Generated-Computers/Locals

What it solves

Running AI models on a Mac for coding agents (like OpenCode, Cline, or Aider) is usually slow and awkward. These agents need to make many tool calls, keep long conversations, and often run in parallel — but typical local model servers are built for simple chat, not for this workload. Locals is a local server designed specifically for agents, making them run up to three times faster on Apple silicon while keeping everything on your own machine.

How it works

Locals runs AI models directly on your Mac's GPU. Instead of treating each request one at a time, it batches requests from multiple agents together, using GPU kernels tuned for a few simultaneous requests. This lets eight agents finish in roughly the time two would take one after another. It also applies a grammar directly on the GPU when generating replies, so tool calls and JSON outputs always match the required schema — no broken or malformed calls. Conversations and system prompts are cached and reused across turns, so long sessions start almost instantly (a 138,000-token conversation's next turn starts in 2.4 seconds). It also adds safety guards: detecting repetitive loops, preventing out-of-memory crashes, and freeing the server if an agent hangs.

Who it's for

Developers who use AI-powered coding assistants on a Mac and want them to be faster and more reliable. It works with tools like OpenCode, Cline, Aider, OpenHands, Claude Code, and others, and with any app that speaks the OpenAI, Anthropic, or Ollama API. You need a Mac with Apple silicon (M1 or later) and macOS 14+. It's especially useful when running many agent instances at once or working with long prompts.

Highlights

• Speed with parallel agents: eight concurrent requests get 3.1x higher token throughput than mlx-lm (271 vs 87 tokens/s on an M2 Pro).
• Always-valid tool calls: replies are constrained by a GPU-side grammar, so fields and arguments always match the schema.
• Fast resumption: prompts and conversations are cached on disk, so an agent's first request after a restart takes 1.7s instead of 16s.
• Single-command setup: locals code opens OpenCode in your project; locals launch connects to many other agent tools.
• Team mode: locals run splits a task across multiple agents, each in its own copy of your project, and applies changes only when your tests pass.
• Works with existing tools: can replace Ollama, LM Studio, mlx-lm, or llama.cpp servers on their usual ports, so existing clients keep working unchanged.