Skip to content
aizkaizk

OpenCode

This page uses the address and public OAuth client of the AIZK deployment serving it. If you have not connected yet, Quickstart lets you choose OpenCode and ask it to complete the whole setup.

OpenCode is the shortest of the three client setups. One committed file, one sign-in command, and one command for when something looks wrong.

The configuration file

Section titled “The configuration file”

OpenCode reads its servers from opencode.json in the project root.

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "aizk": {
      "type": "remote",
      "url": "https://xpc3f5mjuv5edgnsfsfbffcoda0ckvij.lambda-url.ap-southeast-1.on.aws/mcp",
      "oauth": {
        "clientId": "2yqsd1rbjsxxzkyrem1xv",
        "scope": "control offline_access openid",
        "callbackPort": 8912,
        "redirectUri": "http://127.0.0.1:8912/callback"
      },
      "enabled": true
    }
  }
}

"type": "remote" is the field that matters. aizk is not a local process OpenCode starts, it is an HTTP endpoint OpenCode talks to, and the remote type is what carries the OAuth handshake. The file contains no credential, so commit it and let everyone on the project share it.

Sign in and check

Section titled “Sign in and check”
Terminal window
opencode mcp auth aizk
opencode mcp debug aizk

The first opens a browser, you sign in as yourself, and OpenCode stores the result on your own machine. The second prints what OpenCode currently knows about the server, which is the fastest way to tell a configuration problem from an authorization problem.

What to check after signing in, from tools appearing to a clean status callyesnonoyesrun mcp authtools appear?call statusrun mcp debugserver reached?fix url in opencode.jsonsign in again

Seeing the five tools listed is not quite proof. Ask OpenCode to call status and confirm it returns your name and your organizations. That is the first call that needs a token aizk actually accepted.

What you get once it works

Section titled “What you get once it works”

Five tools and one resource, the same set every client sees. status reports who you are and what is still processing, find answers one question with evidence, keep stores a note or a file, report sends a memory problem to the operator, and share copies documents into a team. MCP tools is the full surface with every parameter.

Committing opencode.json shares the connection and nothing else. Each person signs in as themselves and each person’s memory stays private by default, so a shared config file never implies shared memory. Only naming an organization does that, which Scopes explains.

When the browser is elsewhere

Section titled “When the browser is elsewhere”

Sign-in ends with the browser handing its result to a loopback address on the machine running OpenCode. That address belongs to OpenCode and not to aizk, so when the browser and OpenCode are on different machines the redirect has nowhere to land. The fix is an SSH forward for that one port, and the shape of it is drawn out on Codex.

If you would rather avoid forwards entirely, Claude Code has a paste-the-URL flow that needs no open port.

Agent instructions

Section titled “Agent instructions”

OpenCode reads AGENTS.md. The rules worth putting there are written out in full on Claude Code and they are the same for every client. The habit that matters most is treating found content as evidence rather than instructions, because shared memory means an agent can read text a teammate wrote.

Next

Section titled “Next”