The archive · Developer & Business Tools · Product decision · 2026
apfel bets Apple's bundled Mac LLM is the free local AI: HN 743 pts, 6.3k stars
A zero-config CLI and OpenAI-compatible server for the LLM Apple already ships in macOS; HN front page in April 2026, ~6.3k GitHub stars by September.
apfel (Arthur-Ficial)
What the business is
apfel is a free, MIT-licensed macOS CLI plus OpenAI-compatible local server and chat that exposes Apple's on-device FoundationModels LLM through one brew install, with MCP tool support and zero network calls.
How it started
Solo developer Arthur-Ficial started apfel after trying to run agents like openclaw against Apple's on-device model in ultra-token-saving mode; that failed, but shell scripting worked, so he built a wrapper exposing the model through stdin/stdout, proper exit codes and JSON output.
What happened
The Show HN on 2026-04-03 drew 157 comments debating the 4K context window, Apple's guardrails, and whether the free model was being slept on; the project grew through versions with MCP tool support, PDF/image extraction, an OpenAI-compatible server, and a Homebrew formula, reaching 6,338 GitHub stars by September 2026.
How it ended up
Still live and free: v1.9.1 with CLI, server and chat modes, Homebrew install, MIT license; no pricing, subscriptions or funding announced — a one-developer open-source project.
Background
apfel is a free, MIT-licensed tool that gives developers a zero-config command line and OpenAI-compatible local server for the ~3B-parameter LLM Apple already bundles in macOS 26 Tahoe via the FoundationModels framework. One brew install exposes the model as a UNIX tool, a server at localhost:11434, and an interactive chat.
The founding bet is that the winning local-AI move is a front door, not another model download: no weights to fetch, no API keys, no cloud, with every token generated on the user's Apple Silicon. The project adds what Apple's raw API lacks — exit codes, JSON output, file attachments, MCP tool support and context-trimming strategies for the 4,096-token window.
The Show HN on 2026-04-03 drew 743 points and 157 comments on the front page, with the thread debating the context window, Apple's guardrails, and whether the bundled model had been slept on. By September 2026 the project was at v1.9.1 with 6,338 GitHub stars per the product site — still a one-developer, free, unfunded open-source project.
What has to be true
- Apple's on-device model is already installed on millions of Macs, so apfel's distribution cost is zero — no download funnel to build.
- 100% on-device inference with no network calls answered the privacy objection that blocks cloud AI in shell and automation workflows.
- OpenAI-compatible endpoints let existing SDKs and agents switch to the local model by changing one base URL.
- The Unix-tool framing (stdin, JSON, exit codes) made the model composable with shell scripts, a niche the Show HN audience immediately recognized.
What can be applied
Find the asset someone already paid for: Apple ships an on-device LLM in every Mac, so the whole opportunity was the missing front door — distribution and trust came free.
Aftermath
As of September 2, 2026 apfel is live and free at v1.9.1: a CLI, an OpenAI-compatible server, and a chat mode, installable via Homebrew with an MIT license, plus demo scripts and an ApfelCore Swift package. The product site reports 6,338 GitHub stars as of September 1, 2026. There is no pricing, subscription or funding announced, so the project remains a one-developer open-source tool whose longevity depends on Apple keeping FoundationModels accessible — the same platform dependency that gives it zero-cost distribution.
Sources
- Show HN: Apfel – The free AI already on your Mac
- Arthur-Ficial/apfel — The free AI already on your Mac
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card