Local AI Agents on Mac: Free, Offline, Open Source

How do I run a local AI on my Mac that remembers past conversations?

The tools most people reach for first, Ollama and LM Studio, are runners. They load a model and serve replies, and that is where their job ends. Each request carries only the context you hand it, so memory across sessions is something you have to bolt on yourself with a separate library. Osaurus is built the other way around: memory, tools, and identity are part of the app, not an add-on. It is a free, MIT-licensed native Mac app, one of the rare local ai agents on mac that ships the agent layer in the box rather than leaving you to assemble it.

One honest note up front, because trust is the whole point of running local: the parts of Osaurus you can verify from its public repo are the license, the engine, the platform requirement, and the API it serves. The more ambitious agent behavior, the memory that carries across sessions and the autonomous execution, is documented by the maker but not something an outside reviewer has benchmarked. This article keeps those two buckets separate so you know which is which.

A native Apple Silicon runtime, not a llama.cpp wrapper

Built on Apple’s MLX, in Swift

MLX is Apple’s own array framework for machine learning on Apple Silicon, the same lineage as its on-device model work. Osaurus is built purely in Swift on top of it, so inference uses the chip’s unified memory directly instead of going through the cross-platform llama.cpp path that Ollama and Jan use. You install it with brew install --cask osaurus or a signed .dmg from GitHub Releases, and it runs open models such as Gemma, Qwen, GPT-OSS, and Llama, kept in a local models folder. For a Mac owner, that makes it a genuine ollama alternative for mac rather than a port.

The trade-off: Apple Silicon only

Going native has a cost, and it is fair to state it plainly. Osaurus runs only on Apple Silicon and only on macOS 15.5 or later, so there is no Intel Mac, Windows, or Linux build. Ollama and Jan run on all three desktop platforms. If your machine is an M-series Mac on a recent OS, that is no loss. If you need to cover a mixed fleet, one of the cross-platform runners is the safer base.

A drop-in local server for the tools you already use

This is the part that makes it easy to try without rebuilding your setup. The server exposes an OpenAI-compatible /v1/chat/completions, an Anthropic-compatible /anthropic/v1/messages, and an Ollama-compatible /api/chat on the same local port. An editor plugin or script that already targets one of those APIs just needs its base URL changed to your Mac, and the request never leaves the machine. That is a clean way to run llm offline on macbook while keeping the tools you already know.

Tool calling and MCP

Beyond plain chat, Osaurus supports streaming tool and function calls, and it acts as both a Model Context Protocol server and client. MCP is the emerging standard for wiring a model to real tools and data sources, so a model running in Osaurus can call out to plugins and local capabilities rather than only producing text. That is the mechanism behind the agent features in the next section.

The agent layer: memory, identity, and autonomous execution

Persistent memory across sessions

The memory is described as three layers: an identity that holds who you are and how you like to work, facts it has learned, and per-session episodes it can look back on. The point of a local ai agent persistent memory is that the model does not start every conversation from zero, so it can recall an earlier decision or preference. This is the single clearest difference from Ollama, LM Studio, and Jan, which keep chat logs in their window but do not feed a durable memory back to the model on their own.

Approval-gated autonomous execution

For action rather than answers, Osaurus’s agents write a markdown checklist, work through it, and verify the result, with an approval gate before anything actually runs. The maker also describes an isolated Linux sandbox for code execution, an on-device privacy filter meant to scrub sensitive data before any optional cloud call, and a secp256k1 cryptographic identity that gives each agent a verifiable local address. Treat these as the product’s stated design.

Fully offline by default, cloud only when you ask

Offline is the default state, not a mode you switch on. There is no sign-in, no telemetry requirement, and no usage cap, which is what you want from a free open source local llm setup: your prompts, code, and files stay on the machine. Because it is MIT-licensed, you can also read the source and modify it, which is a stronger privacy guarantee than a promise in a policy.

The optional pay-as-you-go Router

Some jobs still want a bigger cloud model, and Osaurus gives you two ways to get there without breaking the local-first default. You can bring your own API keys for providers like OpenAI, Anthropic, Gemini, or OpenRouter, or you can use the Osaurus Router, a hosted path billed pay-as-you-go by credits with no separate provider account. The maker says the Router stores only the usage metadata needed for billing, not your prompt or response content. The exact per-token pricing is not published in the sources here, so check it before you lean on it. Either way, the cloud is opt-in per request, never the default.

How Osaurus compares to Ollama, LM Studio, and Jan

Osaurus vs Ollama

Ollama is the sensible default and the one to beat: MIT-licensed, on every desktop platform, huge model library, and years of real-world use. It is also a stateless server built on llama.cpp, so it forgets between runs and does nothing agentic on its own. Osaurus’s pitch is the opposite trade: give up Ollama’s portability and maturity to get a Mac-native MLX engine and a memory and agent layer in the same app. If you want the proven, run-anywhere base, stay on Ollama. If you are on Apple Silicon and want the agent features without stitching them together, Osaurus is the reason to switch.

LM Studio and Jan

LM Studio is the most polished to click around in, and since July 2025 it is free for personal and commercial use. The catch for this directory is that it is closed source: free to run, but you cannot read or modify it. Jan is the open answer to that, Apache-2.0 and cross-platform, with local chat threads and assistants, but like the others it is a runner rather than an agent runtime with durable memory. Osaurus is the one that is both open source and built around the agent layer, at the price of running only on Apple Silicon.

Who it fits, and who should look elsewhere

Osaurus fits an Apple Silicon Mac owner who wants local, private inference and is drawn to the agent layer: memory that persists, tool calling, and an offline-first default with a cloud escape hatch. The honest caveats are that it is newer and less proven than Ollama, Apple-Silicon-only on a recent macOS, and that its most distinctive features are vendor-described rather than benchmarked. If you need cross-platform reach, a long track record, or a point-and-click GUI above all, Ollama, Jan, or LM Studio are the safer picks. If the pitch is a truly open, native, remembering local AI on your Mac, this is the one to try first, and it costs nothing to find out.

Frequently asked questions

Sources

1. 1

   Osaurus on GitHub (README and LICENSE)

   The maker’s own repo. Source for the MIT license ("Copyright (c) 2026 Osaurus, Inc."), the Swift + Apple MLX engine, the macOS 15.5 and Apple Silicon requirement, the OpenAI / Anthropic / Ollama endpoints on 127.0.0.1:1337, MCP and tool calling, and the documented memory, identity, and autonomous-execution features plus the Router description.

2. 2

   Osaurus press release (GlobeNewswire, July 2026)

   The company’s own announcement, cited for the roughly 7.3k GitHub stars and 185,000+ downloads as of July 20, 2026, and the #2 Product of the Day placement on Product Hunt. Vendor figures, dated here so they can be re-checked.

3. 3

   LM Studio: free for work (lmstudio.ai)

   Confirms LM Studio (Element Labs) is free for personal and commercial use as of July 8, 2025. Its app terms forbid modifying or reverse-engineering the software, which is why the comparison calls it free but closed source rather than open source.

4. 4

   Ollama on GitHub

   Ollama’s repo and MIT LICENSE. Basis for the comparison points that Ollama is free, open source, cross-platform, and built on a llama.cpp / ggml backend rather than Apple’s MLX.

5. 5

   Jan on GitHub (Menlo Research)

   Jan’s repo, licensed Apache 2.0. Source for the comparison note that Jan is free, open source, and cross-platform, with a llama.cpp engine and local chat threads rather than a built-in cross-session agent memory.

6. 6

   Apple MLX on GitHub (ml-explore)

   Apple’s MLX, "an array framework for machine learning on Apple silicon," MIT-licensed. Confirms MLX is the on-device engine Osaurus builds on, and that MLX itself is a framework rather than an end-user app.