Skip to main content
The model proposes work and writes answers; the harness and engine own execution and checks.

​
Definition

The model layer is Fella’s replaceable connection to a chat model. Changing it does not change workspace ingestion, the built-in tool implementations, or deterministic verification. The model receives prompts, registered tool schemas, and selected tool results. It does not receive a direct filesystem, database, or Tauri API.

​
Architecture

Fella currently ships these provider registry entries: Fella is BYOK-only and starts the connection flow on Ollama Cloud. Ollama Cloud uses Ollama’s API shape at a hosted endpoint; it is not an OpenAI-compatible provider. The custom option expects an OpenAI-compatible API root. Some registry rows describe embedding endpoints, but Fella does not currently use embeddings for document retrieval.

​
Flow

Use /login to view and authenticate supported providers. Use /model to inspect or change the active provider, base URL, and model; where supported, Fella fetches the service’s current model list.
For a custom endpoint:
Each tab can select a model. Provider authentication is stored at the application level rather than inside a conversation.

​
Behavior

Ollama connections use /api/chat; OpenAI-compatible connections use /chat/completions. Both stream output into the same harness, which assembles tool calls, executes them through the registry, and returns results in later turns. Transient provider failures retry with backoff. Provider API keys are kept in auth.json under Fella’s application data directory, with mode 0600 on platforms that support it. Secret-bearing configuration commands are redacted from conversation transcripts. The inert /mcp command stores no connector tokens and makes no network request. Document search remains local and direct through grep_files and read_file, so it works even when the provider has no embeddings endpoint. The wire choice is independent of which hosted service you choose:
That is why Ollama Cloud uses the Ollama wire, while OpenAI, Vercel AI Gateway, xAI, OpenRouter, and custom endpoints use the OpenAI-compatible wire. The streamed response is normalized before it reaches agent::run; provider-specific framing does not leak into the tool loop.

​
Limits

  • With a hosted or custom provider, prompts and tool results sent for the conversation leave the machine and are subject to that provider’s policies.
  • The inert /mcp command does not create a connector or make a separate network request.
  • run_python is local execution inside the embedded WASM guest. The guest has no filesystem, network, environment, or subprocess capability; the prompt describes the boundary, while the host ABI enforces it.
  • Provider model catalogs and default model IDs can change. /model is the source of the active service’s available models when listing is supported.
  • The model can make reasoning and tool-selection mistakes regardless of provider. Deterministic checks cover evidence consistency, not every semantic error.

​
Next steps

The harness

See how provider output enters the tool loop.

Privacy and security

Review provider, credential, and local-execution boundaries.

Using Fella

Configure a workspace and ask a first question.