Documentation

Getting started

Conduit is a desktop application. There is no hosted version to sign into — you run it on your own machine, with your own provider key.

Installing

Installers are on the GitHub Releases page — the latest release is the current build for Windows, macOS, and Linux. You can also build from source, which takes about five minutes on a machine that already has the toolchain.

Prerequisites

RequirementVersionNotes
Node.js20What CI builds against
pnpm10.34.4Pinned; corepack enable is the easiest route
RuststableVia rustup, with rustfmt and clippy
Visual Studio 2022 Build Tools—Windows only, for native compilation

On Windows, the repository includes check-setup.ps1, which verifies all four and tells you what's missing.

Build and run

terminal
git clone REPO_URL conduit
cd conduit
pnpm install
pnpm dev

pnpm dev starts the Vite dev server and then the desktop shell. The first Rust build takes a while; subsequent ones are incremental.

CommandWhat it does
pnpm devRun the app in development mode
pnpm buildProduce a native installer for your platform
pnpm checkTypecheck the TypeScript workspace

Supported platforms

Binaries aren't OS code-signed yet. macOS Gatekeeper and Windows SmartScreen will warn on first launch because the bundles have no OS-level signature. Auto-update payloads are separately signed and verified. Building from source sidesteps the warning entirely.

Your first conversation

Conduit won't let you chat until it has a way to reach a model. On first launch, onboarding asks you to pick a provider and supply a credential.

Using a hosted provider

  1. Open Settings → Providers & keys (⌘,).
  2. Choose your provider and paste its API key.
  3. Pick a model from the list Conduit fetches from that provider.

The key goes into your operating system's credential store — Windows Credential Manager, macOS Keychain, or Linux Secret Service. Settings hold only a reference to it, never the value. On machines without a keychain, an encrypted file-backed store is available as an explicit opt-in.

Using a local model, with no key

Install Ollama or LM Studio, pull a model, and select that provider in settings. No credential is needed and nothing leaves your machine. Combined with local-only mode, which refuses cloud providers outright, Conduit makes no internet requests at all.

Attach images and documents with + → Attach files…, by pasting, or by dropping them on the chat; in Chats a drop anywhere on the chat attaches, while elsewhere a drop adds the file to Documents. Images go to models that accept them. A PDF, Word, Excel, text, Markdown, CSV or HTML file reaches every model as its text, extracted on your device; models that read PDFs themselves (Claude, OpenAI, Gemini, and OpenRouter models that list file input) also get the PDF, so charts and scanned pages come through. Each attachment says how it will be sent, or why it won't be.

Edit any message and resend: editing your latest message rewrites it in place, while editing an earlier one forks a new conversation from that point, keeping the original intact. Export a conversation as Markdown or JSON.

Retry keeps every answer. Retrying the last reply keeps the earlier one as a version, and a ‹ 2 / 3 › switcher under the reply moves between them; the chat continues from the one showing. An earlier reply shows its versions but cannot switch, because the messages after it were written from the one it has. The arrow next to Retry opens the model menu: the model you pick answers the last question again, for that reply only, and the chat keeps its own model. The command palette has Retry with another model… too. Compare, next to the switcher, opens the versions side by side in the panel on the right, each with the model that wrote it, its answer and its cost. Use this one picks a version for the last reply, and Fork from this version starts a new chat that continues from any version.

Continue a cut-off reply. When a reply stops at the output limit, the note under it has a Continue button. The model picks up where it stopped, in the same message, with no new message from you. It works with every provider.

The model menu opens with a search field: type to narrow every provider's list, and Enter takes the first match; each provider's models are listed alphabetically. Reasoning effort, in a chat's settings and in Settings → Chat defaults, is Auto, Low, Medium or High. Lower effort answers faster and costs less on thinking models, and Auto leaves the model's default. It is sent only to models that accept it, and is mapped for OpenAI, OpenRouter, Anthropic and Gemini.

Settings → Usage & Cost charts your estimated cost and tokens per day and per model, and counts every model call, including Research runs and workflow steps. Prices come from your own overrides first, then the price OpenRouter lists for its models, then a snapshot of the models.dev catalog bundled with each release; a model with no known price says so instead of showing $0. The same page can alert you when today's estimated spend passes an amount you set: off by default, shown at most once a day after a reply that crosses it, and it never stops anything. Days follow the usage chart, in UTC.

Supported providers

Twenty-three providers ship built in — the OpenAI-compatible endpoint is counted among them, not an addition on top. SambaNova, DeepInfra and Novita AI list their prices, so their models show what they cost.

ProviderCredentialBase URL
AnthropicAPI keyEditable
OpenAIAPI keyEditable
Google GeminiAPI keyFixed
xAIAPI keyEditable
Z.aiAPI keyEditable
Moonshot AIAPI keyEditable
QwenAPI keyEditable
MiniMaxAPI keyEditable
CerebrasAPI keyEditable
SambaNovaAPI keyEditable
DeepInfraAPI keyEditable
Novita AIAPI keyEditable
NVIDIA NIMAPI keyEditable
DeepSeekAPI keyFixed
MistralAPI keyFixed
GroqAPI keyFixed
Together AIAPI keyEditable
Fireworks AIAPI keyEditable
OpenRouterAPI keyFixed
OpenCode ZenAPI keyFixed
OllamaNone — localEditable
LM StudioNone — localEditable
Any OpenAI-compatible endpointOptionalEditable

The last row covers self-hosted gateways such as vLLM and LiteLLM, and any vendor with no dedicated adapter yet: give Conduit a base URL and, if the endpoint requires one, a key.

Adding GPT, Claude, Grok, GLM, Kimi, and the rest

Conduit talks to the model families people actually run in 2026 — GPT, Claude, Gemini, Grok, GLM, DeepSeek, Kimi, Qwen, and MiniMax — through whichever of these three paths fits the model you want.

  1. Built-in adapter. Anthropic, OpenAI, Gemini, xAI, Z.ai, Moonshot AI, Qwen, MiniMax, DeepSeek, Mistral, Groq, Together AI, Fireworks AI, Cerebras, SambaNova, DeepInfra, Novita AI, and NVIDIA NIM each have a dedicated provider: pick it, paste the key, and load the model list. That covers Claude, GPT, Gemini, Grok, GLM, Kimi, Qwen, MiniMax, and DeepSeek directly — each ships with its own default endpoint, so there's no base URL to look up.
  2. A gateway. OpenCode Zen or OpenRouter catalogue the rest behind one key, including whatever else those services are currently serving — Conduit fetches the live list rather than a fixed catalog.
  3. Any OpenAI-compatible endpoint. For a vendor with no dedicated adapter, or a self-hosted proxy such as vLLM or SGLang, point the OpenAI-compatible provider at that endpoint's base URL and supply a key if it needs one.

Qwen's built-in adapter defaults to DashScope's international endpoint, and Moonshot's defaults to api.moonshot.ai — both fields stay editable if you need a different region.

Keyboard shortcuts

Use Ctrl in place of ⌘ on Windows and Linux.

ShortcutAction
⌘KCommand palette — search, commands, models, artifacts
⌘1–⌘9Go to Home, Chats, Slides, Writing, Apps, Workflows, Inbox, Agents, Context
⌘NNew conversation
⌘,Settings
⌘/Open this shortcuts sheet
⌘\Toggle the sidebar
⌘JToggle the document panel
⌘⇧EExpand the artifact panel, or restore the layout
⌘⇧PCycle provider
⌘⇧WToggle web search for this turn
⌘⇧FFork the conversation here
⌘⇧CCopy the last assistant message
EscClose a panel, or stop a running response

The command palette is modal: type > for commands, @ for artifacts, / for models, or plain text to search every message you've ever sent. Settings has its own search box at the top of the sheet, so you can jump straight to a setting by name instead of hunting through sections.

Tools and connectors

Built-in tools

These need no configuration:

A tool asks before it runs unless it is marked read-only. Built-in read-only tools run automatically, and a connector's tool does only when its server marks it read-only; every other tool prompts you before it runs. Web search is off by default and asks for consent the first time; hosted search only runs through the official Anthropic, OpenAI, or Gemini endpoint, and falls back to the local backends above once you point a provider at a custom base URL. Queries to a local search backend go to that service; Exa, the default, may keep them under its terms.

The model's page fetch reaches public https sites only: never localhost or an address on your own network, and every redirect is checked. It returns a page's title and readable text rather than raw HTML, and reads PDFs up to 20 MB. A scanned PDF with no text counts as a page with nothing to read.

MCP connectors

Conduit runs Model Context Protocol servers as supervised connectors. Add a local process on the Connectors page with the command and arguments that start it, or point at a remote MCP endpoint that speaks streamable HTTP.

A connector's tool runs without asking only when its server marks it read-only (the standard readOnlyHint, or Conduit's own permissionLevel field). Every other tool asks first, and you can remember an approval for one chat or always. Before 1.0.0-rc.6, tools on most third-party servers ran without asking, so connectors you already use may now prompt for tools that ran silently before.

On Context → Connectors, each tool of a running connector shows what it declares: only reads, changes things, or always asks. A tool that declares nothing asks before every call; tick Only reads for one you trust to only look things up, and it runs without asking in chats, agents and workflows. A tool that says it changes things, or is sensitive, keeps asking. When a chat would offer more than 20 connector tools, or their descriptions run long, the model gets a search_tools tool instead of the whole list and loads the tools it finds for that turn; with fewer tools nothing changes, and approvals work as before. If your message names a connector, its tools are offered directly.

The runtime restarts crashed local servers with backoff, capped at three restarts in five minutes, limits concurrency, and times out individual calls at thirty seconds. Tool output is size-capped and redacted before it reaches the model, and a structural guard prevents that output from being replayed as though it were an instruction — the standard prompt-injection route through tool results.

Transports: local stdio and remote streamable HTTP ship today. Legacy HTTP+SSE is not supported.

The composer can also pull in a connector's prompts and resources, not just its tools: a prompt picker fills in the arguments a prompt declares before it lands in the composer for you to edit, and a resource picker attaches content to the next message only. Consent is asked once per connector, and a resource that tries to slip in instructions is refused and named rather than silently followed.

Connecting a server on your own machine is the docs problem. Putting that server in front of a customer — without a JSON config and without appearing inside Claude Desktop — is a different one. How to demo an MCP server to a customer compares a hosted chat, Claude Desktop plus config, and shipping a desktop client.

Naming the command on Windows

Connectors are started directly, never through a shell. That is a deliberate guard — connector configuration would otherwise be a route to arbitrary command execution — and it has two consequences on Windows that most published MCP instructions do not account for, because they are written for clients that do use a shell.

Give the executable its real filename, extension included. A bare npx fails with the system cannot find the file specified: without a shell there is nothing to try the entries in PATHEXT on your behalf. Use npx.cmd. The same applies to any tool whose Windows form is a .cmd or .bat shim; node and other real .exe programs resolve normally.

The cmd /c recipe from other clients' documentation will be refused. Configurations in the shape "command": "cmd", "args": ["/c", "npx", …] — the usual Windows advice for Claude Desktop and others — fail with refusing to spawn connector through a shell interpreter. sh, bash, powershell and pwsh are rejected the same way. Point the command at the program itself and pass its arguments in args.

The agent loop

When a model calls a tool, Conduit runs it and feeds the result back for another round within the same turn. You control the bounds under Settings → Agent: a maximum step count from 1 to 50, and a wall-clock budget from 30 seconds to 30 minutes. The chat UI shows a live progress timeline — thinking, tool calls, and text, in the order they actually happen across rounds, with elapsed time and a note when a round stalls — rather than a single spinner.

Your agents

On the Agents page you make an agent once — name, instructions, model, tools, the agents it may hand work to, and a token limit — and use it in chat, in any workflow, or on a schedule. A workflow keeps a reference to the agent, so editing the agent changes it everywhere; deleting an agent gives each workflow that used it its own copy. Four starters come with a new profile: Researcher, Writer, Reviewer and Data analyst. An agent set to Remembers keeps its own notes between runs, which you can see and edit in its editor, and Reads lets it search the knowledge collections you pick.

In a chat, type @ in the message box to pick one of your agents for that turn, or press Ask on its card. Under Settings → Agent you choose which agents the assistant may hand work to. Each hand-off runs on that agent's own model and shows as one line in the reply with its status, tokens, cost and time; Helpers under the reply totals tokens and cost by model. Several hand-offs in one go run at the same time — up to 3 by default, 1 to 4 in Settings → Agent — but helpers that change things, such as using your computer or a connector action, still work one at a time.

Click a hand-off to watch that agent: its task, tool calls, tokens and cost. Stop ends that helper only, and Redirect (or the message box) sends it a new instruction while it works. Run on a schedule… on an agent's card gives it a task to do daily, on weekdays, every few hours, or when a feed or folder has something new; each answer lands in Results, with an optional notification, and Turn into a workflow makes a routine an ordinary workflow you can edit. The document panel's ⋯ menu can send an HTML, Markdown or text document in a chat to your Reviewer agent; if it finds problems, the chat revises the same document.

Artifacts

Longer outputs — documents, scripts, small web pages — open in a panel beside the conversation rather than filling the message scroll. Conduit renders text, code with syntax highlighting, JSON as a collapsible tree, Markdown (including Mermaid diagrams and KaTeX math), and interactive HTML.

Interactive HTML runs inside a sandboxed frame that allows scripts and nothing else: no same-origin access, no popups, no top-level navigation. Forms still work — Conduit hands the page its submit event, and nothing is sent anywhere by submitting. An injected content policy sets connect-src 'none', so the page cannot fetch or open sockets of its own, and Conduit serves it from a sandboxed address of its own. The page can't call the application's own commands. One exception, a known issue on macOS: a page's script can still open a WebRTC connection, which the content policy doesn't cover; Windows and Linux block it in the webview. If you need an artifact to load images or fonts from a specific host, you can widen that yourself in the Artifact Security part of Settings → Privacy & data.

A page that needs live data — an exchange rate, the weather — can still ask for it. Its fetch() calls go to Conduit, which holds the first request to each new site and shows a banner naming it. You can allow that site once, always for that page, or not at all, or let the page reach any public site. Conduit then makes the request itself: https only, no cookies or credentials, no local or private addresses, with a 20-second timeout, a 5 MB response limit, and 120 requests a minute per page. The globe button on the page lists every site it has asked for and every request it made. Permissions belong to that one page and are deleted with it. Local-only mode refuses all of this, and Pages can connect to the internet in the same Artifact Security settings turns it off.

Paste a fenced code block into a conversation and Conduit will offer to promote it to an artifact. Artifacts export to disk, optionally with a metadata sidecar.

Each artifact holds one current payload. There is no version history — editing replaces the contents in place, so export anything you want to keep a copy of.

Documents

Create a collection under Documents in the sidebar, then add files — plain text, Markdown, CSV, HTML, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), or PDF. For an HTML file, Conduit reads the page's title and text and leaves out scripts, styles and other code. Drop a file anywhere on the window outside Chats to add it, or attach a whole collection from the composer for a single chat. A reply shows the files it drew on as clickable chips; click one to see the exact passage.

Just one file? Attach it to the message instead: + → Attach files…, paste it, or drop it on the chat. Nothing is indexed and no embedding provider is involved. The model gets the document's text, extracted on your device, and a model that reads PDFs gets the PDF itself too. Either way the content goes to the chat provider you picked, as part of your message; pick Ollama or LM Studio and it stays on your machine. A PDF over 20 MB goes as text only, and a long document's text is cut at 150,000 characters.

Each collection is embedded by one provider — OpenAI, Gemini, OpenRouter, or Ollama (Anthropic can't embed). Indexing sends a document's full text to that provider, so Conduit asks for consent the first time, remembers your answer per provider, and you can withdraw it later from the Documents page. Pick Ollama and nothing leaves your machine — local-only mode forces this choice. Retrieval itself is hybrid: vector similarity plus keyword search over the same local SQLite database, and the retrieved passages still go to whichever provider answers the chat.

Parsing happens on-device, with no OCR step, so a scanned PDF with no text layer yields nothing — Conduit tells you when that happens rather than importing an empty document. Attached to a message instead, a scanned PDF can still be read by a model that reads PDFs, since that model gets the PDF itself.

Projects

Give every chat in a project the same instructions, documents and skills. Projects in the chat list groups chats one level deep, with no nesting; a project with no settings is simply a folder. A project's ⋯ menu has New chat here, Project settings, Rename and Delete. Deleting asks first and keeps the chats: they move out of the project. Under Project settings:

A project adds to your settings rather than replacing them. Instructions stack in order, each labelled: yours from Settings, then the project's, then the chat's own. Collections and skills are the project's plus any the chat adds. In the composer, a project chip shows which project a chat is in, and the chat settings popover lists what the chat also uses from Settings and the project. A project's settings are read when you send a message, so editing a project changes all its chats, and moving a chat in or out applies from its next message. The working folder is the exception: it is copied when a chat is created, so chats already in the project keep theirs.

Image generation

OpenAI, Gemini, and OpenRouter can generate images inline in a conversation. Conduit asks once, before the first image, since each one is billed by the provider — asking about image generation doesn't itself trigger the prompt. The result is saved locally as an artifact rather than linked from the provider, so it doesn't expire.

Personal apps

A page worth keeping can be saved as a personal app: Save as app in the page's ⋯ menu makes a copy you open from Apps on the rail, or from Your apps on the new-chat screen. The copy keeps working after you delete the chat. Saving asks which of the page's site permissions to keep, one by one, and keeps none you don't tick. When the chat's page changes later, the app offers Update. Eight starter apps come built in; add one from the Apps page, or press Open app on an idea that has one.

A page (an app, or one in a chat) can declare three more things, each checked by Conduit. It can keep data between launches — up to 5 MB per page, stored with the rest of your data and encrypted when encryption at rest is on; Clear data removes it. It can take a few settings, such as a city, which Conduit asks for in a form of its own and passes to the running page without reloading it. And it can ask your AI model, after you allow it for that page: the prompt names the provider and says when text would leave your device, and the page gets text in and text out — no tools, chat history, memory, or documents.

Every personal app has its own Settings page (the gear in the app's header, or ⋯ → Settings): which model answers it, tokens per day with a daily limit for cloud models (100,000 by default; models on your computer are never limited), its saved data key by key with Export and Clear, the sites and providers it may use, and a week of one-line activity records — never prompts, replies, or saved values.

Slides

Describe the story, or ask for a deck in any chat, and Conduit drafts a storyline for you to approve, then builds one live deck: every change edits that deck instead of producing another copy. The deck opens in a studio with the slide large in the middle and the chat beside it. Ask for a change and the assistant edits one slide, or swaps a word across the whole deck.

Built from parts. The assistant doesn't write a slide's HTML: it picks a layout and fills in its parts — a headline, points, numbers with their labels, columns, a quote — and Conduit lays the slide out in the theme. Text too long for its place goes back to the assistant with what to shorten before the slide is made. A free-form layout remains for slides no layout fits. Text that still doesn't fit shrinks a step at a time, down to a readable size, instead of overlapping or running off the slide, and exported HTML and PDF match what you see.

Charts from your numbers. Bar, line and funnel charts are drawn by Conduit from the data the assistant gives: axes that start at zero, bars in proportion, labels that don't collide, sizes readable from across a room. A full-width chart layout gives a chart the whole slide. Diagrams the assistant draws itself are checked for tiny labels.

A layout check after each change. After the assistant changes slides, Conduit looks at how they render. If text runs off, overlaps, crowds the footnote, or a diagram is unreadable, the assistant gets one round to fix exactly that; anything left is shown to you as a note. Text you wrote yourself is never changed by it. The check catches those problems; it doesn't guarantee a well-designed slide.

Script shows every word of the deck as one document; type there, or double-click text on a slide. Text you write is kept when the assistant rewrites a slide, unless you ask it to change that text. History keeps a version for every change, labelled by your request, to restore at any time. Two themes are built in, plus any theme the assistant designs for you.

Present opens the deck full screen, with a separate presenter view — current and next slide, speaker notes, and a timer — to keep private while you share the slides. Export saves a single HTML file that presents itself in any browser, a PowerPoint file, or a PDF on Windows only. Slides adds to your local database: going back to a release from before Slides shipped starts with an empty database (the old file is kept as a backup).

Export to PowerPoint. Choose Export, then PowerPoint — slides you can edit, on any platform. The .pptx file matches the deck as Conduit draws it: the same fonts, sizes, colours, cards, lines and backgrounds, in the same places. Text stays editable, charts are native PowerPoint charts (Edit Data opens their numbers), and speaker notes come along. In the other direction, attach a .pptx file to a message, or add one to Documents, and the model reads its slides in order, with tables, chart data and speaker notes. Macro-enabled .pptm files are refused.

Spreadsheets

Ask for a budget, a price list, a loan calculator or any table that has to add up, and Conduit builds a real spreadsheet that opens beside the chat. The assistant doesn't type in numbers it worked out itself: it writes formulas, and Conduit computes them. Number, currency, percent and date formats, several sheets, frozen header rows, and styles for headers, inputs and totals are part of the sheet. When a formula fails, the assistant sees the error (#NAME?, #REF!) straight away and fixes it.

Change it by asking. Like a deck, it's one spreadsheet: "add a row for the security audit", "make column C euros" or "add a sheet for monthly hosting" edits that spreadsheet instead of making another. A new row takes the formats and formulas of the row above it, and a total that ends just above it grows to include it. Amounts use the currency you ask for. The grid is for reading: click a cell, or move with the arrow keys, to see its formula in the formula bar, and switch sheets with the tabs at the bottom.

Excel in and out. Save as Excel (.xlsx)… in the panel menu writes a file that opens in Excel, Numbers or Google Sheets, with the formulas and their computed values. The other way round, attach an .xlsx file to a chat, or add one to Documents, and the model reads its sheets as tables. The panel shows numbers the way your app language writes them, for example 6,000.00 in English and 6.000,00 in German.

Charts. Ask for a bar, line or pie chart and it appears under the sheet, in the app's colours for both light and dark themes. Charts follow the data when rows are added or removed, and the downloaded .xlsx has real Excel charts. To change a spreadsheet's title, ask the assistant to rename it.

Research

For a question that needs more than a quick search, turn on Research in the composer's + menu, or pick it on Home, and ask. It needs web search turned on and is not available in local-only mode. Research uses a local search backend (Exa, Tavily, Brave, your own SearXNG, or DuckDuckGo), never hosted search, so search queries go to that service and the page text goes to the chat provider you picked.

The brief. Before anything runs, Research proposes a brief: one to six sub-questions it will answer, an optional scope, and a depth. You edit it and approve it. Each depth sets how far it may go:

What it reads. It reads pages in full, three at a time: web pages and PDFs up to 20 MB, parsed on your device. A scanned PDF has no text to read and counts as a page with nothing on it. On a long page or PDF the model sees the opening plus the parts that match your questions, not the whole thing.

The quote check. The model pulls out claims, each with the exact words from the page that support it. A claim counts only if its quote (at least 20 characters) is found in the stored text of that page, and that check is done by the app, not the model. Research then searches again for whatever is still unanswered, and a writer turns the checked claims into the report. The numbered citations and the Sources list are produced by the app, not the model.

Source credibility. Each page is rated for credibility. A weak source is marked in the report's Sources, and its claims are attributed to it ("an unreviewed preprint claims…") or left out, never presented as established fact. Official, public-body and university sites are never rated low. The rating is the model's judgement; the quote check above is done by the app.

The review pass. Before the report is finished, a review pass checks every sentence against the facts it cites and the rest of the report, and corrects or removes what they don't support. Research knows today's date, so "current" means now and earlier events read as past. Figures that differ by year, edition, definition or units are explained rather than reported as a disagreement, and a sub-question it couldn't answer names the sites whose pages couldn't be read.

The report. The result is a Markdown document in the chat's documents: a summary, findings for each sub-question, open questions, and numbered sources, with every finding cited. The chat card shows the summary and how many sources were cited and read, and has an Open report button. Stop ends the run and keeps what was found so far.

Pages can't give it orders. Page text only ever reaches a model call that has no tools, so a page that tries to give instructions can't make it search, fetch or save anything. Research's model calls run in a hidden conversation of their own.

Limits: the model sees an excerpt of a long page, not all of it. A scanned PDF has no text. Your own Documents are not a source yet. A thinking model is slow, and Deep can hit its time limit before it finishes.

Writing

Writing, on the rail (Ctrl+5, ⌘5 on macOS), is for long pieces: blog posts, technical docs, reports, newsletters, essays. Open it from the rail, the Writing tile on Home, or the Write chip in Home's ask box, and describe what you're writing.

Outline first. The assistant proposes an outline: the sections, what each must say, and a target length. You edit, add, remove or reorder sections and approve it. Then it writes the draft section by section into a Markdown editor next to the chat; each section appears as it is written, and the draft keeps going until every section of the outline has text.

Sources. In the Sources tab, a draft can look things up on the web (off by default), draw on your Documents collections, or work from a finished Research report, whose quote-checked facts it prefers. A finished Research card in a chat has Write from this report. Facts from a source are linked in the draft, and the assistant is told to mark anything without one [TODO] instead of making it up.

What you write stays yours. Any paragraph you type or change is marked as yours, and the assistant keeps it word for word unless you ask it to change that text; Let AI edit hands it back. Ask for a change in the chat and only that part is rewritten. Select text for Rewrite, Shorter, Longer, Clearer, Fix grammar, or your own instruction. Show AI-written text tints the paragraphs the assistant wrote that nobody has edited yet.

History keeps a version after every change the assistant makes and after your own editing, and restores any of them. Export saves the draft as a Markdown, HTML or Word (.docx) file, the Word file with real headings, lists, tables, links and code blocks; a draft that links to web pages ends with a Sources list of them.

Limits: web lookups need web search turned on and aren't available in local-only mode. Only a finished Research report can be a source. Unlike Research, the app doesn't check a draft's facts against their sources: the links and [TODO] marks are the model's work, so check them before you publish. Export is Markdown, HTML or Word only.

Workflows

The Workflows page runs routines for you. Steps fetch pages, search the web, run Research, search your Documents, read files from the workflow's own folder, turn CSV or JSON into a table, call read-only tools from your connectors, have the model summarise, and ask you a question. The result can be saved as a document, change a Slides deck or Writing draft, be exported as a file (Markdown, text, CSV, JSON or HTML), or be suggested as a memory — a "Save to memory" step only suggests, and never adds one on its own. Start from a ready-made workflow — a morning briefing, a page watch, a weekly research digest, a weekly numbers deck — or build your own in the step editor. Open any run to see what each step did and which model it used; if you fix a step, rerun from it without repeating the steps before it.

Describe it instead. Say what you want in your own words, on the Workflows page or with the Automate chip on Home, and the chat model drafts a workflow, checks it and fixes what doesn't check out. It opens in the editor unsaved, with a short list of what's left for you to pick, such as the folder or the deck; nothing is saved or scheduled until you say so. Save chat as workflow in the command palette turns what you did in a chat into a workflow you can run again; it sees your messages and which tools were used, not the pages or documents themselves. A workflow can use its own model, and each summarise or agent step can override it — a fast, cheap model for the bulk of the work and a stronger one for the write-up, or a local model for steps that touch private data. A workflow with no model chosen follows the chat model.

Run it when it matters. Run it now, on a schedule (daily, on weekdays at a set time, or every few hours), or automatically when a new post appears in an RSS or Atom feed or a new file lands in its folder: switch on Run automatically, and each new post or file gets its own run. A "Continue only if…" step ends a run quietly when a value hasn't changed, is empty, or doesn't contain some text, and notify and save steps can skip themselves when what they would send or save is the same as last time — so a scheduled run that finds nothing new sends no notification. If a feed or folder can't be read three times in a row you get one notification, and the workflow carries on by itself once it can read again.

Turning on automatic runs first lists everything the workflow will be allowed to do on its own — the sites it reads, web search, the folders it reads, connectors, which model and provider, saving documents — for you to approve. A run that needs more, after an edit for example, pauses and asks: allow once, always allow, or don't allow. A connector tool that just reads runs on its own; one that sends, posts or edits something pauses the run and asks (see Outside actions below). Updates to a deck or draft never change text you wrote yourself, and a version is kept before and after so "Undo this update" in the run, or History, puts it back. Each run has a time and token limit, and you can stop a run in progress. A run missed while Conduit was closed happens once when you open it.

To keep schedules running with the window closed, Conduit can stay in the tray and start there when you sign in; both are off by default, and the tray option is offered once, the first time you switch a schedule on.

See it, and shape it, as a canvas. An open workflow shows its steps as a canvas or a list, and Conduit remembers your choice. The canvas draws the start, each step, loops and the lines between them; during a run each step shows whether it is running, finished, failed, skipped or waiting for you, and a loop counts its items. Select a step to see its input and output from a run. To edit, add steps from a palette, drag them to reorder or into a loop, and drag a wire from one step to another to choose which output goes into which field. Steps with problems are marked, and leaving with unsaved changes asks first. A workflow has Steps, Runs and Results tabs. Every finished run keeps what it saved, with a preview; the Results tab lists them, a Results list in the Workflows sidebar covers every workflow and marks unread ones, and Home shows Fresh from your workflows.

Branches and steps side by side. An If… otherwise step sends the run down a Yes or No path, and the path not taken is recorded as Not taken. A step can wait for any earlier steps, so steps that don't depend on each other run side by side, up to four at once, and a later step can join them; Runs after on each step does the same without dragging. Workflows saved before this ran in order and still do. A workflow can also start when a connector's read-only list tool returns something new, checked every 15 minutes to once a day: each item it hasn't seen before starts a run.

Outside actions. When a step wants to use a connector tool that sends, posts or edits something, the run pauses and shows the exact call. Choose Send, Edit and then send, or Don't send, which stops the run with a reason. Don't ask again for this in this workflow allows that action from now on. In the Start panel, Outside actions turns asking on or off for the workflow (on by default) and lists the actions allowed without asking, each removable. Tools a connector marks as sensitive always ask, and the approvals you keep in chat don't carry over to workflows. An agent step can use up to ten connector tools and can't call one its step doesn't list.

Agents and helpers. A workflow can define helpers, each with its own instructions, model, tools and token limit. An agent step names the helpers it may hand work to, writes each a task and gets the answer back. A helper can hand work on one more level at most, circles of helpers are refused when you save, and a step hands out at most twelve tasks. Helpers need the same approvals a step would. The Agents page shows a card for each agent step and helper: whether it is working, its model, tools, whether it uses your computer, its hand-offs, and its tasks, failures and tokens over the last 30 days, with search and All, Working and Helpers filters. Describing a workflow can now draft branches, side-by-side steps, connector steps and helpers; anything the model invents, such as a connector you don't have, is removed and noted.

Agents that use your computer. This is off until you turn on Settings → Computer, which also holds rules that always ask. A workflow opts in per agent step or helper with Use the computer and must name a folder. The agent can then run commands (a program with arguments, never a shell line), list, read and search files, and write or edit files in that folder. Every command asks first unless one of the workflow's allow rules covers it, such as git status *; the approval card offers the narrowest rule as Allow … from now on. Some commands always ask whatever the rules say: deleting files, git reset --hard, git clean, force pushes, system and security tools, shells, and long or unusual commands, plus your own deny rules. Risky commands get their own card with Don't run as the main button. After an agent reads outside content, such as a web page, a search or a connector result, allow rules stop covering its commands for the rest of that run, and the card says why. File changes show a line diff and ask first, unless you let the workflow change files freely in its folder. Each workflow's Permissions tab lists its command rules and how often each was used, the built-in always-ask list, and rules suggested from commands you approved in the last ten runs. Run details list every command and file change, who approved it, the exit code and the output.

Limits: on Windows, commands are also kept from writing outside the folder, with network traffic throttled to almost nothing; a workflow can turn this off for programs that don't work that way. Rust builds (cargo build, cargo test) are one of them, and other compilers and parallel builds may be too: when a command fails because it was kept in, the approval card says so and offers to run that one command outside the folder. You approve it each time, and it is never remembered. macOS and Linux have no such confinement, so your approvals are the only control there.

The Inbox, and runs that wait. The Inbox lists everything your workflows are waiting on: questions, reviews, outside actions and command approvals. The rail shows a dot when something is waiting, and new requests raise a system notification. A run paused for you carries on after you close and reopen Conduit, with its original deadline; steps that finished don't run again and an action already sent isn't sent twice, and a run whose workflow changed while it waited stops and says so. The time spent waiting doesn't count against the run's time limit.

Browser

Agents can read pages in a real browser, so pages that build themselves with JavaScript can be read, and can click, type and choose on them. Conduit uses your installed Edge, Chrome, Brave or Chromium, in a separate profile of its own, never your everyday one. Turn on Settings → Computer first; the same page lets you pick the browser and reset its profile. The profile keeps cookies until you reset it.

In a chat, the Browser button beside Web and Research lets the assistant use the browser for that chat, one message at a time, in the same tab. Approvals and questions appear in the chat and come back if the window reloads. In a workflow, an agent step or helper gets Use the browser to read pages, and Let it click and type to act on them. Always on this site and Allow on all sites remember your answer for that workflow.

Limits: it asks the first time on each new site, and before anything that spends money, changes an account or signs in. The model never sees or types a password or your saved API keys. When a site needs you signed in, you sign in yourself in the browser window, or use a saved sign-in (below). A human check is yours to finish. Local and private-network addresses are refused. A download asks first (Do it, Always on this site, Don't); programs and scripts, files over 100 MB and more than five downloads in one run are always refused. An allowed file is saved with the conversation, not in your Downloads folder, and the model reads it; Save as… on its card keeps a copy. In a workflow run it is also saved to downloads/ in the workflow's folder, never replacing a file already there. Site permission prompts such as location, camera and microphone are refused. Models that accept images can also take a picture of the page when its text isn't enough; the picture is sent to your model provider. Some sites refuse automated browsers, and Conduit doesn't try to get around that. It browses at a human pace by default: it pauses to read each page, opens at most six pages a minute on one site, and backs off when a site asks it to slow down. A chat or a workflow can choose Fast instead. To sign in to a site before the agent needs it, right-click a link in a chat reply and choose Open in the agent's browser.

Saved sign-ins. In Settings → Computer → Saved sign-ins, add a site, a username and a password. The password goes to your system's keychain; the list shows only sites and usernames. When the agent needs to sign in to that site, a card asks you first, then the app fills in the form itself, on one page or a username page followed by a password page. The model never gets the password: wherever a page shows it, the model reads “(saved password hidden)”. If the saved sign-in doesn't get through, you're asked to finish. A workflow that runs while nobody is watching can use a sign-in only if you turn on its switch, which is off by default.

Voice: read aloud and dictation

Read aloud. The speaker button under a reply reads it out with your system's voices, in the reply's language. Click it again to stop.

Dictation. Open Settings → Voice and choose a transcription provider: OpenAI, OpenRouter or Groq, with your own key, or Automatic to use one you already have a key for. The microphone beside Send then records you; click it again, or wait for the five-minute limit, and the text lands in the composer. Escape discards the recording. Until it's set up, the microphone has a dashed outline and opens Settings → Voice; a setting there hides the button. Your voice goes to the provider you chose, and dictation is off in local-only mode.

Home

Conduit opens on Home, first on the rail. One ask box starts anything: ask for a deck and Slides opens with a storyline; anything else starts a chat. Below it, Needs you lists workflow approvals, workflow questions, and memory suggestions waiting on you (hidden when nothing is), Fresh from your workflows shows results you haven't opened, Pick up where you left off shows your latest chats, decks, drafts, and apps, and every area of the app sits in a tile with a live count — or an example to try while it is empty.

Ctrl+1 to Ctrl+9 (⌘ on macOS) open Home, Chats, Slides, Writing, Apps, Workflows, Inbox, Agents, and Context. Manage has no number; reach it from the rail, Home, or the command palette, which has a "Go to" entry for every area.

Appearance and language

Conduit has one design in two modes. Under Settings → Appearance, choose dark, light, or following your system, and pick the main colour — the accent on buttons, selection, and the active item — separately for each mode; a colour that would be hard to read is refused. The same page sets the interface font size, density, and diagram size.

The same settings page sets the interface language, across all eight locales Conduit ships in. It changes menus, buttons, dates, and numbers — it does not change the language the assistant replies in.

Where your data lives

Conduit stores everything in your platform's application data directory: a SQLite database for conversations, messages, prompts, and usage, plus a content-addressed blob store for attachments and artifacts, deduplicated by hash.

Messages are written as an append-only event log with a materialised view over it, which is what lets a conversation survive the app being killed mid-response. On startup, Conduit reconciles the two.

Optional AES-256-GCM encryption at rest is available under Settings → Privacy & Data, with the master key wrapped by your OS keychain. It is off by default and does not yet cover message text — see the privacy page for the current scope.

Troubleshooting

The build fails on Windows

Almost always a missing C++ toolchain. Install the Visual Studio 2022 Build Tools with the "Desktop development with C++" workload, then run check-setup.ps1 to confirm.

My provider key isn't being saved

Conduit needs a working OS credential store. On headless or minimal Linux systems there may be no Secret Service running. Either start a keyring daemon or enable the encrypted file-backed credential store, which requires supplying a key through the environment.

Ollama models don't appear

Confirm Ollama is running and has at least one model pulled, and that Conduit's base URL for it matches the port Ollama is listening on.

An MCP connector won't start on Windows

Two causes cover nearly all of it, both from connectors being started without a shell. If the log says the system cannot find the file specified, the command needs its extension — npx.cmd rather than npx. If it says refusing to spawn connector through a shell interpreter, the configuration is routing through cmd, sh or powershell, which is refused by design; name the program directly instead. See Naming the command on Windows.

Something else is broken

Export a diagnostics bundle from Settings → Diagnostics and attach it to an issue. It excludes secrets, base URLs, allowlists, and conversation content by design, and requires you to acknowledge what it contains before it's generated.

Read the source