Local AI with Ollama

Most modern productivity applications treat artificial intelligence as a cloud microservice. When you ask software to extract action items from a transcript or synthesize a project update, your sensitive thoughts, internal metrics, and meeting dialogues are packaged into an API payload and dispatched across the public internet to third-party data centers.

For executive briefings, client consultations, legal strategy sessions, and proprietary product planning, that architectural pattern is an unacceptable security compromise.

Oyma takes an entirely different architectural path. By integrating directly with Ollama on macOS, Oyma delivers private, powerful AI summarization and note analysis that runs 100% locally on your own computer hardware.

The architectural shift to on-device intelligence

On-device machine learning is no longer a toy or a compromised fallback. Modern Apple Silicon architectures (M1, M2, M3, and M4) combine high-bandwidth unified memory with dedicated Neural Engines and high-performance GPU cores. This hardware foundation allows Mac computers to execute 7-billion and 8-billion parameter language models at conversational speeds.

Instead of paying a $20 to $30 monthly per-seat fee for cloud-hosted AI notetakers that store your transcripts on remote multi-tenant servers, Oyma pairs with local open-source models to run synthesis right inside your personal workspace.

When a meeting concludes or when you request a synthesis of your daily notes, the inference pipeline executes locally:

1. Local context assembly: Oyma gathers the active note or meeting transcript from your local disk vault.
2. Local HTTP dispatch: The prompt and context payload are transmitted across your Mac loopback interface (http://127.0.0.1:11434) directly to the Ollama runtime.
3. Hardware-accelerated inference: The model processes the tokens utilizing Metal acceleration across your Mac unified memory.
4. Structured Markdown insertion: The resulting action items, executive summaries, and key decisions are streamed back and written directly into your plain text file.

At no point in this sequence does a single packet leave your machine. You can verify this behavior anytime using macOS Network Utility or by reviewing Oyma‘s built-in privacy architecture.

Why local AI belongs inside your Markdown vault

When your notes are stored as plain Markdown files on local disk, routing their contents through proprietary cloud AI services breaks the fundamental promise of data sovereignty. Local AI ensures that your analytical layer respects the same privacy guarantees as your storage layer.

1. Absolute confidentiality for sensitive discussions

Whether you are negotiating vendor contracts, conducting one-on-one team performance evaluations, or reviewing patient histories, confidentiality is non-negotiable. With local Ollama integration, zero third-party telemetry, training data harvesting, or retention policies apply. Your meetings remain as private as an unshared file on your desktop.

2. Complete resilience without an internet connection

Cloud-based AI assistants cease functioning the moment your Wi-Fi disconnects or when you work from an airplane. Because Ollama runs locally on your Mac, Oyma continues to generate structured summaries, extract action owners, and parse discussion notes whether you are at 35,000 feet, working in a remote cabin, or operating behind a strict corporate air-gap.

3. Freedom from vendor subscription fatigue

Cloud AI notetakers routinely impose artificial monthly usage quotas, token caps, or tiered enterprise pricing structures. Running open-source models locally means your usage is governed solely by your computer battery and processing capacity. You can process fifty hours of recordings or hundreds of archived research notes without incurring a single API overage charge.

Supported models and hardware expectations

Oyma is model-agnostic and interfaces with any model installed in your local Ollama library. We have benchmarked several open-weight models to help you select the ideal pairing for your Mac hardware configuration:

• Llama 3.1 8B (Recommended): Exceptional instruction following, sharp action item extraction, and concise executive summaries. Runs at rapid speeds on any Mac with 16 GB or more of unified memory.
• Mistral 7B: Renowned for structural precision, excellent Markdown formatting, and fast time-to-first-token. Performs smoothly even on 8 GB base configurations.
• Gemma 2 9B: Developed by Google, offering high reasoning capabilities and nuanced understanding of technical discussions and complex architectural debates.
• Phi-3 Mini: A lightweight 3.8-billion parameter model that delivers surprising synthesis quality while maintaining ultra-low memory footprints and minimal battery impact on MacBook Air laptops.

Switching models in Oyma requires only a single preference adjustment. As open-source research releases newer, more capable models, you can download them via Ollama and immediately utilize them in Oyma without waiting for an application update.

Structured templates built for action

Generic AI prompts often produce rambling, unstructured paragraphs that clutter your notebook. Oyma utilizes five deterministic summary templates proven in the application codebase (shared/summary-templates.ts):

• Executive summary: A concise high-level overview followed by major decisions made during the call.
• Action items with owners: Direct attribution of tasks, deliverables, and proposed deadlines extracted from conversational context.
• Key decisions log: A bulleted historical record of agreed architectural choices, budgetary commitments, or project pivots.
• Topic breakdown: Thematic headings grouping related conversational threads with supporting timestamps.
• Comprehensive meeting record: An all-in-one synthesis combining context, decisions, and follow-ups ready to be exported or shared with your team.

Every generated summary respects your existing editor typography, automatically integrating with the live-preview Markdown editor and linking seamlessly to referenced notes via wikilinks.

Bringing your own key when you need maximum context

While on-device models handle the vast majority of day-to-day meeting notes and project outlines with ease, certain multi-hour workshop recordings may exceed local context windows.

For these specific scenarios, Oyma gives you the option to bring your own API key for Anthropic (Claude 3.5 Sonnet) or OpenAI (GPT-4o). When configured, your keys are stored securely in the native macOS Keychain (main/ai/key-store.ts). Oyma never proxies requests through an intermediate company server; communications travel directly from your Mac to the chosen provider API, preserving your direct relationship with the platform.

For teams who must ensure zero external communication under all circumstances, Oyma features a strict toggle that shuts down all cloud network endpoints, ensuring only local Ollama inference can ever take place.

Combine local intelligence with our zero-bot meeting recording to transform raw spoken discussions into structured, permanent knowledge on your Mac.