# SweetHive Agents Node The **SweetHive Agents Node** is a desktop app for macOS and Windows that runs your SweetHive agents **on your own computer**. Your agents answer with AI models executed locally - prompts and content never leave your machine during inference - and the app can also connect assistants like **Claude** or **ChatGPT** to your SweetHive content through MCP. > Using OpenClaw? It keeps working as an alternative runtime: it reads the same > config bundles. The Agents Node simply makes everything one guided install. ## Install 1. Download the app: - [macOS (Apple Silicon)](https://downloads.sweethive.com/agents-node/mac/SweetHive-Agents-Node-arm64.dmg) - [macOS (Intel)](https://downloads.sweethive.com/agents-node/mac/SweetHive-Agents-Node-x64.dmg) - [Windows](https://downloads.sweethive.com/agents-node/win/SweetHive-Agents-Node-Setup.exe) 2. Open it, review and **accept the terms**. 3. **Sign in with your SweetHive account** (the same email and password as the web app). The app keeps itself up to date automatically. The macOS builds are signed and notarized by Apple, so they open without warnings. ### Updates - When a new version is out, a banner at the top of the app says **A new version is available** with an **Update** button. - From SweetHive web you can update a node remotely: **My agents** shows an **Update** button next to each node that is behind, and the **My SweetHive** dashboard shows a banner. The node downloads the new version and restarts within a few minutes (version 0.2.15 and later). - Older versions update by hand: *Settings, Check for updates*, then *Restart and update*. ## Add context agents A *context agent* is a scoped connection to the hives and contexts you choose - it can never see more than you can. You can add one from either side, and both stay in sync: - **From the app** - *Context agents → New context agent*: name it, pick the hives, choose a capability and expiry. It appears in SweetHive → **My agents** instantly. - **From SweetHive web** - *My agents → Connect an agent*: create the token, download the config bundle and import it in the app (*Context agents → Import config bundle*). Chat with your agents from **My agents** in SweetHive - replies are generated on your computer. See [Chat with your agent](/en/chat-with-your-agent/). ## Local models The AI runtime ships **embedded in the app** - nothing else to install. The *Local models* tab shows a curated catalog organised by machine size, with a recommendation for **your** computer (current families such as Qwen, Gemma, Granite, Nemotron and Mistral): one click installs a model (with live progress), one click uninstalls it, and installed models show a badge. Models that are no longer in the catalog appear only if they are already installed, at the bottom under *Previous models on this computer*. If you uninstall a model that one of your agents is using, the app asks what to use instead (the default model or another installed one), so no agent is left without a model. In *Settings → Storage folders* you can choose where models, temporary files and tool libraries are stored. ## Render workers (SweetReach) The node can also run **render workers**: plug-ins that turn a marketing brief into a finished video for **SweetReach**, SweetHive's marketing app. In the *Marketing* tab, connect the node to the marketing backend with your node secret, then install a worker: - **HeyGen** - photoreal AI avatars via the HeyGen API. No GPU needed; paste your HeyGen key. The node renders the video, downloads it locally, then uploads the finished file to SweetHive storage. - **Local avatar (GPU)** - a talking-head avatar rendered entirely on your GPU, offline, with a local voice. Best on a workstation or server. Video jobs reach the node as SweetHive [jobs](workflows-and-jobs): only a node with a ready worker (a HeyGen key or a GPU worker) takes them, renders the video with the worker you chose and hands the file back to SweetReach. The script is written separately by the agent connected to the campaign's context (which may run on a different node), so each job runs on the right machine. New workers appear in the catalog without an app update. ## Jobs: workflows and app requests The node runs the steps of your [workflows](workflows-and-jobs) and the requests of SweetHive apps with its local models: reading documents for SweetTimetable, rendering videos for SweetReach, generating and extracting text. It takes the work by itself as soon as SweetHive has something for it: there is nothing to configure app by app. The **Jobs** section of the app shows, in plain words, what the node is doing now (with its progress) and what it did recently, including why something failed (for example a model that is not installed or a document too long for the model). The same list is in SweetHive web under **My agents, Jobs**. ## Reading documents (SweetTimetable) The scans and photos you upload to SweetTimetable become jobs in **My agents, Jobs**, and they are run by your node, by the hive's shared node or by your organization's [Agents Server](agents-server). In SweetTimetable you see the result: the proposed expenses to check in *Upload documents*. - PDFs that contain real text are read directly on the server, with no node needed. - A page with several receipts is read in sections (each motorway toll is its own expense): with a vision model such as qwen2.5vl:7b it takes about two minutes on a recent Mac, and the full transcription arrives after the proposed expenses. - If a job waits, SweetTimetable and My agents tell you what to do: turn the computer on, update the Agents Node, or choose a model that can see images. - Always check the proposed expenses before creating them. ## Connect Claude or ChatGPT The *Connect AI apps* tab turns the node into an MCP server for your assistant - now with **one-click setup**: - **Connect Claude Code** - finds your `claude` CLI and registers the node automatically; open a new session and ask about your SweetHive content. - **Connect Claude Desktop** - writes the entry into Claude Desktop's configuration for you; restart Claude Desktop and the tools are there. - **Developer options** - the JSON snippet and a secret-protected local HTTP MCP endpoint for other clients. ## SweetHive Desktop - chat with your computer The Chat tab includes **SweetHive Desktop**: conversations with the machine itself, organised like a modern AI chat (a titled thread list, *New chat*). Answers come from your default local model and your data stays on the computer; the conversations sync to your account, so you can also open them **from SweetHive web** (*My agents → SweetHive Desktop*) and talk to your computer remotely - it answers within seconds whenever the app is running. ## Privacy & control - Tokens are scoped, expire, and can be revoked anytime from **My agents**. - Secrets are stored encrypted with your operating system's key store. - Model inference runs entirely on your machine.