> Markdown twin of https://conduitllm.com/ > A free, open-source desktop AI app for research, writing, slides, apps and workflows, on your own API keys or local models. Steer every run, see every step. --- [New Writing, and costs for every current model](https://conduitllm.com/changelog.html#v1-rc-7) # The open-source desktop AI app for research, writing, slides, apps and workflows. Use your own API keys or local models, and ask across your own files. Steer runs mid-answer, approve each tool before it acts, see every step and source, and track what you spend each day. Get Conduit [View the source](https://github.com/runningpixels/conduit) - Free and open source (AGPL-3.0) - No account, no subscription - Windows, macOS, and Linux A release candidate: installers aren't OS code-signed yet, so expect a SmartScreen or Gatekeeper prompt on first launch. Details Key held in your OS keychainThe interface never sees it Every fact cited to your docsFiles parsed on your machine Saved on your diskLocal SQLite — no Conduit server What it does ## What you can do with it Real screenshots of the real app, following one made-up project: Ferry, an open-source sync app built with Rust and Tauri. Everything shown here ships in the current build. ### Research Every claim backed by a real quote Ask a question and Conduit proposes a brief — the sub-questions, the scope, how deep to go — for you to edit and approve before anything runs. Then it searches, reads the full pages, and writes a report with numbered sources. Each claim in the report comes with the exact words it is based on, quoted from its source. Conduit checks those words are really on the cited page, and leaves out any claim whose quote it cannot find, so a source never gets credit for something it does not say. The report notes how many pages it read and how many it cited. - Quick, Standard, or Deep: from 8 searches to 50 - Reads web pages and PDFs, up to 20 MB each - Stop at any point and keep what it found ### Writing Better blog posts, docs and reports Writing is built for long pieces: blog posts, technical docs, reports and newsletters. You agree the outline first — the sections, what each must say, how long — so the structure is right before a word is written. Then the draft fills in section by section, with links to the pages, documents and Research reports behind its facts. Revise it the way you would with an editor: ask for a change and only that part moves, or select a paragraph to make it shorter or clearer. Your own sentences stay word for word, and History keeps every version. - Cites the pages, documents, and Research reports it used - Marks a missing fact as a TODO instead of inventing one - Export Markdown or print-ready HTML, with the sources at the end ### Slides A deck you edit by asking, not a file you regenerate Describe the story and the assistant drafts a storyline for you to approve, then builds one live deck. Ask for a change and it edits that slide — or swaps a word across the whole deck — instead of producing another copy of the file. History keeps a version for every change, named after your request. Script shows every word of the deck as one document you can type in, and text you write stays yours when the assistant reworks a slide. Present full screen with a private presenter window for notes, the next slide and a timer. - Two built-in themes, plus any theme the assistant designs for you - Export a single HTML file that presents itself in any browser, or a PDF on Windows - Ask for a deck in any chat and it opens in Slides ### Apps Keep what it builds, and open it any time A page the assistant builds doesn't have to scroll away with its chat. Save it as a personal app and open it from Apps on the rail whenever you need it — it keeps working after you delete the chat. Eight starter apps come built in, from a Pomodoro timer to a budget tracker and a live weather dashboard. Each power a personal app has is declared by the page and checked by Conduit: the sites it may reach, a little storage of its own, a few settings you fill in, and — only if you allow it — a question to your AI model, answered with text and nothing else. - Save a page from its ⋯ menu; site permissions carry over only if you tick them - Data kept between launches, up to 5 MB per app, with Clear data one click away - Model access asks first, names the provider, and sends no chats, memory, or tools ### Workflows Routines that run on a schedule, and ask first Workflows fetch pages, search the web, have the model summarise, and save the result as a document. Start from a ready-made one — a morning briefing, a page summary, a topic watch — or build your own step by step, then run it now or on a schedule. Before a schedule turns on, Conduit lists everything the workflow will be allowed to do on its own for you to approve, and a run that needs more pauses and asks. ### Connectors Plug in any MCP server, on your own terms Conduit speaks the Model Context Protocol, over local stdio and remote streamable HTTP. Install a server from the official MCP registry in one click, and the assistant can use its tools in any chat, next to web search, your documents and everything else. Tools marked read-only run without interrupting you. Anything else stops and asks first, showing the exact data it will send, and you can approve it once, for the chat, or for good. - OAuth for remote servers, with tokens kept in your OS keychain - Server prompts and resources, right in the composer - Remembered approvals you can review and revoke [Conduit as your MCP client](https://conduitllm.com/use-cases/mcp-client.html) ### Steering An agent you can steer, not just watch Every run shows its work as it happens — thinking, each tool call, and the answer, in order, with the time it has been running. Queue a follow-up while the model works, or interrupt it and redirect it. When it needs something from you, it puts a real form in front of you instead of guessing. Afterwards the reply folds its steps into one line — "2 steps · 1 site" — that opens the inspector: every tool call and site contacted, and every source the answer used. Leave a chat that's waiting on you and the chat list says "Needs you." - Approve a tool for this chat or for good; sensitive ones always ask - `SKILL.md` skills per conversation — their scripts are never run - Memory stays pending until you choose to save it ### White-label Make it your own Put your name, logo and colours on Conduit and ship it as your product's assistant, or your team's. One `brand.md` file holds the whole brand, so it is a configuration, not a fork, and it updates with Conduit. Because it is local-first and runs on your users' own keys, their data never touches your servers. - Rebrand any install today, in Settings - Light and dark palettes, checked for readable contrast - Branded installers built from source, in early access [What ships today](https://conduitllm.com/whitelabel.html) [Every feature, with its limits](https://conduitllm.com/features.html) ## Try it on your own keys Free and open source, on Windows, macOS and Linux. Bring a key, or point it at a model you already run. Get Conduit [View the source](https://github.com/runningpixels/conduit) Models ## The models you already pay for Seventeen providers built in. Paste a key and pick from the list the provider returns, or skip the key and [run a model on your own machine](https://conduitllm.com/use-cases/local-llm.html). ### Model labs Your own key, at the lab's own price. ### Gateways and fast inference One key for hundreds of models, or the quickest hosts for open-weight ones. ### On your machine No key and no cost. With local-only mode on, no internet either. MiniMax, and whatever shipped this week, comes through OpenRouter, OpenCode Zen, or any OpenAI-compatible endpoint. Cost ## Know what you spend Requests go from your machine straight to the provider you picked, at that provider's own price. Conduit adds nothing on top and resells no tokens. Settings charts your estimated cost per day and per model. Prices come from your own overrides first, then the provider's live listing, then a bundled price table, and a model with no known price says so instead of showing $0. - Your provider's per-token price, with no markup - Set any model's price yourself - Local models through Ollama or LM Studio cost nothing [Will pay-as-you-go beat a $20 plan for you? Try the calculator](https://conduitllm.com/api-vs-subscription.html) Get Conduit ## Download Conduit v1.0.0-rc.7 is the current tagged build — installers for Windows, macOS (Apple silicon and Intel), and Linux. It is a release candidate rather than a 1.0: expect rough edges, and read the note on installers below before you run it. ### Windows 10 / 11 · x86_64 [Installer (.exe)](https://github.com/runningpixels/conduit/releases/download/v1.0.0-rc.7/Conduit_1.0.0-rc.7_x64-setup.exe) ### macOS 11+ [Apple silicon (.dmg)](https://github.com/runningpixels/conduit/releases/download/v1.0.0-rc.7/Conduit_1.0.0-rc.7_aarch64.dmg) [Intel (.dmg)](https://github.com/runningpixels/conduit/releases/download/v1.0.0-rc.7/Conduit_1.0.0-rc.7_x64.dmg) ### Linux x86_64 [AppImage](https://github.com/runningpixels/conduit/releases/download/v1.0.0-rc.7/Conduit_1.0.0-rc.7_amd64.AppImage) [Debian (.deb)](https://github.com/runningpixels/conduit/releases/download/v1.0.0-rc.7/Conduit_1.0.0-rc.7_amd64.deb) **A note on installers.** These builds are not code-signed by Apple or Microsoft, so the first launch shows a Gatekeeper or SmartScreen warning. On macOS, right-click the app and choose Open. The update payloads are signature-verified regardless — that part does not depend on OS signing. [All releases and signatures](https://github.com/runningpixels/conduit/releases) [What changed in each release](https://conduitllm.com/changelog.html) ### Or build from source terminal ``` # clone, install, run git clone REPO_URL conduit cd conduit pnpm install pnpm dev ``` You'll need Node 20, pnpm 10.34.4, and a stable Rust toolchain — plus Visual Studio 2022 Build Tools on Windows. Full instructions, including how to add your first provider, are in the [documentation](https://conduitllm.com/docs.html). More about Conduit: [use cases](https://conduitllm.com/use-cases.html) · [how it compares](https://conduitllm.com/compare.html) · [every network flow](https://conduitllm.com/privacy.html) · [white-label](https://conduitllm.com/whitelabel.html) FAQ ## Questions, answered plainly **Is Conduit free?** Yes, and it's open source. You pay your model provider directly — Conduit never proxies your requests or resells tokens, so there is nothing to subscribe to. **Do I need an API key?** For hosted providers, yes — you bring your own. If you'd rather not use one, run Ollama or LM Studio locally and Conduit will talk to them with no key and no cost. **What will it cost me to run?** Whatever your provider charges, and nothing on top — Conduit takes no cut, adds no subscription, and sends you no bill at all. What you actually spend depends on which model you pick and how hard you use it, which is why the app charts your cost and tokens per day from your own usage rather than leaving you to find out on a statement. Running Ollama or LM Studio locally costs nothing. **Which models does Conduit support?** The ones people actually use — GPT, Claude, Gemini, Grok, GLM, DeepSeek, Kimi, Qwen, MiniMax — and anything else your provider lists. Seventeen providers are built in, including dedicated ones for xAI, Z.ai, Moonshot AI, and Qwen. MiniMax, and whatever shipped this week, comes through OpenRouter, OpenCode Zen, or any OpenAI-compatible endpoint. A model you already run in Ollama or LM Studio needs no key at all. **Where does my data go?** Your conversations, attachments, documents, and artifacts stay in a local SQLite database and blob store on your machine. What leaves it is what a feature needs, sent from your machine straight to the service you picked — your messages to your chat provider, a document's text to the embedding provider when you add it to a collection, and a search query when you turn web search on. The [privacy page](https://conduitllm.com/privacy.html) lists every one of those flows and its default. **Can I chat with my own documents?** Yes. Add text, Markdown, CSV, Word, or PDF files to a collection under Documents, attach it to a chat, and the reply names the files it drew on — click one to read the exact passage. Files are parsed on your machine and the index lives in your local database. To be searchable, a document's text is sent once to the embedding provider you choose (OpenAI, Gemini, or OpenRouter, after you agree to it); pick Ollama and it never leaves your machine. Scanned PDFs without a text layer can't be read. [How it works, step by step](https://conduitllm.com/use-cases/chat-with-documents.html). **Does it phone home?** No. The app has no background telemetry, no analytics, and no usage reporting. Update checks are manual by default — nothing contacts the update server until you press Check now, and background checking is something you switch on yourself. Update payloads are signature-verified before anything is applied. (This website is separate and does use Google Analytics — see the [privacy page](https://conduitllm.com/privacy.html).) **Will my system warn me when I install it?** Yes. The binaries aren't OS code-signed yet, so Windows SmartScreen and macOS Gatekeeper flag them on first launch. Building from source avoids this entirely. Update payloads are separately signed and verified, so the update path is protected even while the installer isn't. **Can I run a completely local model?** Yes. Ollama and LM Studio are supported directly, as is any OpenAI-compatible server you host yourself. Combined with local-only mode, and with update checks switched off, Conduit will make no internet requests at all. [Running models locally](https://conduitllm.com/use-cases/local-llm.html) covers the setup. **Is there an open-source Claude Desktop alternative?** Conduit is one. It is AGPL-3.0 licensed, keeps conversations on your own machine, and talks to seventeen providers rather than one — but Claude Desktop is more mature and its installers are code-signed, which ours are not yet. [The full comparison](https://conduitllm.com/compare/claude-desktop.html) lists the rows each one wins. **What MCP clients run on Windows?** Several, including Claude Desktop, Cursor, and Conduit. Conduit runs on Windows 10 and 11 with no feature differences from macOS or Linux, and speaks MCP over both local stdio and remote streamable HTTP with OAuth. [More on Conduit as an MCP client](https://conduitllm.com/use-cases/mcp-client.html). **Can I white-label an AI chat client?** Partly, today. You can change the name, logo, and colours of a stock Conduit install from Settings → Branding; producing your own signed installer with your connectors pre-wired is not finished yet. [The white-label page](https://conduitllm.com/whitelabel.html) has the capability table and says plainly what is missing. **How is this different from a browser-based front end?** Conduit is a native desktop application, not a web app you host. That's what makes the OS keychain, a local database, supervised local tool processes, and true offline use possible — none of which a browser tab can do.