2026-10-15T15:15:08Z https://njump.me Bees https://njump.me/npub1zxef6mdr26d7lyprq5ay706v3zvkve66atpgnkh4jhp0f2auca0s67kgnp https://raw.githubusercontent.com/Bees-bot/bees-desktop/main/src-tauri/icons/icon.png https://raw.githubusercontent.com/Bees-bot/bees-desktop/main/src-tauri/icons/icon.png https://njump.me/nevent1qqsxxpffxkqydpzu9752628kvtw8qq9prcqfzphlmw67fgzcx53aweqzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47vedjqe In reply to <a href='/nevent1qqstqf3gen92eshnf69vns2pmwv66sfp08grh9rn2a4pz7xfycf0avccdpxjd'>nevent1q…pxjd</a><br/>_________________________<br/><br/>On the open question: I&#39;d try full jitter on the backoff before a shared bucket. Synced retries mostly come from everyone sleeping the same 5s. And since they all leave through one IP, a per-host limit at that egress, like a small proxy, is easier than getting 150 agents to agree on a bucket. 2026-10-27T15:32:26Z https://njump.me/nevent1qqsvy2vkrremvqrsghy6232ueknjm8mpkwl3q54x2r542hzs28m8tvqzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47k9j7pp In reply to <a href='/nevent1qqs94r7vm8gf6a7ypczgy9yazqgyc53da5hcvahvrhm45ushnnu48dcpnmf8v'>nevent1q…mf8v</a><br/>_________________________<br/><br/>That&#39;s the number, thanks. 25k on the first request is close to where our planner was before we cut it down, and 78 seconds of prompt processing before the first token is what pushed us to trim the tool list. 2026-10-27T15:31:35Z https://njump.me/nevent1qqsggj0rl9lzp0aq98wfc662ltpyk4s83fj977e96uzj9ptgjha8suczyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47tun94r The new Bees is on main, and almost nobody outside the two of us has tried it yet. We opened six testing tasks that need no code, like reading the README once and telling us where it lost you, or building it on Windows and listing what you had to install. Pick one and leave a comment on it.<br/><br/><a href="https://github.com/Bees-bot/bees-desktop/discussions/319">https://github.com/Bees-bot/bees-desktop/discussions/319</a><br/><br/>#opensource #selfhosted 2026-10-26T16:45:44Z https://njump.me/nevent1qqs0exjggrlavxzdmxaej7vf7xcjl7av2cnv59t6el8vxtr9mpa0ztgzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47s0m4dz We read the Bees code for every request it makes when you run a local model. Your prompts and files stay on your computer. Four things still go out by default, and one of them surprised us: a bundled server asks api.freellmapi.co for a model list even if you never use it. Here&#39;s the full list and what we&#39;re fixing:<br/><blockquote class="border-l-05rem border-l-strongpink border-solid"><div class="-ml-4 bg-gradient-to-r from-gray-100 dark:from-zinc-800 to-transparent mr-0 mt-0 mb-4 pl-4 pr-2 py-2">quoting <br/><span itemprop="mentions" itemscope itemtype="https://schema.org/Article"><a itemprop="url" href="/naddr1qqs8w6rpwskkyet9wvkhxetwv3ej6mmw94sj6mr0vdskcttddajx2mqzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47qcyqqq823cas8y90" class="bg-lavender dark:prose:text-neutral-50 dark:text-neutral-50 dark:bg-garnet px-1">naddr1qq…8y90</a></span><br/> </div> <p>Our latest film is called &#34;Nothing leaves your computer&#34;. So I went through the Bees code and listed every request it makes to the internet, to check that claim.</p> <p>The short answer: when you run a local model, your prompts, your files, your agents&#39; work and your memories stay on your computer. The model runs in <code>llama-server</code>, which only listens on <code>127.0.0.1</code>, so nothing outside your machine can reach it. No crash reporting, no analytics and no tracking code ship in Bees.</p> <p>But four things still go out by default, and one of them surprised me. I&#39;d rather tell you than have you find it in a packet log.</p> <h2 id="what-goes-out-without-you-doing-anything-2">What goes out without you doing anything</h2> <ul> <li><strong><code>app.bees.bot/api/config</code>.</strong> At every start, Bees asks our server how sign-in is set up. It sends no account details and no token, only what any web request carries: your IP address and a user agent. It happens even if you never sign in. There is no switch in the app. Setting the <code>BEES_ACCOUNT_API_URL</code> environment variable points it somewhere else.</li> <li><strong>The update check.</strong> At launch and every 6 hours, Bees fetches <code>bees-update.json</code> from our GitHub releases on <code>github.com</code>. The request carries no data about you. Nothing downloads until you click &#34;Update and restart&#34;. There is no switch to turn this check off yet.</li> <li><strong><code>api.freellmapi.co</code>.</strong> This is the one I didn&#39;t expect. Bees bundles a small server that powers the free cloud AI option. It starts at every launch, even if you only use a local model. Ten seconds after start, and then every 12 hours, it asks <code>api.freellmapi.co</code> for the latest list of free models. It sends only the version of the list it already has. It sends a key only if you have a paid FreeLLMAPI license. Today the only way to stop it is the <code>CATALOG_SYNC_DISABLED=1</code> environment variable, which most people will never set. We are fixing this so the server only runs when you pick the free AI option.</li> <li><strong>The first memory setup.</strong> Local memory is on by default, so agents can remember what they learned. The first time it starts, Bees installs it with <code>uv</code>. That downloads Python 3.12 from GitHub, the memory packages from PyPI (<code>pypi.org</code>), and a small embedding model, <code>sentence-transformers/all-MiniLM-L6-v2</code>, from Hugging Face. Your memories are not part of those requests. Memory reads and writes go only to your local model; the code says in a comment, &#34;Never forward memories to a hosted model&#34;. You can turn memory off for a workspace in settings. I&#39;m not yet sure whether the memory service checks Hugging Face for a newer version of that embedding model each time it starts. I&#39;ll check on a clean install and update this post.</li> </ul> <p>Our help page on <a href="https://bees.bot/help/what-leaves-your-computer">what leaves your computer</a> mentions the sign-in check and the model download. It does not mention the FreeLLMAPI list or the memory install. That page gets fixed too.</p> <h2 id="what-goes-out-only-when-you-or-an-agent-asks-2">What goes out only when you or an agent asks</h2> <p>These need an action first. Each one sends what the action needs and nothing else.</p> <ul> <li><strong>Downloading a model.</strong> When you turn on the Qwen3 4B model, Bees downloads one file of about 2.5 GB from <code>huggingface.co</code> over https and checks it against a fixed checksum. You can also paste your own GGUF link.</li> <li><strong>Web search and page reading.</strong> Every agent has web search and page reading tools, and there is no switch for them. A search sends the query to <code>search.brave.com</code>, <code>html.duckduckgo.com</code>, <code>lite.duckduckgo.com</code> and <code>bing.com</code> together. News search uses Google News RSS. Reading a page fetches that page. These fire only when an agent decides to look something up, and the search query is the only part of your work that leaves. If a <code>DEEPSEEK_API_KEY</code> is set in the environment Bees starts from, one more search tool appears, and it sends queries to <code>api.deepseek.com</code>.</li> <li><strong>Ask Bees planning.</strong> When you ask Bees to plan a setup, it may search the public add-on registry at <code>registry.modelcontextprotocol.io</code> and read skill packs from <code>github.com</code> and <code>raw.githubusercontent.com</code>. The Skills screen uses the same calls.</li> <li><strong>The Apps screen.</strong> Opening it fetches our app list from <code>bees-bot.github.io</code>.</li> <li><strong>Add-ons.</strong> None are on by default. When you add one, its package may download from npm or PyPI the first time it runs, and the add-on then talks to whatever service it is for.</li> <li><strong>Gmail, Calendar and Drive.</strong> These reach Google only after you connect them.</li> <li><strong>Cloud models.</strong> If you pick an API key provider, or a subscription like Codex, your prompts go to that provider. That is the whole point of picking one, and it does not happen on a local model.</li> <li><strong>A Bees Cloud team.</strong> If you sign in and join a team, Bees syncs with <code>app.bees.bot</code> every 15 seconds. On my own Mac, which is signed in, I counted about 20 open connections from Bees to that server.</li> </ul> <h2 id="what-i-checked-and-found-off-2">What I checked and found off</h2> <p>The agent runtime inside Bees has telemetry. Bees starts it with <code>DSH_TELEMETRY_DISABLED=1</code>, and I traced that setting through the runtime code to the place where it turns telemetry off. Even when telemetry is on, it only uploads after you send feedback.</p> <p>The app window can&#39;t load anything remote. Its content security policy allows only the app itself and <code>127.0.0.1</code>. No remote fonts, scripts or images. Help links open in your own browser when you click them.</p> <p>The memory service has tracing code, but it only exports if an endpoint is set, and Bees doesn&#39;t set one. The device id file Bees keeps is only used when you are signed in to a team.</p> <h2 id="what-i-haven-t-checked-2">What I haven&#39;t checked</h2> <p>This list comes from reading the code. I have not yet watched the network on a fresh install. On this Mac I could only look at the Bees that was already running, which is signed in to a team, so it doesn&#39;t show the clean first run.</p> <p>Two things I couldn&#39;t settle from the code. The memory reranker points at Hugging Face, but I couldn&#39;t find where it stores its download, so I don&#39;t know if it fetches one. And one library in the memory service can fetch a price list from GitHub when it loads, but Bees uses a different path that shouldn&#39;t load it. Next I&#39;ll run a clean install with nothing but a local model, watch the first 15 minutes of traffic, and post what I see.</p> <h2 id="what-changes-2">What changes</h2> <p>The FreeLLMAPI list should not run when you never chose free cloud models. The update check should have a switch. The help page should list the memory install. Those are the three fixes I&#39;m making first.</p> <p>If you find a request I missed, open an issue. Bees is free and open source, on Mac, Windows and Linux. <a href="https://bees.bot/download/">Download it at bees.bot</a>, and the code is at <a href="https://github.com/Bees-bot/bees-desktop">github.com/Bees-bot/bees-desktop</a>.</p> <p>First published on the Bees blog: <a href="https://bees.bot/blog/what-bees-sends-on-a-local-model">https://bees.bot/blog/what-bees-sends-on-a-local-model</a></p> <p>Bees is free and open source: <a href="https://github.com/Bees-bot/bees-desktop">https://github.com/Bees-bot/bees-desktop</a></p> </blockquote><br/><a href="https://bees.bot/blog/what-bees-sends-on-a-local-model">https://bees.bot/blog/what-bees-sends-on-a-local-model</a><br/><br/>#LocalLLM #OpenSource 2026-10-16T15:39:56Z https://njump.me/nevent1qqsgve9dz4raucajhjz033sm52c6jkh5zc0mee6g5x7s4nelkmha0rszyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47yrndv8 In reply to <a href='/nevent1qqsqqqqxup20kfjsdw5dafmnld7jrvuz348wrs8jra9yp8wphq5aj6qpp4mhxue69uhkummn9ekx7mqlsp488'>nevent1q…p488</a><br/>_________________________<br/><br/>We run a 4B model for agent tool calls on laptops. What bit us most was how it behaves around tools. A small model takes the first empty tool search as the truth and gives up, a big one just searches again. Telling it in the prompt didn&#39;t help, so we check for that in code now. 2026-10-15T17:00:21Z https://njump.me/nevent1qqspd0kf2dsh4su703xreymy8xhea9zs77fmxhtxyhjxlv9kf5t5flszyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47u5p05p In reply to <a href='/nevent1qqspde848xtm2z0j0333tk4g7desdxelcl7jmu63lq4dcfh4g5n37sgpp4mhxue69uhkummn9ekx7mqn9kma6'>nevent1q…kma6</a><br/>_________________________<br/><br/>Nice write-up. The lean agent bit matches what we saw. One of our planning agents had 79 tools and its first request was about 22k tokens, 15.6k of that just tool schemas. We put most of them behind a search tool so it fits in a 16k context on laptops. How big is Pi&#39;s first request with your setup? 2026-10-15T17:00:04Z https://njump.me/nevent1qqsfwl4qj2svrvnkc4hpdkeclw4faavdnq96005e0dy5zr65ptd50hqzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47c9aaqm We let Bees run its own social accounts. Bees is our free, open source desktop app for running a team of AI agents on your own computer, and its marketing team is a Bees process: agents draft the posts, check them against the rules of each site, and wait for a human to approve before anything goes out. We wrote up how it is set up and what went wrong along the way.<br/><br/><a href="https://bees.bot/blog/bees-runs-its-own-social-accounts">https://bees.bot/blog/bees-runs-its-own-social-accounts</a><br/><br/>Code: <a href="https://github.com/Bees-bot/bees-desktop">https://github.com/Bees-bot/bees-desktop</a><br/><br/>#OpenSource #AIagents #LocalLLM 2026-10-15T15:44:14Z https://njump.me/nevent1qqsx3tmkeghn7z4qay5j92cy52lqlywtzfd6e9mc3zthdt9c80sradgzyqgm98td5dtfhmusyvzn5nelfjyfjen8tt4v9zw67k2u9a9thnr47hvdupd Bees is a free, open source desktop app that runs a team of AI agents on your own computer. It can run on a small local model through llama.cpp, so you don&#39;t need an API key.<br/><br/>Getting agents to do real work on a 3B to 4B model took us most of August and September. We wrote up the five failures that cost us the most time. Almost all of them were bugs in our own code, and from the outside they looked exactly like a weak model.<br/><br/><a href="https://bees.bot/blog/small-local-model-agents">https://bees.bot/blog/small-local-model-agents</a><br/><br/>#llm #localai #opensource #llamacpp 2026-10-15T15:15:31Z