Connect a model¶
The five-minute quickstart uses the offline echo provider. To use a
real model, choose the account or local server you already have. Colossus can list
available models and create a configuration for the one you select.
1. Open the workspace you want to use¶
Run setup from a repository or folder that does not already contain
.colossus/config.yaml. The --local command creates a configuration there and will
not overwrite one.
Already using the quickstart folder or another configured workspace? Run
colossus config effective to find its active file, then follow the matching
provider guide to update that configuration.
2. Choose a connection¶
Select one method. Setup loads the provider's model catalog and asks you to choose a model by number or exact ID.
Sign in through the official Codex CLI, then choose a model available to your subscription:
A ChatGPT/Codex subscription uses its own sign-in; it does not use an OpenAI API key. See subscription setup for account and model details.
Set OPENAI_API_KEY in the current terminal, then run:
This uses OpenAI API access and billing, which are separate from a ChatGPT subscription. See OpenAI API setup.
Set an API key for this terminal
The key value stays out of CLI arguments and YAML.
The value lasts for the current shell session. Use your secret manager for persistent or unattended runs.
Set OPENROUTER_API_KEY in the current terminal, then run:
Select an exact model from the catalog. See OpenRouter setup.
Set an API key for this terminal
The key value stays out of CLI arguments and YAML.
The value lasts for the current shell session. Use your secret manager for persistent or unattended runs.
Start your model server and load a model first. For an unauthenticated Ollama server on its default port:
The connection depends on the server and selected model supporting the required API calls. See local model setup for LM Studio, compatibility, and manual model entry.
For Groq, Together AI, DeepSeek, Mistral, LM Studio, or a custom compatible endpoint,
run colossus provider presets to see the choices, then
colossus provider setup --local for the interactive selector. The
provider guides cover manual and existing configurations.
3. Check the connection¶
The route command should name the selected model and provider. models doctor sends
one bounded generation probe and reports "ready": true when the selected route
works. If the model catalog is unavailable but you know the exact model ID, add
--model MODEL_ID to your setup command.
For authenticated providers, the generated file contains a credential reference,
not the key itself. Setup also uses the standard full-access and plaintext-storage
defaults; review
colossus config effective and choose your access,
sandbox, and storage settings before
giving an agent real work.
What's next?¶
-
Run a model turn
Send a bounded prompt and inspect its result.
-
Provider details
Configure an existing workspace or troubleshoot a connection.