using turn, and see exactly what it observed and did. Then add a rule in Observe mode and see what Rippletide would have blocked, without interrupting your agent.
Supported today: existing Node.js 18+ JavaScript/TypeScript agents and MCP servers. Python and other runtimes are not supported by this connection path yet.
This page walks the agent path; an MCP server follows the same motion with the MCP guide.

​
Coding-agent set up

Move into the repository that contains your agent or MCP server, open Codex, Claude Code, or Cursor there, then paste this prompt:
The CLI owns the setup complexity and guides the coding agent through the connection.

​
1. Create the connection and its key

In Agents & integrations, click Connect → Agent (or Connect → MCP). Rippletide creates the agent and mints its Connection key — copy the generated .env block immediately into a .env file at the root of your repository; the full key is shown only once:
This is the agent-scoped Connection key, not the Platform API key from Settings — never put a Platform key in an agent repository. Keep .env git-ignored. The dialog also gives you a setup prompt for step 2.

​
2. Wire up your repository

The fastest path is to let a coding agent (Cursor, Codex, Claude Code…) do the wiring: with the .env saved, paste the setup prompt into the coding agent. It installs the SDK, declares your agent’s prompts and tools, and starts the runtime worker. This should take 3 to 5 minutes. If you’d rather stay in a terminal, the CLI does the same thing:
Either way, the details are in Connect an agent.

​
3. Run one real turn

Start your agent the way you always do — the wiring changes nothing about how it runs, and it still needs its own secrets (model keys, tool credentials) in place. Send it one message that makes it use a tool.

​
4. Verify

  • Open Harness management: you should see your agent’s identity, prompts, tools, and connections.
  • Open Runtime Events: the proof completes when the agent has registered, a heartbeat has arrived, and a tool-using turn has been recorded. (MCP servers run no agent loop, so a tool call is enough — no turn.)

​
5. Your first decision

This is the moment the product clicks:
  1. On the Rules page, pick the tool your agent just used and write a rule that would forbid what it just did.
  2. Save it and set the release to Observe mode.
  3. Send your agent the same message again.
The action goes through untouched — and the Rules page shows a would block decision with the rule that fired. You have just seen what enforcement would do, with zero risk to your agent. When you trust the rule, switch that release to Enforce: see observe vs enforce.

​
If something doesn’t appear

  • Node version: the SDK needs Node.js 18 or newer.
  • .env location: it must sit at the repository root (of the package that owns the agent, in a monorepo), and stay git-ignored.
  • Your agent’s own secrets: Rippletide adds none — model keys and tool credentials must be configured as before.
  • Is the process actually running? Evidence only flows while your agent (or its worker) is running.
  • Check the wiring from the repository root:
It reads the .env exactly like the SDK and prints PASS with what the platform can see. If you’re stuck, report it to your Rippletide contact with the verify output — never share the key itself.