opencode-auth-provider

A reusable runtime that mirrors OpenCode's provider discovery logic so plugins can reuse the credentials stored in ~/.local/share/opencode/auth.json (API keys and OAuth) to run clean AI calls via the Vercel AI SDK.

Features

• Loads the same opencode.jsonc configuration layers (global + workspace) and merges provider overrides
• Reads ~/.local/share/opencode/auth.json and environment variables just like the CLI
• Installs provider SDK packages on-demand (via Bun) and exposes ready-to-use LanguageModel instances
• Boots the same OAuth auth plugins as the CLI (Anthropic + Copilot by default, plus anything listed in your opencode.jsonc, e.g. opencode-gemini-auth) so refresh logic matches the CLI
• Keeps models in sync with models.dev and applies the same filtering rules (e.g. experimental flags)

Usage

import { OpencodeAI } from "opencode-auth-provider"
import { generateText } from "ai"

const runtime = new OpencodeAI({ workspaceDir: process.cwd() })

const { providerID, modelID } = await runtime.getDefaultModel()
const model = await runtime.getLanguageModel(providerID, modelID)

const result = await generateText({
  model,
  prompt: "Summarize the most recent changes",
})

console.log(result.text)

Selecting providers directly

const githubModel = await runtime.getLanguageModel("github-copilot", "gpt-4o-mini")
const response = await generateText({
  model: githubModel,
  messages: [{ role: "user", content: "Explain the PR diff" }],
})

Refreshing state

Call runtime.reset() if you update opencode.jsonc or auth.json during runtime.

Demo script

A ready-to-run script lives at scripts/run-model.ts. It uses the package to pick a model and print the response.

Prerequisites

1. Bun ≥ 1.1 installed.
2. OpenCode credentials configured (e.g. ~/.local/share/opencode/auth.json plus any opencode.jsonc files you rely on).
3. This repo checked out (e.g. /home/dan/.config/opencode/opencode-auth-provider).

Steps

cd /home/dan/.config/opencode/opencode-auth-provider
bun install           # install deps (already done if you ran it once)

Quick single-shot call

bun scripts/run-model.ts --prompt "Say hello"

• Use --provider <id> and --model <id> to override the defaults (otherwise it uses whatever opencode.jsonc marks as default).
• Use --cwd /path/to/project if you want the runtime to load configs from another workspace.
• The script outputs which provider/model were chosen and prints the model text. Any OAuth refresh handling happens automatically.

Interactive picker

bun scripts/chat-with-provider.ts

• Shows every detected provider + model and lets you pick by number or ID.
• Prompts for the message text and prints the resulting completion.
• Press Enter to accept defaults (the same ones OpenCode would use).

Notes

• The package requires Bun (matching OpenCode) because it relies on Bun.spawn, Bun.hash, and Bun's package installer for on-demand provider SDKs.
• Additional OAuth plugins can be added by listing them in opencode.jsonc and installing their npm packages alongside this library.
• By default the Anthropic and Copilot auth plugins are always loaded (unless OPENCODE_DISABLE_DEFAULT_PLUGINS=1).
• A bundled snapshot of models.dev is included so cold starts work offline; when the network is available, it continues to refresh the cache hourly just like the CLI.