Agents and Chat

An agent is a reusable AI teammate with its own instructions, skills, and conversation history. A new installation comes with an assistant ready for everyday work, and you can add specialists when you want a separate role.

Each conversation keeps its agent, workspace, files, model, and earlier messages, so you can keep working on the same thing without re-explaining it.

Start a conversation

1. Start from Home, or open Chat to continue an earlier conversation.
2. Use your assistant, or choose another agent. If it lives on another device, that device must be online.
3. Add files, images, or a project folder if the work needs them.
4. Ask normally. Progress, setup steps, approvals, and results appear in the conversation.

Choosing a model is optional. Open Advanced to change it and other detailed controls, and select Keep visible to keep them open across restarts.

You can send another message while the agent is still working; each message starts its own run. Chat history is kept across restarts in both the desktop app and the browser.

Choose a mode

Switching modes keeps the conversation and attachments. No mode grants extra permission to act or spend. Chat mode blocks changes; to act, switch to Do or Create yourself. If an agent can’t guarantee Chat’s no-changes rule, that mode is unavailable for it.

Permission requests in the chat

When an agent needs something it isn’t allowed yet, such as running a program on your computer or reading your finances, the chat shows a card with the choices right there:

• Just this time allows this one request.
• For today allows it until midnight.
• Always allows it from now on.
• Not now declines this time; the agent may ask again later.
• Never declines and stops the agent from asking again.

Your choice applies to that agent only. Choosing Always later undoes a Never. For finances, you can review or change every agent’s access in Finances → ⋯ → Agent access. Sending money, trading, and other payments always ask you each time, with no Always option.

Use Work instead when a request should be assigned, tracked, reviewed, or kept with formal deliverables.

Choose the right agent

• A local agent for files and apps on this computer.
• A remote agent when the project or tool lives on another trusted machine.
• A managed cloud agent when work must continue while your computers are off.
• A specialist for its instructions and skills, not just its name.

If something isn’t set up, Chat shows the setup right in the conversation and keeps your prompt.

Know what an answer rests on

A model answers from a compressed memory of its training. That is fine for common, stable facts and unreliable for anything recent, niche, or numeric. Chat shows you which kind of answer you got:

• Model badges. The model picker marks models that can use Tools (search, files, commands, connected apps), read Images, or do Thinking on hard problems. A badge appears only when HivemindOS is sure. If your model can’t use tools, the picker says so, and the agent answers from memory instead of pretending to look things up.
• Where the answer came from. Each reply is marked From memory or Used tools, with a count of files or images it read. Treat “from memory” answers about recent events, exact numbers, or obscure topics as a draft to confirm.
• Start a fresh chat. Everything in a conversation is sent with each new message. When a chat grows long or drifts to a new topic, a Start a fresh chat suggestion appears. Dismiss it to keep going.
• Read images as text. Under Advanced, switch Images from as-is to read as text. Each image is first transcribed word for word as a step you can check. Use this for labels, receipts, and anything with numbers.
• Examples. In agent settings, Custom instructions has an Examples section: pairs of when I ask something like… and answer it like this. One example steers format and tone better than a paragraph of rules.

Use the thumbs-up and thumbs-down buttons on a reply to rate it.

Use a model on this computer

Your first assistant, and any assistant left on the default model, uses a model running on this computer when one is loaded in LM Studio, Ollama, or another local model app. Otherwise it uses the free online model. Choosing a specific model turns this off.

If the online model can’t be reached or today’s free allowance is used up, a model loaded on this computer answers instead, if there is one. A reply that needs payment still asks first. The line under each reply names the local model that answered and why. Incognito conversations always stay on the private online model.

While Incognito is on, a switch above the message box chooses what the chat keeps afterwards:

• Don’t save (the default): the conversation is not added to your chat history or your brain and is not synced. It disappears when you leave the chat.
• Save on this computer: kept in your chat history on this computer only. It is not synced to your other devices, not uploaded, and not turned into memories.
• Save to my account: saved like any other chat.

A chat never becomes less private than the choice it started with. Turning Incognito off later leaves a chat saved on this computer where it is.

When a reply fails, Retry replaces the failed reply with a new attempt in the same place, so the conversation shows one answer, or one error if the new attempt fails too.

A local model answers only when the whole conversation fits in its context, so your assistant’s instructions are never cut off. Many local apps start at 4,096 tokens, which is too small. Load the model with a context of 16,384 or more. In Ollama, set OLLAMA_CONTEXT_LENGTH=16384 before starting it.

Attach documents and build apps

Attach a document for one conversation, or add it to Hive Superbrain so agents can find it later. Text inside a document or web page never becomes an agent instruction. See Documents And Brain Drop for formats and privacy.

Ask for a new app, site, dashboard, or game and Chat creates a project for it in the built-in App workspace. The project stays attached to the conversation, and Preview opens that exact project on the chosen machine.

Phrases that explicitly ask to teach the hive something create a Brain Review proposal. Nothing becomes shared memory until it is approved.

Sphere view

Turn on Sphere view at the top of Home to try an experimental layout with a large sphere and a floating input. It is off by default. With Chat here on, tasks you start from Home open around the sphere; Focused view switches to a quieter layout, and Return to sphere switches back with your conversation and draft intact.

Show app preview and Show terminal open side panels. The terminal shows which computer your commands run on; if none is selected, choose one there or select Connect a computer.

Swarm commands

• /swarm [number] <task> asks a team of your ready agents for role-specific passes and returns one combined answer. Add a number to cap the team size.
• /swarm-goal <build request> expands a rough build idea into a full brief and hands it to Queen Bee as tracked, evidence-checked Work. See Queen Bee Swarm Goals.
• /swarm-sim <scenario> starts a MiroShark simulation and shows its queued status.

The swarm button in the composer only fills in /swarm ; nothing runs until you send. See Slash commands for the full list.

Automatic thread titles

Open title settings from the Chat sidebar to keep first-message previews or turn on Local or Cloud titles.

• A title is created from your first real message, alongside the reply. Greetings like “hello” wait for the next message.
• Local uses a small Qwen3.5 0.8B model or a larger 4B option, installed through LM Studio in the app.
• Cloud lets you pick from your configured models.
• Only your first and latest real messages are used, up to 600 characters each. Tool output and attachments are excluded, and text sent to the cloud is scrubbed of secrets first.
• A title is set once and isn’t replaced later.

Voice and calls

• Talk with Scout. In a chat with Scout, the waveform takes Send’s place while the message box is empty. Select it to open a voice chat over the whole window: say what you want, and Scout answers out loud while what you both said shows in small bubbles. Talk over him to stop him. Mute or end the chat at the bottom, and use the gear to pick how you talk, including live voices that answer in their own voice. What you say is sent as your message in that chat, so the conversation is all there afterwards. The first time, the app asks before anything is sent.
• Hear Scout’s replies. Turn on Speak replies in Settings > Voice to have Scout read his replies aloud as they arrive, in the voice you pick. Each voice has a free sample. Spoken replies and voice chats use your HivemindOS credits.
• Voice chat with Queen Bee. Choose Voice Chat with Queen Bee from the menu bar icon or Navigation menu (Cmd+Shift+V). The window shows a glow and a live transcript while you talk. Replies use OpenAI speech when an OpenAI voice key is set up, or your computer’s built-in voice otherwise.
• Voice in a conversation. Select the voice icon in Chat to talk with your agent. You can type during the call or use the camera button to share what you see.
• Agent calls. Start an in-app voice call with an agent, or have a scheduled call ring your paired phone. See Calling for free and paid options.

Advanced: runtimes and providers

You don’t need this section for normal chat. It is a reference for choosing a specific agent engine, model provider, or local model app.

Runtimes

Managed background tasks. Codex, Claude Code, and HivemindOS can run background tasks that keep their folder, model, status, logs, and completion check in one record. A signed-out Claude Code install is not shown as ready. HivemindOS tasks stay inside the selected folder, follow the task’s permission setting, and can be stopped or resumed. Running codex or claude yourself in a terminal is outside HivemindOS; report that work through the Work board if it should count. See Agent Evaluations.

AEON is optional. Choosing it for a Zero Human Company links that one company to one AEON workspace and skill; it doesn’t take over other agents or settings. See Using AEON With Zero Human Companies.

ChatGPT sign-in and API billing

When ChatGPT sign-in (OAuth) is connected, HivemindOS uses it for OpenAI chat, including delegated Hermes runs and Queen Bee. This is checked again when each run starts, so an older tab or synced device can’t switch you to API-key billing. If a signed-in run fails, you see the failure instead of a silent retry on your API key.

To use API-key billing for OpenAI chat on purpose, set OPENAI_PREFER_API_KEY=1, or disconnect ChatGPT sign-in. Separate OpenAI features such as speech or image generation may still use their API key.

Local model apps

• LM Studio can be installed, started, and controlled from agent settings: download, load, unload, and test models, and find models shared through LM Link.
• Local Model Advisor, inside the LM Studio panel, checks your CPU, GPU, and memory on this computer and recommends models that fit. It ranks all 7,738 GGUF candidates from a built-in 9,810-model snapshot and downloads your pick through LM Studio (1,564 entries have a direct download route; the rest use LM Studio’s search). It needs no extra install, and nothing is downloaded until you choose a model.
• FreeToken serves large Mixture-of-Experts models. HivemindOS finds its server on port 1919 on this machine or another machine in your hive. FreeToken 0.1.2 needs a Linux x86_64 host with an NVIDIA GPU, driver r580 or newer, and CUDA 13; a Mac can use a supported machine in your hive. Install it yourself, start it with ft serve --model <path-or-hf-id>, then refresh Local models.
• SIE is for a shared GPU server. Point the SIE provider at its gateway (default http://127.0.0.1:8080) to see its models, health, queue, and GPU memory. Generation models can be used in chat, and generation and embedding models can be warmed from the panel. SIE decides which models stay loaded. See the SIE project.

Adaptive routing

The Adaptive provider picks a model for each message from OpenRouter and any configured OpenAI-compatible providers, instead of using one fixed model.

• Candidates are scored on availability, health, cost, speed, task fit, privacy, context size, and past results. Choose Balanced, Fast, Quality, Cheapest, Reliable, or Private to change the weighting.
• Best free never uses paid models. Free then fallback tries free routes first, then only the fallback model you entered.
• Optional cost and speed caps are strict: under a cost cap, paid routes with unknown pricing are skipped; under a speed cap, routes with unknown speed are skipped.
• Rate-limited or failing routes are paused for a while, and models that keep giving poor answers rank lower. Each run records which route was chosen and why.
• Long Hermes tasks can stay quiet for up to 15 minutes while working; the chat limit is 20 minutes. If one route fails mid-task, the next resumes the same session.

Hive Fusion compound model

The Hive Fusion provider sends your prompt to a panel of your configured models at once. A judge model compares their answers, and a synthesizer writes one final answer from that analysis. If only one provider is set up, it still works with a one-model panel. The panel, judge, synthesizer, and panel size (default 3) can be customized. OpenRouter’s hosted version is offered separately as openrouter/fusion under OpenRouter.

Hive Compute models

In the HivemindOS provider’s All models list, Auto, Fast, and Deep carry SALE badges. They try marketplace GPU workers first and fall back to the matching hosted model when marketplace capacity is unavailable. See Hive Compute to earn from your own spare GPU.

Related

• Documents And Brain Drop
• Work and automations
• Slash commands
• Calling