Fx

Fx is a coding agent from Vercel Labs, written in Zig as a single native binary and model-agnostic. Custom model connections are defined in the user-level ~/.fx/settings.json; point the base URL at aicentos.

One pitfall to call out first: .fx.json in the project root is a different thing. It deliberately disallows most fields (including models) so that a cloned repository cannot swap out your model or endpoint. Custom endpoints must be written in ~/.fx/settings.json — putting them in .fx.json has no effect.

PrerequisitesCopy anchor link

First create a dedicated key for Fx on the console's "API Keys" page (naming it fx makes usage easier to track):

• Console: https://aicentos.com
• Create a key: https://aicentos.com/keys
• Top up when the balance runs out: https://aicentos.com/wallet

Confirm Fx is installed:

curl -fsSL https://fx.sh/setup.sh | bash
fx --version

Configuration to fill inCopy anchor link

Setup stepsCopy anchor link

1. Edit the user-level settingsCopy anchor link

In ~/.fx/settings.json, register a connection following the "custom model connection" format. The field structure of the connection object is defined by the official documentation:

• Documentation: https://fx.sh/docs
• README: https://github.com/vercel-labs/fx

Set the API address of the connection to:

https://aicentos.com/v1

For the API Key, use the one you created in the console; take model names from the Model Gallery in the console.

2. Point Fx at the connectionCopy anchor link

Fx lets you specify the provider and model at launch:

fx provider local FX_MODEL=gpt-4o fx ask "Answer in one sentence: what is 1+1?"

Or specify it with an environment variable:

export FX_MODEL="gpt-4o"

3. Confirm the current providerCopy anchor link

Fx remembers your choice per provider; the next launch reuses the previous provider.

TestCopy anchor link

First verify connectivity with a side-effect-free prompt:

fx ask "Answer in one sentence: what is 1+1?"

A normal response means the integration works. On authentication errors, first check whether the base URL includes /v1 and whether the API Key has stray spaces.

NotesCopy anchor link

• Do not touch .fx.json. The project-level file only allows a few fields such as provider_order and provider_strict, has no model field, and is not used to configure custom endpoints.
• provider_order / provider_strict are Vercel AI Gateway routing parameters; they only apply to Gateway requests — custom model connections ignore them. Do not expect them to switch the upstream.
• Official support for custom model connections in Fx is still evolving; older versions may need a newer build to recognize them. After upgrading, run the test command above once before real use.
• The binary install script downloads from the network; verify that you trust the source before running it.

Related linksCopy anchor link

• Project home: https://github.com/vercel-labs/fx