opencode

My setup in the console prints this block with your own routes in it. See Set up your coding tool. This page explains what it prints.

The block

opencode is configured by a file, ~/.config/opencode/opencode.json. There is no environment-variable form.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "panorama": {
      "name": "Panorama",
      "npm": "@ai-sdk/anthropic",
      "options": {
        "baseURL": "https://panorama.your-company.internal/v1",
        "apiKey": "pnt_..."
      },
      "models": {
        "standard": {
          "name": "standard",
          "headers": { "user-agent": "opencode (panorama)" },
          "limit": { "context": 1000000, "output": 64000 }
        }
      }
    }
  },
  "model": "panorama/standard",
  "small_model": "panorama/fast"
}

Merge it into your existing file if you have one.

• baseURL ends in /v1. Claude Code’s does not. The two tools build their request paths differently, so the same gateway has two addresses on My setup on purpose.
• Every route you may use is declared as a model, with its real context window and output limit beside it. Copy these numbers from My setup and do not edit them: they come from the model your route runs on.
• model and small_model are your everyday route and your background route. Without them opencode has a provider and no model selected.
• user-agent is how Panorama recognises opencode traffic.

If your routes include models that are not Claude, more providers appear beside the first, and each route is listed under one of them:

Provider	Package	For routes on
panorama	@ai-sdk/anthropic	Claude models
panorama-openai	@ai-sdk/openai	Models that speak OpenAI’s Responses API, such as GPT and gpt-oss
panorama-chat	@ai-sdk/openai-compatible	Models that speak only Chat Completions, such as Qwen and DeepSeek

A GPT model is sent the Responses API because it refuses the tools opencode sends on Chat Completions. If opencode fails on a GPT route with “The model provider rejected this request”, your file sends that route to Chat Completions: copy the block from My setup again.

Keep the token out of the file

Replace the apiKey value with a reference to an environment variable, and export the token in your shell:

"apiKey": "{env:PANORAMA_TOKEN}"

export PANORAMA_TOKEN=pnt_...

That way the file can be shared or committed and the credential cannot.

The token lasts 30 days and nothing renews it. Mint a new one on My setup and update the variable. See Set up your coding tool.

Web search

When your administrator has turned on web search, the block My setup prints also has an mcp section, beside the providers:

"mcp": {
  "panorama": {
    "type": "remote",
    "url": "https://panorama.your-company.internal/mcp",
    "enabled": true,
    "headers": { "Authorization": "Bearer {env:PANORAMA_TOKEN}" }
  }
}

It gives opencode a web_search tool, served by Panorama. opencode mcp list shows it as connected. It carries the same token as the providers, so the same {env:PANORAMA_TOKEN} reference works for both.

Switching routes

Use opencode’s model list. Every route you may use is in it. You will need this when a budget passes its soft limit and your route stops: the refusal names the routes that still work. See When a request is refused.

How opencode compares to Claude Code here

Better: it knows your route’s real context window. A route name is not a model any tool has heard of. opencode is told the window outright in its provider block. Claude Code has to be told through two extra environment variables, and assumes 200k when they are missing.

Worse: attribution stops at the repository. opencode sends its working directory and no branch, so Panorama can tie your work to a repository and not to a ticket. Claude Code sends the branch. If ticket-level cost matters to your team, that is the difference. See How attribution works.

Worse: no exact-attribution step. opencode’s headers are fixed strings in a file, so a branch written into it would be wrong at your next checkout.