For local-model users
Conduit doesn't run, download, or manage models — it's a desktop app that connects to the local server you already have running: Ollama or LM Studio directly, or any OpenAI-compatible endpoint such as a self-hosted vLLM or SGLang server. No key, no bill, and the rest of the app — documents, MCP tools, artifacts, skills, memory — works exactly the same as it does with a cloud model.
The problem
You already have a model running — pulled into Ollama, loaded in LM Studio, or served off your own vLLM box. What's missing isn't a model. It's everything around it: a place to search old conversations, attach a file, call a tool, or reach a cloud model without switching to a different window entirely.
What you get
Conduit doesn't touch your model files. It's the client, and your local server stays the thing doing inference.
Conduit doesn't download or manage models — it connects to the server you already have running, so there's no credential to paste and no usage to meter. Local models price at $0 in the cost chart, because they cost you nothing.
Documents with Ollama embeddings, MCP tools, artifacts, skills, saved prompts, full-text search, apps, and workflows — all work the same whether the model behind them is on your machine or in the cloud.
When a question needs a frontier model, switch providers without leaving the conversation. Stay on the local provider and that option is simply never used.
How it works
Three steps. There's no model to pull through Conduit itself at any point.
Launch Ollama, LM Studio, or your own vLLM or SGLang server the way you already do, with whatever model you want loaded or pulled.
In Settings → Providers & Keys, pick Ollama or LM Studio directly, or choose the OpenAI-compatible provider and paste your server's base URL. Neither local provider needs a key.
Choose it from the list Conduit fetches from your server. Documents, MCP tools, artifacts, and everything else in the app work exactly as they do with a hosted model.
Pick your runtime
Exact screens, and any pre-filled defaults, are in the documentation — this is what each one needs from you.
| Runtime | What you paste | Credential |
|---|---|---|
| Ollama | Its base URL — editable if you changed the port or run it elsewhere | None — local |
| LM Studio | Its local server's base URL | None — local |
| Any OpenAI-compatible server — vLLM, SGLang, a proxy such as LiteLLM | That endpoint's base URL | Optional, only if the endpoint requires one |
All three are editable-base-URL providers, so pointing Conduit at a non-default port or another machine on your network is a settings change, not a rebuild. Conduit vs Ollama's own app goes deeper on what a separate client adds on top of Ollama specifically.
Limits
The full network breakdown, for every feature, lives on the privacy page.
No — Conduit never runs, downloads, or manages model weights itself. It's a desktop chat app that connects to a local server you already have running — Ollama or LM Studio directly, or any OpenAI-compatible endpoint such as a self-hosted vLLM or SGLang box. Pulling, deleting, and upgrading models stays with that server, not with Conduit.
It's a client that sits on top of Ollama rather than a GUI built by the Ollama project itself. Ollama has to stay installed and running — Conduit calls its API to list the models you've pulled and stream chat against them, with no separate key or account needed. See Conduit vs Ollama's own app for where each one is the better daily driver.
Yes. Pick LM Studio from the provider list in Settings, the same way as Ollama — no credential needed, and Conduit fetches whichever models it's serving. Point the base URL field at LM Studio's local server address if you're not using the default.
Yes, through the OpenAI-compatible provider — the seventeenth provider Conduit ships, built for exactly this. Point it at your server's base URL and add a key only if your setup requires one; the same route covers any self-hosted proxy that speaks the OpenAI chat completions API.
With a local provider selected and update checks left on their default manual setting, yes — the model runs on your machine, and local-only mode additionally switches off web search and cloud document indexing, so nothing else in the app reaches out on its own. Turning on background update checks, or picking a cloud provider, is what changes that.
Tool use and vision depend on the model, not on Conduit — MCP and built-in tools work with any model whose adapter supports tool calls, and image attachments go to models that accept them; a text-only model such as DeepSeek gets an attached image dropped with a stated reason rather than a faked answer. Image generation itself only runs on OpenAI, Gemini, or OpenRouter today, never on a local model.
Related
Your own API keys for every hosted provider, for the days a local model isn't quite enough.
Bring your own key For developersLocal and remote servers, OAuth, and consent graded by what a tool can do — for a local model with something to do.
Conduit as your MCP client For research and analysisIndex a collection with Ollama embeddings and retrieval stays as local as the chat model does.
Chat with your documentsFree and open source. No account, no key needed for Ollama or LM Studio.
A release candidate: installers aren't OS code-signed yet, so expect a SmartScreen or Gatekeeper prompt on first launch.