Setup Adapter Build it with your AI

Build it with your AI

guide

You built an AI — a CLI agent, a local model, something entirely your own. The tool that built it can also connect it to Conduck: one brief tells your AI coding tool how to build the adapter, verify it against Conduck's own checks, and hand a working setup back to you.

1Hand your AI the brief

Paste this into the AI coding tool that knows your engine — Claude Code, Codex, or whatever you build with — and fill in the one placeholder:

paste into your AI tool
Read https://conduck.com/setup/adapter/build.md and follow it through the adapter build and verification handoff. Stop before HTTPS exposure or pairing; leave those steps to me. The AI engine to connect: <describe it - e.g. "the agent in this repo" or "the Claude Code CLI on this machine">.

The brief carries the whole workflow — here's the flight plan it walks your AI through, so you know what's happening while it works:

1

Reads the contract

Fetches the adapter spec and records the exact revision it builds against.

2

Inspects your engine

Whether it can take a whole conversation at once, how a turn runs, whether it can see images, what can run in parallel — then agrees the details with you: port, token, optional file folder.

3

Builds the adapter

An HTTP server in front of your engine, reachable only from the machine itself — plus a harmless echo mode used purely for testing.

4

Loops on the adapter check

Runs Conduck’s conformance checks against echo first, then against your real engine — fixing and re-running until ordinary and deep runs both exit green.

5

Tests what checks can’t see

Its own tests for timeouts, process cleanup, file confinement, and that nothing sensitive ever reaches a log.

6

Installs it always-on

Under a supervisor (launchd, systemd), so the adapter survives restarts — a terminal window left open is not a deployment.

7

Hands back to you

A README and self-test write-up, the values you’ll need for pairing — and the pairing step itself, which is deliberately yours.

How much of that is real work depends on your engine. One that can take a whole conversation at each turn keeps the adapter small; one that only accepts the next message inside a session it owns may need a change on the engine side — the brief has your AI raise that with you before it starts, not after.

The full, exact instructions are the adapter contract plus the brief itself — this is just the map.

2It builds until the adapter check passes

The finish line is not your AI's opinion — it's Conduck's automated conformance check. The conduck-connect script's --check-adapter mode (and its deeper --deep run) sends app-shaped requests plus contract edge-case probes and verifies the replies, exiting 0 only when everything conforms. The brief instructs your AI to loop on that check and to report the check's results, not its own summary.

  • Ask for the deliverables the brief requires: a README, a SELF-TEST.md with the adapter-check transcript, and a SPEC-FEEDBACK.md noting anything unclear in the contract.
  • Spec feedback is welcome upstream — the brief has your AI draft a fully redacted report and asks you before anything is submitted.

3Continue into setup and pair

Your AI stops after it builds, verifies, and installs the adapter. It hands you the port, token, and working folder, but it does not expose the server or pair a device. Start setup yourself in a real terminal:

on the adapter's machine
curl -fsSLO https://github.com/gigaduckai/conduck-connect/releases/latest/download/conduck-connect.sh && bash conduck-connect.sh --setup

Setup reports any OpenClaw or Hermes installation it finds, then asks what you want to configure. Choose the adapter your AI built. It helps put HTTPS in front (Tailscale, cloudflared, or a front you already run), verifies with real requests, and prints the setup code as a QR code. Whichever route you take, the certificate has to be one your devices already trust — a self-signed one is refused outright, and no app setting can override that; the certificates and TLS guide lists three free ways to get a trusted one. Setup needs a real, interactive terminal and asks a few questions — the port, the token, the working folder. Your AI will hand you those values; you type them and scan the code. That's deliberate: pairing is the moment a person confirms what goes live.

The code is an open format - a gateway may mint it natively instead of running the script; the minting rules live with conduck-connect on GitHub.

What to expect

  • This flow is tested. AI coding tools with no prior knowledge of Conduck have built adapters from the public contract alone — more than twenty independent builds on live servers across several rounds. What those builds got wrong is what revision 1.10 of the contract now spells out, including one fault a loopback check can never see.
  • The adapter check changes no host configuration. It sends live requests that may use compute or enter server history; the optional --files profile also writes and deletes small probe files in the configured folder. Setup starts only after you accept the continuation question, and it asks before each configuration change.
  • Deployment is part of the job. The brief has your AI install the adapter under a supervisor so it survives restarts — Conduck expects an always-on server, and a terminal window left open is not one.
  • Your AI, your devices. The adapter connects your own engine — your key, your login, your machine — to your own devices. Don't expose it to third parties or resell access.