Connect OpenAI Codex

verified on 2026-10-09: MCP in Codex · Configuration reference

1. 1

   Add the server

   Add the entry with only url (see the file below without bearer_token_env_var).

2. 2

   Log in

   Run codex mcp login filmo; the browser opens Filmo's consent screen.

• If the server advertises scopes_supported, Codex prefers those scopes during login; otherwise it uses the scopes in config.toml. Filmo's resource metadata advertises every scope: choose in the consent screen.
• By default Codex uses a client metadata document when the server supports it and dynamic registration otherwise.

• Only the *:read scopes the client asked for are checked (every read scope when it asks for none).
• Write scopes, workflows:run and ai:ask always start unchecked, even if the client asks for them; the ones it asked for carry the label “Requested by the integration”.
• Check by hand only what the assistant needs, and leave ai:ask unchecked unless you want it to spend AI credits.

To change the scopes later, remove the connector and add it again.

Before you approve, read Authentication.

1. 1

   Export the key

   Set FILMO_API_KEY in the environment Codex starts from.

2. 2

   Name the variable

   bearer_token_env_var holds the name of the variable, never its value.

[mcp_servers.filmo]
url = "https://app.filmo.io/api/v1/mcp"
bearer_token_env_var = "FILMO_API_KEY"

Rental operations (21): orders:read, orders:write, availability:read, items:read, contacts:read, account:read. It includes the order writes, cancelling included: pair it with the approval mode below, or pick a preset marked (read).

Every preset and its size are in Limits and context.

default_tools_approval_mode accepts auto, prompt, writes and approve; writes prompts for tools not marked read-only. Per tool: approval_mode.

The tools that must ask for approval are in Safe writes.

• Scope step-up and listing tools again after login: not documented by the vendor.

Ask the assistant: “Which Filmo account am I connected to?” It should call filmo_account_get, a read tool that returns no personal data.