Authorization: Bearer header). It gets tools for site files, builds, the catalogue and — if you enable them — orders and discounts. All its changes land in the draft in Workspace.
Step by step
1
Create a workspace token
In Workspace → MCP → New token pick the smallest sensible set of scopes — usually
site:write and catalog:read. You see the token (iswt_…) only once. Scopes are described on the AI Workspace page.2
Set the environment variable
The configuration never contains the token — it refers to the
ITEMSHOP_TOKEN variable.3
Pull the site project
In an empty folder download the current draft (CLI):
4
Save the MCP configuration
Save the file from the next section in the project folder.
5
Start the agent
MCP configuration file
.codex/config.toml
.codex/config.toml works in projects you marked as trusted. Prefer global settings? Append the same section to ~/.codex/config.toml. bearer_token_env_var names the variable, not the token itself.
In Workspace → Agent choose the same agent and the panel shows ready commands, files and a zip bundle for your shop. Offline you can generate them with:
Instructions file
The panel also generatesAGENTS.md — the file OpenAI Codex CLI reads at start-up. It holds your shop’s context (name, id, API and panel addresses), the list of enabled MCP tool groups and the project rules: keep itemshop.site.json, data only through the SDK, checkout via redirect with two consents, no secrets in code, build and file limits, publishing policy. Copy it into the project root.
Check the connection
In the agent runcodex mcp list — the itemshop server should be connected. You can also click Test connection in Workspace → MCP. Start with something safe, e.g. “Show the shop context and the list of site files”. Tool calls then show up in the MCP log and on the activity timeline (source mcp).
Skills
Install agent skills with a command, pointing at this agent’s skills folder:Troubleshooting
All agents
Comparison and shared rules.
MCP
Tools and scopes.