Skip to main content

Connect OpenCode to DIAL

In this tutorial you register DIAL as a custom provider in OpenCode, the terminal coding agent, then run a prompt against it. OpenCode speaks Chat Completions, so it talks to DIAL directly — no proxy needed. Every turn then runs on your DIAL deployment. This tutorial is for developers who already have a DIAL host and OpenCode installed.

Prerequisites​

What you'll build​

DIAL registered as a custom OpenCode provider. By the end, opencode run sends a prompt to your DIAL deployment and prints the reply.

Step 1: Set your DIAL key​

Bash or Zsh:

export DIAL_API_KEY="<DIAL_API_KEY>"

PowerShell:

$env:DIAL_API_KEY = "<DIAL_API_KEY>"

Step 2: Add DIAL as a provider​

Create opencode.json in your project directory (or globally at ~/.config/opencode/opencode.json; on Windows, %USERPROFILE%\.config\opencode\opencode.json). DIAL's deployment id is part of baseURL, so add one provider/model entry per deployment.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "dial": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DIAL",
      "options": {
        "baseURL": "https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.5-2026-04-24",
        "apiKey": "{env:DIAL_API_KEY}"
      },
      "models": {
        "gpt-5.5-2026-04-24": {
          "name": "GPT-5.5 (DIAL)",
          "limit": { "context": 200000, "output": 128000 }
        }
      }
    }
  }
}

When you give a model a limit block, OpenCode requires both context and output. Set output to the deployment's maximum completion tokens — 128000 for gpt-5.5-2026-04-24. DIAL rejects a turn that asks for more than the ceiling with max_tokens is too large, so never set it higher than the deployment allows.

Note

Pick your deployment with the wire in mind. @ai-sdk/openai-compatible sends a plain max_tokens and no api-version, which gpt-5.5-2026-04-24 accepts. Some other GPT-5.x deployments reject a plain max_tokens (wanting max_completion_tokens) or require an api-version, and OpenCode supplies neither — see Azure-backed deployment settings. DIAL accepts the key as a Bearer token via apiKey; if your DIAL only accepts the Api-Key header, add "headers": { "Api-Key": "{env:DIAL_API_KEY}" } under options.

Step 3: Run OpenCode​

opencode run --model dial/gpt-5.5-2026-04-24 "Reply with exactly: PONG"

Verify:

You see PONG. Launch plain opencode for the interactive TUI and pick the DIAL model from the model switcher.

What you learned​

  • How to register DIAL as a custom @ai-sdk/openai-compatible provider in opencode.json.
  • How to set a model's limit.output to the deployment's completion-token ceiling.
  • How to run a prompt against your DIAL deployment from the terminal.

Good to know​

  • OpenCode uses Chat Completions, so it reaches DIAL directly — no proxy needed, unlike Codex. Deployments exposed only over the Responses API are not reachable this way.
  • When a model has a limit block, OpenCode requires limit.output. Set it to the deployment's max completion tokens; DIAL rejects a value above that ceiling with max_tokens is too large.
  • Add one provider entry per DIAL deployment — the deployment id lives in baseURL.
  • Tool calling and vision work only when the DIAL deployment advertises those capabilities.

Next steps​