Codex with Sign in with ChatGPT
Use providers/codex to call the Responses API directly with an eligible ChatGPT
plan. You need a browser for the first sign-in. Loom sends the conversation and
tool definitions on each model call, so you can run your graph without installing
the Codex CLI. OpenAI’s direct inference guide
Run the example
From the repository root, start the interactive coding assistant. It opens a browser on first use, then reuses the saved registration on later runs.
go run ./examples/codexTo run one turn, pass a prompt. The shared shell_command tool lets the
assistant inspect the current directory.
go run ./examples/codex -prompt "Use ls to list the files in this directory."The example selects the first visible model in the account catalog by default.
Set CODEX_MODEL to choose a default or pass -model for one run. To see
available IDs and start a session with a particular model, run the following commands.
go run ./examples/codex -list-modelsgo run ./examples/codex -model MODEL_IDThe example keeps one account registration in loom/codex-example/session.json
under the OS user configuration directory. Run one instance at a time so a
single process owns token refresh. Use -login again to reauthorize the saved
account. Chat history lasts for the current process. Example source
Register and sign in
Use this integration in eligible open-source and locally hosted applications.
Generate a host ID with loomcodex.NewHostID() once and save it before starting
sign-in. Keep each account registration in a separate protected record keyed by
the issued client ID. Persist credentials atomically with owner-only permissions
(0600 on Unix). Pass a saved registration in LoginOptions.Previous to
reauthorize it. Registration and sign-in
The following function runs first-time sign-in and builds a provider. The application supplies a stable host ID and a function that saves the selected registration. Pass an application-lifetime context for the refresh source.
import ( "context"
loomcodex "github.com/masterkeysrd/loom/providers/codex")
func connectCodex( ctx context.Context, hostID string, save func(loomcodex.Credentials) error,) (*loomcodex.Provider, error) { creds, err := loomcodex.Login(ctx, loomcodex.LoginOptions{ AgentName: "My Loom App", HostID: hostID, }) if err != nil { return nil, err } if err := save(*creds); err != nil { return nil, err } source, err := loomcodex.NewTokenSource(ctx, creds, save) if err != nil { return nil, err } return loomcodex.NewProvider(source, nil)}On subsequent runs, load the selected credential record and call
NewTokenSource directly. It renews access near expiry and saves the rotated
token set before returning the refreshed access token. Use one source per
registration and coordinate refreshes if multiple processes share a credential
record. Applications that already manage direct ChatGPT OAuth can pass their
existing oauth2.TokenSource to NewProvider. Accounts and sessions
Choose a model and bind tools
provider.ListModels(ctx) returns the visible model slugs for the selected
account. Pass the desired profile’s ID to
llm.NewModel(provider, modelID, nil). Use model.BindTools(myTool) and the
usual Loom tool loop. If your graph checkpoints a pending call, persist the full
assistant message. On resume, execute the tool by name with its saved arguments
and attach the original call ID to the tool result. Send the restored history
and tool result together on the next model invocation.
Each inference request uses streaming, sends the conversation history, and sets
store: false. Loom preserves encrypted reasoning items in the serialized
assistant message so the next request can include them.
Preview limits
The provider omits MaxTokens, Temperature, and TopP on this route.
Function tools are supplied as Responses additional_tools. Hosted file
search, Code Interpreter, tool search, and audio or video inputs are unsupported.
Web search availability depends on the model and account policy.
Responses API preview limitations