Use TensorFold,
step by step.

Find it in the app

Click a screenshot to enlarge

123Enlarge ↗

1. 1

   Choose Minimal for the small local-server interface.

2. 2

   Choose Full for chats, projects and connected tools.

3. 3

   Continue with the mode you selected.

Open full-resolution image ↗

Follow the steps

1. Choose Download for Mac and open the downloaded DMG file.

2. Drag TensorFold into Applications, then open TensorFold from Applications.

3. On the first screen, choose your language and select Minimal for a local model server or Full for the chat and work interface.

4. Continue to Your model and choose the built-in TensorFold engine to run a model on this Mac; use a compatible model that passes the hardware checks.

5. Complete Get started, review the suggested first task or choose a blank chat, then send your first message when you are ready.

Your first conversation

The app includes its engine. You do not need the separate CLI installation to chat in the app.

1. Open Find, download and load a model to choose a compatible checkpoint.

2. Wait for its download to finish, load it, and check that its state is Running.

3. In Full mode, choose Chat and ask a short question such as "Explain what you can help me with in two sentences."

Find it in the app

Click a screenshot to enlarge

SettingsMinimalChat layoutComplete workspace

123Enlarge ↗

1. 1

   Open App & onboarding under Preferences.

2. 2

   Choose Minimal or Full under How you use TensorFold.

3. 3

   In Full, choose Chat layout or Complete workspace.

Open full-resolution image ↗

Choose the app mode

1. Open app settings

   Open Settings, then choose App & onboarding under Preferences.

2. Choose Minimal or Full

   Under How you use TensorFold, choose the mode you want. Your models and conversations stay available.

If you chose Minimal

Use these controls when you choose Minimal.

1. Use the Minimal tabs

   Server controls the model, Models manages downloads, and API connection provides the connection details for another app.

2. Hide the window or open Full

   Choose Run in menu bar to keep the server out of the way, or Open Full to return to chats and the rest of the app.

If you chose Full

These two layouts are available inside Full mode.

1. Choose a Full layout

   Back in App & onboarding, find Full workspace layout. Choose Chat layout for conversations or Complete workspace for projects and developer views.

Find it in the app

Click a screenshot to enlarge

Complete workspaceChat layoutMinimal

1234Enlarge ↗

1. 1

   Start a task in a folder or open a new chat.

2. 2

   More opens Assistant, Rooms, Automations, Plugins and Command Centre.

3. 3

   Models manages downloads and the local engine.

4. 4

   Open Settings to configure the app.

Open full-resolution image ↗

Follow the steps

1. For the full navigation, open Settings → App & onboarding, choose Full, then Complete workspace.

2. Use New chat for a conversation, New task for work in a chosen folder, and Search for saved messages.

3. Open Models for the model library, or Bots for saved model-and-instruction presets.

4. Expand More in the sidebar for Assistant, Rooms, Automations, Plugins and Command Centre.

5. Open a project to see Overview, Rooms, Kanban, Solo, Files, Terminal and Settings.

6. Open Settings for AI & Models, Assistant, Behavior and Preferences groups.

What to do

1. Open Settings → Runtime and find the app update controls.

2. Choose Check for updates and review the displayed version and release notes.

3. When an update is available, choose Download update and wait for downloading and verification to finish.

4. Finish or stop active work, then choose Install and restart when offered.

5. If automatic installation is unavailable, choose Download, verify and open or Open verified installer, then replace TensorFold in Applications.

6. After the app reopens, reload your model if needed and copy fresh API connection details for other clients.

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Open Models to download or manage a model.

2. 2

   Settings contains modes, accounts and connectors.

3. 3

   Switch to project navigation and developer panes.

4. 4

   Choose a model and enter the task or message here.

Open full-resolution image ↗

Follow the steps

1. In Full, choose New chat and select a ready model in the composer.

2. Type your request, then send it; Shift + Enter adds a new line, and Settings → App & onboarding lets you change the send key.

3. To include a file, open the attachment menu and choose Attach document or image; complete any extraction review or resolve any file warning before sending.

4. To reuse a previous input, open the clock icon labelled Recent prompts and choose Insert; review the appended draft before sending.

5. To correct a sent message, stop any active run, choose Edit below your message, review the effect, then choose Save changes or Save as edited branch.

6. Use Stop while a response or task is running when you want it to finish early.

What to do

1. Choose Search in the chat navigation or press Command + Shift + F to search all saved conversations; Command + F searches the active conversation.

2. Enter a phrase, question or decision and select All saved conversations or Current conversation.

3. Use the project, archive and other filters to narrow the results, then open the matching message.

4. In the conversation, choose Export readable text.

5. Choose Markdown, Plain text or Self-contained HTML, and select the active conversation or specific messages.

6. Choose Preview export, inspect the text and omissions, then choose Save reviewed export and pick a local file destination.

What to do

1. Open Settings → App & onboarding and find Language and reading comfort.

2. Choose Language, reading direction, chat text size, code text size and spacing, then check the preview.

3. Enable Use this language for new assistant replies if you want future replies to follow the interface language, then choose Save.

4. For speech playback, open Settings → Voice and turn on Speech bridge and Text-to-speech.

5. Choose Kokoro · offline voices, then Download Kokoro; after it finishes, choose a voice and Preview voice.

6. For dictation, enable Speech-to-text with System dictation, when available, then choose Voice input in the composer's More options menu.

What to do

1. Open Settings → Security and find Encrypted workspace backup; save or discard any pending settings edits first.

2. Choose Create backup, then select conversations, projects, bots and the optional preferences, unfinished messages and managed attachments.

3. Choose Preview selected backup and review its counts, contents and warnings.

4. Enter a backup password of at least eight characters, then choose Save encrypted backup and a local destination.

5. To restore, choose Restore backup, select Keep existing records or Replace matching records from the backup, enter the password and choose Choose backup and preview.

6. Review the conflicts and warnings, confirm the reviewed choices, then choose Restore reviewed items.

What to do

1. In a chat or new-task composer, open More options → Memory.

2. Choose Use saved memory to allow recall, and Propose new memories to allow suggestions for your review.

3. Choose Save choices or Use for new chat; the selection applies to the next turn.

4. For a temporary new conversation, enable Incognito on the home composer before starting it.

5. Use an encrypted workspace backup when you want a recoverable copy of saved records.

6. Only when you intend to erase local records, open Settings → Runtime → Power & local controls → Purge memory and read the confirmation in full.

What to do

1. Open a conversation and choose Outline above the chat, or press Command + Shift + O.

2. Filter by words or choose an entry type such as Your message, Answer, Recorded outcome or Pinned decision.

3. Choose an entry to jump to its original location in the active conversation.

4. Use Previous entries and Next entries when the outline spans several pages.

5. If the outline reaches its scan limit, use Find in chat to search the remaining conversation.

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Discover compatible models to download.

2. 2

   Add a model folder you already have.

3. 3

   Follow current downloads here.

4. 4

   Start a compatible cached model in Minimal.

Open full-resolution image ↗

Follow the steps

1. Open Models, then choose Discover.

2. Use a supported starting point or enter a model name or owner/model in Search Hugging Face.

3. Review the compatibility result, memory estimate and download-space check before downloading.

4. Open Downloads to follow the queue, cancel a download or retry one that failed; completed models appear in My models.

5. Choose Use model in Full, or Start server in Minimal, and wait for the Running state; in Full, choose Chat to open a conversation with it.

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Discover compatible models to download.

2. 2

   Add a model folder you already have.

3. 3

   Follow current downloads here.

4. 4

   Start a compatible cached model in Minimal.

Open full-resolution image ↗

Follow the steps

1. Open Models → Model folders, or Settings → Model folders.

2. Turn on Scan the Hugging Face cache, or choose Add folder and select the model folder in the file picker.

3. Choose Scan now and wait for the compatibility check; choose a smaller folder if the scan reaches its limit.

4. Read any reason shown beside the checkpoint, then choose Load when it is available.

5. To stop checking a folder later, choose Remove folder.

What to do

1. Open Models → My models and click the model name to open its details.

2. Choose Settings, then select a profile or choose Add small context, Add large context or Create profile.

3. Set Context window for the prompt and reply together, then set Maximum reply below that window.

4. Adjust Temperature; open Advanced sampling and acceleration for Top P, Top K, Min P and Speculative drafting.

5. Choose Save & apply and wait for the confirmation; if the model was loaded, reconnect API clients using its fresh connection details.

What to do

1. Load a model and wait until the engine is ready.

2. Open Models → API in Full, or API connection in Minimal.

3. Choose Copy URL and paste it into your client's OpenAI-compatible base URL field.

4. Choose Copy API key and paste it into that client's API key field; use the Model ID shown in the panel.

5. For a terminal check, choose Copy authenticated request and run the copied streaming curl command.

6. After an engine restart or model reload, copy the connection details again.

App API or standalone CLI?

The app uses a new local port and API key for each engine session. The standalone CLI normally uses http://127.0.0.1:8080 until you choose another host or port. Always copy the app connection values when connecting to the app.

What to do

1. For a comparison, install two models confirmed to fit this Mac and finish active local work.

2. Open Models → My models → Compare models, select model A and model B, enter one prompt, then choose Compare models.

3. Read both answers and their timings; the app restores the previously loaded model when the comparison finishes.

4. For repeatable measurements, open Models → My models → Benchmarks and choose a compatible cached model.

5. Optionally enable Serial baseline, then choose Run benchmark; reopen this panel to follow or cancel the job.

6. When it finishes, choose Copy basic results or Review and export JSON; review the exported report before sharing it.

What to do

1. Start your model server first and note its API address, port and any API token.

2. Open Settings, expand AI & Models, and select Models & Accounts.

3. Find Add provider or node and select Local node.

4. Choose This Mac or Network under Where is it running, then choose your Runtime.

5. For a custom OpenAI-compatible server, choose vLLM / llama.cpp and enter the server port and API path, usually /v1; the name covers custom compatible endpoints too.

6. For Network, enter Host / IP; for This Mac the host is fixed to 127.0.0.1, then enter an API token if your server requires one.

7. Check the Endpoint preview, select Test response, then select Add and connect.

8. Return to a chat and select that provider and its model in the model picker.

What to do

1. Obtain an API key from your chosen provider using that provider's own documentation.

2. Open Settings → AI & Models → Models & Accounts.

3. In Add provider or node, choose Cloud and HTTP endpoint.

4. Choose Cloud provider, optionally give the account a Display name, and paste its API key into the password field.

5. Select Test response, read the result, then select Add and connect.

6. Return to your chat and choose the added provider and model.

What to do

1. Install the selected provider CLI using its official instructions if it is not already installed.

2. Open Settings → AI & Models → Models & Accounts and choose Cloud.

3. Under Provider type, choose Codex CLI or Claude CLI.

4. Choose Single under How many subscriptions and let TensorFold check the installed CLI.

5. If the CLI is detected and signed in, select Connect; if installed but signed out, select Add & sign in and complete the browser sign-in.

6. If TensorFold cannot find the CLI, install it or set its CLI binary in Advanced, then select Re-check.

7. Select the connected provider in a chat, room, bot or automation.

What to do

1. Switch Full workspace layout to Complete workspace if you want the project sidebar visible, then add your local project.

2. Load the local TensorFold model you want to use, then open Models → Workers.

3. Choose a Git project with an initial commit, enter an optional Task name and describe What should this worker do?.

4. Include how the result should be checked, or choose Use document task for the built-in documentation starter.

5. Choose Start worker and follow its state; open it when Needs you appears to review approvals.

6. When it finishes, open its changes for review; use Stop to cancel a running worker while preserving its worktree.

What to do

1. Open Bots in Full and choose New bot.

2. Enter a Name, Responsibility and Instructions that describe the job and when the bot should ask you.

3. Choose its Model, then select Read only or Review actions with workspace rules under Permissions.

4. Choose This Mac under Execution for the ordinary workspace, then choose Save bot.

5. Choose New chat on the bot card to start with those saved instructions, or Open chat to resume an existing conversation.

6. Use Edit for future conversations, Pause to prevent new work, or Archive to keep its history while removing it from the active list.

Run a bot in a separate Linux VM

This optional route requires Apple silicon, macOS 26 or later, at least 10 GiB of free disk space, and a supported model running on this Mac with Cortheon off.

1. In the bot editor, choose Separate Linux VM under Execution and select Save bot.

2. On the saved bot card, choose Set up bot VM and review the proposed setup.

3. Wait until the VM reports VM setup complete, then choose New chat.

4. Check the execution target before starting work; a failed VM does not silently fall back to your Mac.

What to do

1. Install the supported Nemotron 3.5 Lightning MLX 4-bit checkpoint through Models and finish active local work.

2. In Full, open Models → Learning.

3. Turn on Enable Living Weights and wait while TensorFold prepares the independent private copy.

4. If the panel says Private copy ready, choose Load learning model.

5. Enter a small, explicit statement under Teach a fact and choose Teach this fact.

6. Review the reported result and learning graph; use Stop lesson if needed and wait for the safe save boundary.

What to do

1. Open Models → Compute and clusters and choose Add host, or Discover hosts to look for enrollment suggestions.

2. Review the address and credentials, verify the SSH fingerprint when using SSH, and save the host.

3. Use the SSH setup controls only after checking prerequisites and reviewing the proposed per-user headless installation.

4. Choose Inspect to read the host's engine version, cached models, memory and actual supported profiles.

5. Choose New plan, select a ready host and model, enter its immutable revision, and choose the implemented execution profile and generation settings.

6. Choose Review plan, inspect admission results, then use the reviewed plan's staging and start actions; use the endpoint only after its authenticated canary succeeds.

What to do

1. Configure a source in Settings → Connectors, then open More → Assistant in Complete workspace.

2. Choose Sync to fetch updates, or Sync email for email; Refresh reloads the current app snapshot.

3. Select a connected source in the sidebar and a conversation in the Threads list.

4. Choose Brief me for a summary of the selected source or all attention items.

5. Enter an instruction and choose Ask TensorFold to create an assistant task; set its preferred model in Settings → Models & Accounts → Assistant model.

6. For a generated outgoing message, open its approval, verify the account, recipient and exact body, then choose Approve and send once only when it is correct.

What to do

1. Open More → Rooms, choose or create a project, then open its Rooms tab and choose New room.

2. Add an agent from Room agents, or use Recommended trio or Planner + builder + reviewer to start with roles.

3. Choose the provider and model for each seat and review its role; local builders and operators perform work, while external models follow planning restrictions.

4. Choose Collaboration and Deliberation settings, then set the Worker budget for Concurrent workers and Starts/run.

5. Enter a problem in the room composer and choose Start; resolve any attachment warnings first.

6. While the room runs, send text guidance with Steer room, answer its questions and review work; choose Stop to end the current run.

What to do

1. Open More → Automations and choose New automation.

2. Enter its Name and Instruction, then choose Manual, Interval or Daily under Schedule.

3. For Daily, set the clock time and Timezone and check Next scheduled time; Use this device's timezone fills the current zone.

4. Choose an Automation model, then select Chat for one completion, Agent for a tool task, or Digest for an activity brief without a model.

5. Choose Inbox as the delivery destination, or configure a supported external destination and recipient.

6. Choose Save, turn on Enabled for scheduled work, and use Run now to check the result; open History to inspect previous runs.

What to do

1. Open More → Command Centre in Complete workspace, or use Open full activity from Chat layout.

2. Choose Refresh and review attention items, Approvals and Running Work.

3. Open the underlying task before approving a tool call or reviewing a finished result.

4. Check Readiness Doctor for blocked providers, connectors or remote access, then choose Full doctor for setup details.

5. Use Sources to manage tool connections and Run Brief to run the configured brief automation.

6. For an interrupted job, inspect its current state and use only the recovery action offered for that job.

What to do

1. Open More → Command Centre and choose New mission in Mission Control.

2. Enter the Mission title, Task it repeats and a concrete Stop condition.

3. Choose the repeat mode and whether each run starts fresh, builds on the previous run or carries notes.

4. Set the role, fan-out, Max runs, cost cap and expiry appropriate to the task, then choose Create.

5. Use Start to run the mission and Inspect to read its checkpoints, journal and scratch notes.

6. Use Pause or Cancel when needed; mark Complete only after reviewing the actual result.

What to do

1. On the host Mac, open Settings → Orchestrator and enable Orchestrator server under Host this TensorFold as an Orchestrator.

2. Wait for Listening, then review the host URL and Network settings.

3. For a second machine, establish an SSH tunnel or a TLS front-end to the loopback server before pairing.

4. For another TensorFold app, choose Show pairing on the host, then on the client enable Connect to an Orchestrator, enter the reachable URL and token, and choose Test connection.

5. For a browser, name the device in Paired devices and choose New pairing link; open the link before its ten-minute expiry.

6. Choose a saved host as a project's run location or enable Run everything on an Orchestrator, then check the composer's Runs on indicator before starting new work.

What to do

1. Load the model and open its details. Under Capability checks, choose Text and Check selected capabilities, then read the result.

2. Open Settings → Models & Accounts → Optional helper models and select a qualified local model for Conversation naming, Document facts or Source answers.

3. Load the selected built-in helper model before running the job.

4. In an eligible local conversation, choose Local helpers above the chat.

5. Choose Conversation title, Extract document facts or Answer from local sources, then enter only the source text or question you want the helper to use.

6. Choose Run local helper and review the result; Copy result saves it to the clipboard, and a title result also offers Use as conversation title.

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Open Assistant → Connectors.

2. 2

   Google connects Gmail and Calendar.

3. 3

   Use the Slack or Discord card for a selected-channel bot.

4. 4

   Telegram uses your own bot and explicit allowlists.

Open full-resolution image ↗

Choose a connection

1. Open Connectors

   In Full mode, open Settings. Under Assistant, choose Connectors.

2. Pick one service

   Choose the service card you need, then follow its guide. Start with one connection.

Connect and check it

1. Choose its access

   Follow that service’s setup steps to add your credentials, select its accounts or channels, and save.

2. Check the connection

   Request a sync or test read in the service’s panel. Read the result before asking the Assistant to use it.

Before anything is sent

Review the exact action whenever a task will change an external service.

1. Review outgoing changes

   Before an email, reply or calendar change goes out, check the account, recipient and full content in its approval.

Pick your source

• Gmail and Calendar, using your own Google Desktop OAuth client.
• Slack or Discord, using a bot limited to selected channels.
• Telegram, with explicit sender and chat allowlists.
• WhatsApp, an experimental phone-linked connection.
• Documents and browser snapshots, using only the local sources you select.
• Email, iMessage and other workspaces, with service-specific credentials and scope rules.

Find it in the app

Click a screenshot to enlarge

Import clientChoose access

123Enlarge ↗

1. 1

   Create or select the Google Cloud project you control.

2. 2

   Enable the Gmail and Calendar APIs.

3. 3

   Import the Desktop OAuth client JSON you downloaded.

Open full-resolution image ↗

Prepare your Google client

1. Open Google setup

   Open Settings → Assistant → Connectors, then choose Google.

2. Create your Google project

   Choose Google Cloud setup. In Google Cloud, create or select your project and configure its OAuth consent audience.

3. Enable Gmail and Calendar

   Choose Enable APIs, then enable the Gmail API and Google Calendar API in that project.

4. Import your Desktop client

   Create a Desktop app OAuth client in Google Cloud and download its JSON. Back in TensorFold, choose Import Desktop client JSON and select that file.

5. Add yourself as a test user

   If the Google project is in Testing, add your Google account as a test user before signing in.

Connect your account

1. Choose the permissions

   Select Read Gmail and Read Calendar. Turn on Send reviewed emails or Apply reviewed calendar changes only if you need those actions.

2. Sign in with Google

   Choose Sign in with Google, complete consent in your browser, then return to TensorFold.

3. Check what Google granted

   Check the account shows Connected and read the permissions under Google granted before opening Gmail or Calendar.

Official setup links

• Google Cloud setup
• Enable Google APIs

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Choose whether to allow reading Gmail.

2. 2

   Choose whether to allow reading Calendar.

3. 3

   Sending emails and changing events are separate permissions.

4. 4

   Open Google consent after choosing the permissions.

Open full-resolution image ↗

Follow the steps

1. Open Settings → Assistant → Connectors → Google and select the connected Google account.

2. Choose Gmail, enter a query such as is:unread in Search Gmail, and select Search.

3. Use Write an email to compose a message; Save local draft stores it in TensorFold without sending it or creating a Gmail draft.

4. To send, enable Send reviewed emails for the account, select Review email before sending, check the full account, recipients and content, then choose Review and approve on this Mac.

5. For a schedule, choose Calendar, choose a calendar and set Calendar range start and Calendar range end.

6. Select View events or Check availability; use New event or Edit or cancel when your account has calendar write access.

7. Enter the event details, time zone, guests and Guest updates choice, then select Review new event, Review event changes or Review event cancellation.

8. Review the exact change and confirm it on the Mac; read Recent Google action outcomes to see the result.

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Enter your own Slack bot token.

2. 2

   Save and verify the token, then choose channels.

3. 3

   Only include channels you want the bot to access.

4. 4

   Connect after verification and channel selection.

Open full-resolution image ↗

Create the Slack bot

1. Open Slack setup

   Open Settings → Assistant → Connectors → Slack, then choose Create Slack app.

2. Set up the workspace bot

   Follow Permission guide to add the required bot scopes. Install the app in the workspace and invite its bot to the channels you want to use.

Connect it to TensorFold

1. Add its bot token

   Back in TensorFold, choose Add workspace bot and paste the token into Bot token.

2. Verify the bot

   Choose Save and verify token. When verification succeeds, choose Choose channels.

3. Choose the channels

   Select the channels to allow, or enter their exact IDs in Allowed channels.

4. Connect and read

   Choose Connect selected channels. Then choose Read selected channels and check the read result.

Stop access later

1. Stop access when needed

   Choose Disconnect to stop this saved connection, or Remove to delete its entry from TensorFold.

Official setup links

• Create Slack app
• Slack token documentation

Find it in the app

Click a screenshot to enlarge

1234Enlarge ↗

1. 1

   Enter the bot token from your Discord application.

2. 2

   Verify the token before choosing channels.

3. 3

   Choose the exact allowed channels.

4. 4

   Connect the selected channels when setup is ready.

Open full-resolution image ↗

Prepare your Discord bot

1. Open Discord setup

   Open Settings → Assistant → Connectors → Discord, then choose Discord Developer Portal.

2. Create a Discord bot

   Follow Discord’s setup instructions to create an application and bot. Copy the bot token.

3. Set its permissions

   Enable Message Content Intent if you need ordinary message text. Install the bot with the channel permissions listed below.

Connect and read

1. Add and verify the token

   In TensorFold, choose Add server bot. Paste Bot token, then choose Save and verify token.

2. Select channels

   Choose Choose channels, select the channels to allow, then choose Connect selected channels.

3. Read the selected channels

   Choose Read selected channels and check the connection status and read result.

Stop access later

1. Stop access when needed

   Choose Disconnect or Remove in TensorFold. Revoke the bot token through Discord if you also want to end provider-side access.

Official setup links

• Discord Developer Portal
• Discord bot permissions

Find it in the app

Click a screenshot to enlarge

Bot setupRemote access

1234Enlarge ↗

1. 1

   Enter your own BotFather token.

2. 2

   Add the exact chat IDs you allow.

3. 3

   Also add the exact sender IDs you trust.

4. 4

   Connect after verifying the token and completing both lists.

Open full-resolution image ↗

Connect your bot

1. Create your Telegram bot

   In Telegram, use verified @BotFather and /newbot. Then start a private chat with your new bot.

2. Verify its token

   Open Settings → Assistant → Connectors → Telegram. Paste Bot token and choose Save and verify token.

3. Allow trusted chats and people

   Enter Allowed chat IDs and Allowed sender IDs. A command must match both lists.

4. Connect the bot

   Choose Connect bot, then Check now. Turn on Check for new commands automatically if you want background polling.

Optional: share chats and rooms

1. Enable optional remote access

   Under Chief of staff and room access, turn on Enable Telegram remote access if you want the bot to use desktop chats or rooms.

2. Share selected chats or rooms

   Choose Only the chats and rooms I select, tick the targets to share, then choose Save remote access.

3. Choose a remote conversation

   In the private bot chat, use /chats or /rooms to list shared targets. Use /chat id or /room id to switch, and /chief to return.

Stop access later

1. Turn access off

   Choose Turn remote access off to revoke remote sessions. Choose Disconnect to stop the bot connection and polling.

Official setup links

• Open verified BotFather
• Telegram bot setup

What to do

1. On the host Mac, open Settings → Assistant → Connectors → WhatsApp.

2. Under WhatsApp · Link my phone, select Add account and then Link phone.

3. On the phone, open WhatsApp → Settings → Linked Devices → Link a Device and scan the QR code shown on the Mac.

4. Wait for Connected, then enable My own WhatsApp chat or enter international phone numbers under Selected chats.

5. Choose Save chats; groups require their exact WhatsApp group ID.

6. Use Disconnect to stop the local connection and Reconnect to restore a saved phone link.

7. To end the phone-side authorization, remove the linked device in WhatsApp as well.

What to do

1. For files, open Settings → Assistant → Connectors → Documents.

2. Choose Add folder and select the folder in the native picker.

3. Check that the source is enabled, then choose Sync documents.

4. Read Last scan to see ready, failed, unsupported and skipped counts before asking the assistant about those files.

5. For browser context, open Connectors → Browser and enter the tab URL, title and excerpt, then choose Add current tab.

6. Alternatively provide a local JSON history export and select Add history export.

7. Choose Sync browser, then use the saved sources in Assistant.

8. Disable a source or use its remove control when you want TensorFold to stop using that source.

What to do

1. Open Settings → Assistant → Connectors and choose the service.

2. For Email, select Add folder for EML, EMLX or Maildir files, or Add IMAP account and enter the provider's host, port, TLS, username, app password and mailbox.

3. For iMessage, enable iMessage connector, select Check access, and follow Open System Settings if Full Disk Access is required; restart TensorFold after granting it.

4. For Mattermost, use Add server, enter Server URL and Access token, and set Allowed channels before Verify token and Sync Mattermost.

5. For Matrix, choose Add account, enter Homeserver URL and Access token, and set Allowed rooms before Verify token and Sync Matrix.

6. For Google Chat, choose Add space, enter the Webhook URL for sending and a separate Access token if you need reading, then set Allowed spaces and select Verify.

7. Check the saved source and scope before choosing Sync email, Sync iMessage or the service's sync button.

8. Turn off the source or remove it in TensorFold when finished; revoke its credentials at the service when you also want provider-side access removed.

What to do

1. Open Plugins and choose Built-in to search the available tool cards.

2. Use Check setup on a card and read its Available, Disabled or Worker required status.

3. Choose Connections to open a service's Set up or Manage flow, or Installed to inspect configured scripts and plugins.

4. For an MCP tool server, choose Advanced → Add server.

5. Enter Name, Command, Arguments and Working directory from that server's own setup instructions.

6. Choose the intended launch mode, keep approval mode Ask for reviewed actions, and select Save server.

7. Select Connect, then inspect the connected server's tools, resources and prompts.

8. Use Stop to disconnect a running server, or Delete to remove its configuration.

What to do

1. In a chat, enter /computer-use followed by the task, or mention @computer with your request.

2. Grant TensorFold macOS Accessibility permission when prompted.

3. Review the requested app access and approve the app for that task.

4. Watch the control indicator at the top of the screen while TensorFold opens the app, reads controls and performs the requested actions.

5. Take over for passwords or actions the app cannot safely identify.

6. Select Stop on the indicator or send /computer-use stop when you want control to end.

What to do

1. Start a chat that contains no private phone or Google account history.

2. Ask a specific public-news question or provide the public page URL you want read.

3. Let the model call the available web tools and review the links returned with the answer.

4. If the page cannot be fetched, check it in a browser instead of assuming the model saw it.

What to do

1. For Android USB, install Android Platform Tools, enable Developer options and USB debugging, attach a data cable and approve the computer on the phone.

2. Open Settings → Assistant → Phone control and select the adb executable if it was not detected.

3. Open Phone beside the chat model picker, refresh devices and connect the intended phone.

4. Choose a model running on this Mac and select Allow this chat.

5. Review each requested device action; take over for authentication, payments, sending or other committing steps.

6. Select Stop to cancel work and clear the chat's control grant and preview.

Official setup links

• Android Platform Tools
• Apple iPhone Mirroring requirements
• WebDriverAgent real-device setup

What to do

1. On an Apple Silicon Mac, install the Homebrew formula below, or upgrade your existing installation.

2. For Linux, download the release archive for your CPU architecture and its companion SHA-256 file, verify the checksum, and unpack it.

3. Keep the archive's bin, lib and share directories together, then add its bin directory to your PATH or call bin/tensorfold-native directly.

4. Run --version and capabilities --json; the latter reports the families and flags compiled into your installed build.

brew install ashhart/tensorfold/tensorfold
tensorfold --version
tensorfold capabilities --json

brew update
brew upgrade tensorfold
tensorfold --version

V=1.1.0
P=linux-x86_64
curl -fLO "https://github.com/ashhart/TensorFold/releases/download/v$V/tensorfold-$V-$P.tar.gz" &&
curl -fLO "https://github.com/ashhart/TensorFold/releases/download/v$V/tensorfold-$V-$P.tar.gz.sha256" &&
sha256sum -c "tensorfold-$V-$P.tar.gz.sha256" &&
tar -xzf "tensorfold-$V-$P.tar.gz" &&
cd "tensorfold-$V-$P" &&
bin/tensorfold-native --version &&
bin/tensorfold-native capabilities --json

What about the curl installer?

The website also offers the command below. When this guide was checked, the engine 1.1.0 release did not contain the .sha256.sig files that this installer requires, so a signature download can fail. Use the Homebrew or manual archive route above for this release; do not disable the installer verification.

curl -fsSL https://tensorfold.dev/install.sh | sh

Linux ARM64

Use the Linux archive commands above with P=linux-aarch64 instead of P=linux-x86_64. Run the checksum check successfully before extracting, then keep bin, lib and share together.

Make the archive command available

From inside the extracted release directory, the command below adds its bin directory to your current terminal session. Archive users should replace tensorfold with tensorfold-native in the rest of these examples.

export PATH="$PWD/bin:$PATH"
tensorfold-native --version

What to do

1. Check your build with tensorfold capabilities --json and use a native recipe for the exact model, checkpoint format and hardware.

2. Run tensorfold pull with one Hugging Face repository ID; rerun the same command if an interrupted download needs to resume.

3. List cached models with tensorfold models, then inspect the chosen checkpoint with tensorfold info.

4. For Qwen3.8-27B, pull the DFlash2 drafter separately and pass it when serving.

tensorfold pull mlx-community/Qwen3.5-2B-MLX-4bit
tensorfold models
tensorfold info mlx-community/Qwen3.5-2B-MLX-4bit

tensorfold pull TensorFold/Qwen3.8-27B-MLX-4bit
tensorfold pull z-lab/Qwen3.8-27B-DFlash2
tensorfold info TensorFold/Qwen3.8-27B-MLX-4bit

tensorfold pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit
tensorfold info TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit

What to do

1. Pull the model first, then start serve in a terminal and leave it running.

2. Use --name to give clients a short stable model ID; the example calls it local-model.

3. Wait until startup completes, then check /health and /v1/models from another terminal.

4. Send the chat request below; read the reply in choices[0].message.content.

5. Use Ctrl-C in the server terminal to stop it; start it again with different flags to change launch settings.

tensorfold serve mlx-community/Qwen3.5-2B-MLX-4bit \
  --name local-model --parallel 1 --context 8192 --temperature 0 --no-thinking --dashboard

curl -fsS http://127.0.0.1:8080/health
curl -fsS http://127.0.0.1:8080/v1/models

curl -fsS http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"local-model","messages":[{"role":"user","content":"Say hello in one sentence."}],"max_tokens":128,"temperature":0}'

What to do

1. Start the engine with --name local-model and check the exact advertised ID in /v1/models.

2. For an OpenAI-compatible client, set the base URL to http://127.0.0.1:8080/v1 and the model to local-model.

3. For an Anthropic Messages client, use http://127.0.0.1:8080 as the base URL; the route is /v1/messages.

4. If another machine needs access, create a restricted API-key file, bind to the intended interface and configure the client with that key.

ANTHROPIC_BASE_URL=http://127.0.0.1:8080 ANTHROPIC_API_KEY=local claude --model local-model

mkdir -p "$HOME/.config/tensorfold"
(umask 077; openssl rand -hex 24 > "$HOME/.config/tensorfold/api-key")

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit \
  --name local-model --host 0.0.0.0 --port 8080 \
  --api-key-file "$HOME/.config/tensorfold/api-key"

What to do

1. Use --context for the total prompt-plus-reply window and --max-tokens for the default reply limit.

2. Keep the checkpoint's generation defaults first; add --temperature, --top-p, --top-k or --min-p when you want an explicit policy.

3. Use --thinking or --no-thinking to choose the model's reasoning mode where its template supports it.

4. Increase --parallel only after the startup memory plan leaves enough room and your model supports shared requests.

5. Check the effective context in /v1/models and the actual allocation in /memory.

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --name local-model \
  --context 32768 --max-tokens 1024 --parallel 2 \
  --temperature 0.7 --top-p 0.95 --top-k 20 --no-thinking

curl -fsS http://127.0.0.1:8080/v1/models
curl -fsS http://127.0.0.1:8080/memory

tensorfold capabilities --json
tensorfold serve MODEL --help

What to do

1. Pull the public checkpoint before starting; the benchmark never downloads a model.

2. Run tensorfold benchmark with the checkpoint and an output filename.

3. Inspect the saved receipt and check that every measured attempt completed.

4. Open the benchmark board, choose Upload result, pick a public display name and select the JSON file.

5. Review the public fields, then confirm the upload; incomplete or nonstandard runs are not automatically ranked.

tensorfold benchmark TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --output benchmark.json
tensorfold benchmark inspect benchmark.json

tensorfold benchmark suites

tensorfold benchmark TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --reps 1 --tokens 64 --temperatures 0

tensorfold benchmark publish benchmark.json

What to do

1. Use a complete checkpoint with the required MTP head for families that draft internally.

2. For Qwen3.8-27B, download the target and DFlash2 separately, then pass --drafter explicitly.

3. Keep the default 4-bit drafter preparation unless the model recipe calls for another setting.

4. For a serial comparison, restart with --no-drafts or send "draft":false for one request.

tensorfold pull TensorFold/Qwen3.8-27B-MLX-4bit
tensorfold pull z-lab/Qwen3.8-27B-DFlash2
tensorfold serve TensorFold/Qwen3.8-27B-MLX-4bit --name local-model \
  --drafter z-lab/Qwen3.8-27B-DFlash2 --context 32768

tensorfold serve TensorFold/Qwen3.8-27B-MLX-4bit --name local-model --no-drafts --context 32768

What to do

1. Use the same model and unchanged conversation prefix to let the normal in-memory prompt cache resume a later turn.

2. Leave the native cache planner in charge first; use --prompt-cache-gib only when you have a reason to choose a budget.

3. On supported Metal families, add --learn to retain shared system prompts and tools on disk.

4. Use --learn-conversations instead when you want each retained conversation state written to disk for restart/resume.

5. Stop with Ctrl-C or SIGTERM so queued conversation writes can finish.

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --name local-model --context 8192 --learn

tensorfold serve TensorFold/Qwen3.8-27B-MLX-4bit --name local-model \
  --drafter z-lab/Qwen3.8-27B-DFlash2 --learn-conversations --learn-gib 32 --learn-daily-gib 32

curl -fsS http://127.0.0.1:8080/memory

What to do

1. Add --recall to a supported serving setup.

2. Use the same tokenizer identity for servers that should share a record.

3. Read /v1/recall to check record size and how many copied tokens landed.

4. For comparisons, keep Recall off or use an empty record in read-only mode.

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --name local-model --recall

curl -fsS http://127.0.0.1:8080/v1/recall

What to do

1. Choose a context window that fits your hardware.

2. Add --compact-at auto to enable compaction when the prompt and reserved reply would exceed available context.

3. Use --compact-keep to control how much recent conversation stays verbatim.

4. If your client handles the compaction extension, send back the compacted conversation and preserve its memory note.

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --name local-model \
  --context 32768 --compact-at auto --compact-keep 8192

What to do

1. Use a complete native vision checkpoint with its processor files and image tower.

2. Start it with --vision, allowing memory for both the text model and image encoder.

3. Send a PNG or JPEG as an OpenAI image_url data URL alongside a text question.

4. Use --vision-max-images for the history's image count and --vision-image-tokens for their shared visual-token budget.

tensorfold serve TensorFold/Qwen3.8-27B-MLX-4bit --name local-model \
  --drafter z-lab/Qwen3.8-27B-DFlash2 --vision --context 32768

python3 - <<'PYIMAGE'
import base64, json, urllib.request
from pathlib import Path
image = base64.b64encode(Path('photo.png').read_bytes()).decode()
body = {
  'model': 'local-model',
  'messages': [{'role': 'user', 'content': [
    {'type': 'text', 'text': 'Describe this image.'},
    {'type': 'image_url', 'image_url': {'url': 'data:image/png;base64,' + image}}
  ]}],
  'max_tokens': 256
}
request = urllib.request.Request('http://127.0.0.1:8080/v1/chat/completions',
  data=json.dumps(body).encode(), headers={'Content-Type': 'application/json'})
with urllib.request.urlopen(request) as response:
  print(json.load(response)['choices'][0]['message']['content'])
PYIMAGE

What to do

1. Start a supported model with --cortheon.

2. Keep your client's usual local model and API base URL.

3. Ask a question that needs current information; the server can search, fetch pages and feed the evidence into its reasoning steps.

4. Set BRAVE_SEARCH_API_KEY if you want Brave search; without it, Cortheon uses DuckDuckGo.

tensorfold serve TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit --name local-model --cortheon

What to do

1. Use the Nemotron 3.5 Lightning MLX 4-bit checkpoint on a qualified Mac or GB10, with disk space for a separate teaching copy.

2. Copy the real model files into a new folder rather than teaching the shared Hugging Face snapshot.

3. Create a private API-key file and start the copy with --slide and --slide-graph.

4. POST a fact to /v1/slide/learn and read the streamed result to see whether the lesson was retained.

5. To undo all lessons, stop the server and run slide rollback against that teaching folder.

mkdir -p "$HOME/models"
cp -cRL "/absolute/path/to/the/exact/snapshot" "$HOME/models/nemotron-living"

(umask 077; openssl rand -hex 24 > "$HOME/models/slide-key")
tensorfold serve "$HOME/models/nemotron-living" --slide \
  --api-key-file "$HOME/models/slide-key" \
  --slide-graph "$HOME/models/nemotron-living-facts.json"

curl -N http://127.0.0.1:8080/v1/slide/learn \
  -H "Authorization: Bearer $(cat "$HOME/models/slide-key")" \
  -H 'Content-Type: application/json' \
  -d '{"text":"My preferred colour is blue."}'

tensorfold slide rollback "$HOME/models/nemotron-living"

What to do

1. For paired Mac Flash Next or GLM, use the same engine release and checkpoint on both machines.

2. Set up MCDMA and the Thunderbolt link using its fabric guide, then create one rank settings file per Mac.

3. Start rank 1 first and rank 0 second; send API requests only to rank 0.

4. For CUDA Nemotron Super VL on two GB10s, use its NCCL-based native recipe instead of the Mac --speed-up configuration.

5. Check the model-specific memory and failure-recovery notes before increasing context or concurrency.

tensorfold serve "$HOME/models/flash-next-6bit" --name flash-next \
  --speed-up rank1.json --context 32768 --temperature 0 --no-thinking

tensorfold serve "$HOME/models/flash-next-6bit" --name flash-next \
  --speed-up rank0.json --context 32768 --temperature 0 --no-thinking

What to do

1. Open Settings → Readiness, or Command Centre → Readiness Doctor → Full doctor, and inspect Blocked or Needs review entries.

2. If a model cannot load, open its details and check the memory verdict, supported format and missing-file reason; try a smaller compatible model when it needs more memory.

3. If a download failed, open Models → Downloads and use the displayed retry action; for gated repositories, download through your own authenticated tool and use Model folders.

4. If model settings or comparisons are disabled, finish active local chat, API, worker or learning work before trying again.

5. If another client lost its connection, load the model and recopy the URL and API key from API connection after the server restart.

6. If a draft could not save, use Retry save and keep the draft visible until the app reports it is saved.

What to do

1. Run tensorfold --version and tensorfold capabilities --json; confirm you are using the native 1.1.0 engine and the flags it actually supports.

2. If a model cannot be found, run tensorfold models and tensorfold info with the same repository ID, then pull the missing checkpoint or use its complete local directory.

3. If startup refuses memory, read the requested and available bytes in the error, close other large workloads and try fewer streams, a shorter context or a smaller supported checkpoint.

4. If a client cannot connect, check the serving terminal, /health, the port, the model ID and its configured API key.

5. For a report, include the engine version, public checkpoint ID and revision, exact command, a minimal public request and the relevant redacted error.

tensorfold --version
tensorfold capabilities --json
tensorfold --help

tensorfold models
tensorfold info mlx-community/Qwen3.5-2B-MLX-4bit

tensorfold serve mlx-community/Qwen3.5-2B-MLX-4bit \
  --name local-model --port 8081 --parallel 1 --context 8192 \
  --temperature 0 --no-thinking --dashboard

curl -fsS http://127.0.0.1:8081/health
curl -fsS http://127.0.0.1:8081/v1/models