Models and routes
Where the NALA agent gets its inference: your own API key, a local model, or NALA-provided supply.
The NALA agent needs a model. Where that model comes from is the route, and it determines behaviour, cost, and who holds the credential.
The three sources
| Source | Available | Credential | Cost |
|---|---|---|---|
| Your own API key | Now | Yours, on your machine | Your provider bills you |
| A local model | Now | None | None |
| NALA-provided | Planned | NALA-issued token | A NALA plan |
PREVIEW
NALA-provided inference is not available yet. Today the agent runs on your own API key or a local model. This page documents all three so the shape is clear; the third is marked planned and should not be relied on.
Your own API key
You supply a key; the agent uses it.
SECURITY
Requests using your key go direct from your machine to that provider. They do not pass through NALA servers. Your key is stored in the OS credential store, never in a config file, never in diagnostics, and never transmitted to us.
This is the fully self-sufficient path. Nothing about it depends on a NALA account.
A local model
No credential, no network, no cost. Capability depends entirely on the model and your hardware.
Useful when the repository must not leave the machine — see privacy.
Your existing provider subscriptions
This is the question people ask first, so it is worth being exact.
Your Claude, ChatGPT, Cursor, or other subscription is usable in NALA — by launching that vendor's own agent, which authenticates itself and runs as its own process. That is free, supported, and what NALA has always done.
IMPORTANT
Those subscriptions cannot fuel the NALA agent. NALA never holds, extracts, or reuses a provider credential, and several providers explicitly prohibit reusing consumer-plan credentials in third-party tools. See account and authentication.
So: subscriptions run their own agents; the NALA agent runs on your key, a local model, or NALA-provided supply.
Switching route
/provider
/model
/profile
Route can change mid-session. That is deliberate and consequential — later turns in the same transcript can behave differently from earlier ones.
NOTE
When output quality changes unexpectedly, check whether the route changed before concluding the model got worse. Two turns in one transcript are not necessarily comparable.
A route is more than a model
| Element | Meaning |
|---|---|
| Harness | The agent program being driven |
| Provider | The vendor behind it |
| Account | Which identity with that provider |
| Model | The specific model |
| Effort | Reasoning depth, where exposed |
| Tool set | What the agent may call |
| Platform | Where the process runs |
| Billing class | Whose money is spent |
See providers, models, and routes for the full model.
What NALA cannot tell you
Regardless of route, NALA cannot report your provider's plan tier, remaining quota, or rate-limit state. Those are not exposed to it, and it does not guess.
An error naming a model or a limit is answered by the provider, not by NALA.