Ollama Cloud vs. Local Architecture

Introduction

The ollacc project provides a specialized architecture for integrating Claude Code with Ollama. While standard Ollama deployments typically rely on a local server (localhost:11434) to proxy requests to the cloud or run models locally, ollacc implements a "Direct-to-Cloud" path. This architecture eliminates the local Ollama overhead by pointing the Anthropic SDK directly at the Ollama Cloud API.

Sources: README.md:5-15, README.md:20-25

Architectural Comparison

The primary distinction between standard Ollama usage and the ollacc approach lies in the network path and resource utilization of the local machine.

Standard Local-Routed Architecture

In a standard setup, requests from Claude Code are sent to a local Ollama server. Even when using cloud models via ollama pull <cloud-model>, the local server acts as a proxy. Every request and response consumes local CPU and RAM as it funnels through the local daemon.

This diagram illustrates the proxying behavior where the local server remains in the critical path.
Sources: README.md:12-19

Ollacc Direct-to-Cloud Architecture

The ollacc architecture removes the intermediate local server. By configuring specific environment variables, the Anthropic SDK communicates directly with https://ollama.com. This is particularly beneficial for "swarming" scenarios where multiple parallel sessions would otherwise be bottlenecked by a single local process.

The diagram shows how ollacc acts as a configuration wrapper that enables a direct connection between the client and the cloud API.
Sources: README.md:20-27, lib/auth.sh:51-55

Key Components and Logic

Configuration and Environment Management

ollacc manages the state necessary to bypass the local server through a configuration file and specific environment variables required by the Anthropic SDK.

Sources: lib/auth.sh:10-13, lib/auth.sh:51-55, README.md:104-115

Dynamic Model Discovery

Instead of relying on locally pulled manifests, ollacc dynamically discovers available cloud models by intersecting popular models with the user's authorized catalog.

1. Catalog Fetch: Retrieves the authorized tags from https://ollama.com/api/tags.
2. Popular Scrape: Scrapes https://ollama.com/search?c=cloud for current trending models.
3. Intersection: Filters the popular list to only include models present in the user's authorized catalog.

Sources: lib/catalog.sh:22-60

This sequence shows the process of populating the model selection menu without querying a local database.
Sources: lib/catalog.sh:22-55

Execution Flow

The ollacc system follows a specific startup sequence to ensure the environment is prepared before handing over control to the Claude Code process.

The flow ensures that by the time claude is executed, the local environment is indistinguishable from one pointing to a native Anthropic endpoint, despite routing to Ollama Cloud.
Sources: install.sh:65-71, lib/auth.sh:18-49, lib/menu.sh:7-35, README.md:30-36

Persistence and Security

The architecture relies on a local hidden directory ~/.ollacc/ for state persistence.

• Config File: ~/.ollacc/config stores the API token and default model. It is created with chmod 600 to restrict access to the current user.
• Permissions: The directory itself is chmod 700.
• Idempotency: The installer ensures that the ollacc alias is added to shell profiles (.zshrc, .bashrc, etc.) only once.

Sources: install.sh:41-43, install.sh:65-71, lib/config.sh:44-46

Conclusion

The ollacc architecture serves as a performance-oriented alternative to the standard Ollama local-proxy model. By leveraging the Anthropic SDK's environment-based configuration, it enables direct communication with Ollama Cloud, reducing local resource consumption and improving scalability for parallel AI agent tasks.