DEV Community

Yuji Ooya for i-plusplus

Posted on Edited on AI-assisted

Local AI Fabric 0.0.1: One Plugin, Two Clients

If you have a model running on your own machine, you don't have to open another chat window to use it. Local AI Fabric (LAF) connects ChatGPT and Codex to a local inference engine you configure. For example, you can send a short passage to a suitable local model for summarization, check the job, and read the result without leaving your client. ChatGPT and Codex remain the clients; the configured model handles the inference request.

In my first LAF post, I wrote about the work that starts after a local model answers its first prompt. For 0.0.1 I picked a smaller question: if I change one tool, can both clients pick up that change without rebuilding the connection?

If you only have a minute

  • Fresh ChatGPT and Codex sessions picked up the new tool description through the same private plugin entry. An already-open Codex session still showed the old version. If a tool refresh looks like it failed, try a fresh session first.
  • 0.0.1 is a source preview. You can inspect it with Git using the commands below. There is no public plugin installer for this version.

When you need a tunnel: If you want ChatGPT on the web to reach an LAF MCP server on your own machine without making that server public, set up your own private tunnel using OpenAI’s Secure MCP Tunnel guide. Codex CLI running on the same machine can use local stdio and does not need a tunnel. The private connection described here is not shared with readers or bundled with the source preview.

The first update was one sentence

I changed the Router status tool's description to make clear that it checks readiness without loading a model. Then I refreshed the tool definitions on the existing plugin entry. Both fresh clients showed the new description and the Router returned ready. That's a useful check when I'm troubleshooting the connection: I can verify what the clients see without starting inference. This check didn't cover a first-time install.

Inspect the source

From a terminal with Git and an unused destination directory:

git clone --branch 0.0.1 --single-branch https://github.com/i-pp/laf.git
git -C laf rev-parse HEAD
git -C laf show HEAD:README.md
Enter fullscreen mode Exit fullscreen mode

The commit should start with 7e413d9. These commands inspect the source; they don't install a plugin, start a service, or download model weights. The README links to host setup and model management. To run LAF, you'll need to supply a compatible backend and model separately.

Next I want to make first-time setup as easy to follow as that status check. If the setup guide leaves you wondering what to do next, open an issue and tell me where you got stuck.

Project: Local AI Fabric on Project-KISS · Source tag 0.0.1

Prepared with AI assistance. Editorial Policy v2.0.

Top comments (0)