Getting Connected
Freddy is scaffolded but the brain isn't wired yet. This chapter is the shortest and the most consequential: providers, models, and thinking levels decide the cost and quality of every session that follows.
The problem
Most people run their coding agent on exactly one model — whichever their vendor gave them — at one setting, forever. That's like owning a car with one gear. Model choice is a per-task decision: the model that plans Freddy's search architecture and the model that renames a variable should not cost the same. Elyra is provider-agnostic precisely so this becomes a dial you actually turn.
Logging in
Inside the session:
/login
Two paths, pick per provider:
- Use a subscription — OAuth login for Claude Pro/Max, ChatGPT Plus/Pro, GitHub Copilot. If you already pay for one of these, you have an agent budget already.
-
Use an API key — any of the
supported providers (Anthropic, OpenAI, Google, xAI,
Mistral, Moonshot, DeepSeek, Groq, OpenRouter, and
more). Keys are stored in
~/.elyra/agent/auth.json, or picked up from environment variables if you prefer.
Choosing models
See what you have and switch anytime:
/model
A searchable selector opens — type to filter
(/model sonnet jumps straight to matches).
For Freddy's build work, a strong general coding model is
the daily driver. Switching is mid-session and
instant; the conversation carries over. Nothing is lost
when you change models — this matters more than it
sounds, and Chapter 8 exploits it for cross-vendor
review.
A lineup, not a model
Now define your rotation — the short list you'll
cycle through with Ctrl+P:
/models
For Freddy, a sensible three-model lineup:
[x] claude-sonnet-5 daily driver: features, edits
[x] gpt-5.6 second opinion, cross-vendor review
[x] claude-haiku-4-5 cheap: commit messages, quick questions
The habit this enables: before sending a prompt, ask
yourself "is this a haiku task or a sonnet task?" and tap
Ctrl+P accordingly. Renaming things, writing
a commit message, asking what a file does — cheap
model. Designing, implementing, debugging — strong
model. This single habit routinely halves a project's
model spend.
Thinking levels
The second dial: how long the model reasons before
answering. Cycle with Shift+Tab —
levels run from off through
minimal, low,
medium, high, to
xhigh. The current level is always visible
in the footer.
- low/medium — Freddy's everyday setting. Edits, small features, questions.
- high — architecture decisions, debugging something genuinely weird, Chapter 9's search feature.
- xhigh — rare. You'll know the moment when it comes.
Thinking is money and time: a high-effort answer to a trivial question buys you nothing but latency. Match the dial to the question.
The safety net you get for free
Because you logged into two providers, something quietly
important is now true: a provider outage no
longer ends your session. If a turn fails
because a provider is overloaded, rate-limited, or a
model vanishes, Elyra classifies the failure and retries
the turn on an equivalent model from your other provider
— automatically, visibly, and modality-aware. It's
on by default (providerFailover setting);
you did everything required by having two logins.
Check the whole setup before moving on:
elyra doctor preflight
✓ All default models resolve against the registry.
What you learned
- Log into two providers, not one. Options are the point — for cost, for review quality, and for the failover net.
-
/modelto switch,/models+Ctrl+Pfor your lineup. Cheap model for cheap tasks is the highest-ROI habit in agent work. -
Shift+Tabsets thinking effort. Default low-medium, reach for high deliberately. - Failover is free insurance — already on, powered by your second login.
-
elyra doctor preflightconfirms the wiring before real work starts.