Agentfly
Agentfly DocsSearch docs

The assistant

Connect a model

The assistant · 3 min read

Choose ready models in Free · no API key, or use Bring your own key to connect a provider or named API endpoint.

When you would want it

A key that was pasted wrong should say so immediately, not silently fail the first time you ask a question.

Where to find it

  • All Features → Connect a model
  • Connect
  • Settings

What you get out of it

  • A mistyped key fails at setup, with a clear message - not silently on your first question.
  • Bring-your-own means the model bill comes from your provider at their price, with no markup.
  • Keys live in your operating system's password store, never in a plaintext settings file.

How to use it, step by step

  1. Open Connect a model from the Assistant or command palette - or choose Connect on the banner that says Agentfly works directly with your codebase.
  2. The Free · no API key tab lists included models and ready local models. Choose Use this model to connect immediately. Included models need a verified Agentfly account and share its daily allowance.
  3. Switch to Bring your own key for your personal provider connections. Administrators control both model catalogs, including new models, hidden models, recommendations and defaults.
  4. Follow the three steps at the top: choose a provider, check and save its key, pick a model. The steps light up and the progress bar fills as you complete them. Add a named connection with its OpenAI-compatible base URL and exact model ID for anything else. Use the NVIDIA NIM preset for https://integrate.api.nvidia.com/v1.
  5. Choose a requests-per-minute limit that fits your provider account and save the connection. All agents using that connection share the same queue.
  6. Paste your API key and choose Check and save. Keys are encrypted using this computer's operating-system credential store.
  7. Choose the model for chat - the first one listed is the recommended default - or assign the connection to a named agent. Waiting and cooldown status explain when a request is queued.
  8. Set thinking effort if you want reasoning models to think harder than their default: Auto lets the provider decide, Max costs the most. For a custom address, follow the three numbered hints and choose Save endpoint.
  9. If the provider still returns 429, allow its cooldown to finish. RPM pacing cannot override token quotas, account limits, or requests made outside Agentfly.
  10. For local file tasks, choose an Ollama model that supports tools. Agentfly recovers complete tool calls returned as plain JSON or tool_call envelopes and runs them through the usual file controls. Examples and incomplete replies stay text. If a model keeps inventing tools or returning invalid arguments, the turn stops after three identical failed calls; try a fresh chat or another tool-capable model.

Why this beats the usual way

Subscription AI editors lock you to one vendor's models and resell access at a margin. Agentfly takes no cut: connect Anthropic, OpenAI, Google, a local Ollama, or any gateway - and switch between them per question if you like.

Find it in Universal Search

Connect a model · The assistant · api key · provider · model · free models · BYOK · bring your own key · tabs · local ai · NVIDIA NIM · custom endpoint · connections · requests per minute · RPM · rate limit · 429