OpenCode + WordPress: MCP setup for the free, open-source coding agent

What OpenCode is

OpenCode is an open-source AI coding agent for the terminal, built by SST, with its home at opencode.ai. It speaks the Model Context Protocol, which is the part that matters here: every server listed under mcp in its config becomes a set of tools it can call. Respira is one of those servers, and the site it acts on is your own WordPress install.

Two things set it apart from the other agents on this site. It ships free models through OpenCode Zen, with ids such as opencode/big-pickle, next to paid provider models such as anthropic/…. And its config shape is its own, which is where most first attempts go wrong. Both are covered below.

Looking for the no-key, no-subscription version? freecode runs the OpenCode engine and reads the same file, so the block on this page works there too.

The setup, in five steps

1. Install OpenCode.

   Get it from opencode.ai. It runs in the terminal and its free models need no API key, so the only value you have to find is the Respira one in the next step.

2. Open the Respira setup page and choose OpenCode.

   The page reads the sites on your account and builds an opencode.json with all of them inside. If you would rather write the file yourself, copy the RESPIRA_CONFIG_B64 value and paste it into the block below. Either way one entry covers every connected site, and you move between them by naming the site in chat.

   Open dashboard setup How the MCP server works
3. Save it where OpenCode looks.

   opencode.json at the root of the project you open OpenCode in, or ~/.config/opencode/opencode.json for every project. If OpenCode is already running, start it again so it reads the file.

4. Check the connection.

   opencode mcp list

   The respira-wordpress entry should report connected. If it does not, the shape section below is the first place to look.

5. Run a first task.

   opencode run "using respira-wordpress, list my connected WordPress sites and which page builder each one runs"

   Real site names coming back is the test. A tool list that looks right but returns nothing is usually an access level set to read-only when you wanted content.

The config

This is the block that was tested. Replace the placeholder with the base64 value from your dashboard and keep everything else as it is.

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "respira-wordpress": {
      "type": "local",
      "command": ["npx", "-y", "@respira/wordpress-mcp-server"],
      "environment": {
        "RESPIRA_CONFIG_B64": "<paste from the Respira dashboard>"
      },
      "enabled": true
    }
  }
}

The file at the project root travels with the project. If the project is in git, add opencode.json to .gitignore, or use the global file under ~/.config/opencode/ instead, so the credential never sits in a repository.

Free models, and what one did on a real run

OpenCode Zen carries a set of free models next to the paid providers. On the day this page was written the ids included:

• opencode/nemotron-3.5-lightning-free
• opencode/mimo-v2.5-free
• opencode/big-pickle

On a real machine, with the block above in place, a free model was asked to list the connected sites. It called respira_list_sites and listed five WordPress sites correctly. A second run called respira_get_site_context and respira_get_builder_info. The tool calls worked every time; the model itself wandered between calls and needed a retry or two before it settled.

That is the honest picture of a small free model against a live site. It is fine for a look around, a text change, a page duplicate, or a scan. For a multi-page build, a paid model inside OpenCode holds the thread better. Whatever the model, three habits keep the result clean:

Keep each task small and specific.

Name the site, name the page, describe one change. "On site A, duplicate the pricing page and change the headline to X" gets done. "Improve the site" does not.

Read the duplicate before you approve it.

Every structural write lands on a copy. The Approvals and Undo screen in wp-admin shows what changed. Publish from there, not from the chat.

Let the agent retry, and stop it when it loops.

A free model that repeats a call is not a broken server. Restate the task with the page name in it, or move the task to a paid model.

Before you point it at a client site

The config value is a live credential.

RESPIRA_CONFIG_B64 carries access to every site in it. Keep the file out of git and off shared machines. Rotating a site token in the Respira dashboard cuts off every client holding the old one, OpenCode included.

Pick the access level per site.

Read-only for a look around, content for text, builder for layouts, full only when you mean it. The level is set on the Sites page in your dashboard, and a free model cannot exceed it however it wanders.

Nothing goes live from the chat.

Duplicates and drafts wait in the Approvals and Undo screen in wp-admin until you publish them. If the agent got it wrong, throw the copy away and the client never sees it.

Why OpenCode plus Respira fits

A free, open-source agent against a production-safe backend.

OpenCode runs on your machine on models that cost nothing. Respira runs on your hosting. Neither side asks you to move the site anywhere.

Native builder writes.

Elementor stays Elementor, Bricks stays Bricks, Divi stays Divi, across 17 builders. Nothing is flattened to HTML in transit.

Snapshots and duplicates on every write path.

A small model moves fast and sometimes sideways. Undo has to be faster.

The same account everywhere.

One Respira account works in OpenCode, freecode, Claude Code, Cursor and Codex. OpenCode can also shell out to the Respira CLI for JSON output, and when you would rather not open a terminal at all, Respira AER is the same builder intelligence in a chat in your dashboard.