Get things done with Sessions AI

Set up for a class or project

If you're a student, create one persona per class and load its reading into a knowledge base. Every chat with that persona can then draw on the material — without anything being uploaded.

1. 1
   Create a persona for the class

   In Settings → Personas, add a persona named for the course (e.g. “Biology 101”). Give it a system prompt like “You are my biology study partner.”

2. 2
   Add the reading to its knowledge base

   Open the persona's Knowledge base and add your textbooks and notes — pick local files, or paste a URL to fetch and convert a reading on-device.

3. 3
   Start a chat with that persona

   New chats use the active persona, so the model can search that material as it answers — “Quiz me on chapter 3” or “Explain meiosis using my notes.”

4. 4
   Attach anything one-off

   For a single assignment, attach the PDF directly to the chat instead of adding it to the knowledge base. Attachments live with that chat only.

Have a conversation

The fastest way to start. Pick a model once, then chat as much as you like.

1. 1
   Choose and download a model

   In Settings, pick a model from the catalog. Smaller models respond faster; larger ones handle more complex tasks but need more storage.

2. 2
   Start a new chat and send

   Type your message and send. The model runs on your device, so the first reply may take a moment as it loads.

3. 3
   Regenerate or cancel anytime

   Not happy with a reply? Regenerate to try again. Mid-generation, tap Stop — the partial reply is discarded so you can edit and resend.

Ground answers in your own documents

Add documents to a knowledge base and the assistant searches them locally, then grounds its answer in what it finds. Nothing is uploaded.

1. 1
   Add files or fetch a URL

   In the persona's Knowledge base, add text, Markdown, or PDF files — or paste a URL and Sessions will fetch and convert it on-device.

2. 2
   Sessions indexes it locally

   Documents are chunked and indexed with on-device full-text search. Saving re-chunks and re-embeds the document.

3. 3
   Ask, and check what was used

   Relevant chunks are added to the prompt as context. “Last retrieval” in Settings shows which of your documents informed the most recent answer.

Take action with tools

Ask in plain language and the model reaches for on-device tools — no setup required. A few things to try:

Calendar, reminders, and contacts use your device's own data with your permission. See the tools reference for the full list.

Create images

Switch the composer to Image mode and describe what you want. Generation runs locally — your prompts and pictures never reach a server.

1. 1
   Download an image model

   In the model catalog, choose an image model (e.g. SDXL). Image models are larger, so they need more storage and memory.

2. 2
   Switch to Image mode and describe it

   Toggle the composer to Image, then describe your picture — “A watercolor sunset over the bay.” Tune size, steps, and seed in generation defaults.

Models

Sessions runs open language and image models locally. Browse the in-app catalog (powered by Hugging Face), download what you need, and remove models to free space. See the full model catalog.

• Text models — pick by size and capability. Smaller is faster; larger handles complex reasoning.
• Image models — for Image mode. Larger downloads, more memory needed.
• Storage — models range from hundreds of MB to several GB. Keep only the ones you use.

Personas

A persona is an assistant identity with its own setup. Switch the active persona to change which one new chats use.

Knowledge base

Per-persona document storage that grounds answers in your content. Add text, Markdown, or PDF files, or fetch a URL — everything is chunked and indexed on-device with full-text search. Nothing is uploaded.

Tools reference

MCP servers

Model Context Protocol lets you connect third-party servers that expose extra tools — a database query or code-execution tool, for example. Add MCP server URLs per-persona in Settings; Sessions connects over HTTP and registers their tools.

Settings & sampling

• Max tokens — the maximum length of each reply.
• Temperature — higher is more creative; lower is more focused.
• Advanced sampling — top-k, top-p, min-p, typical-p, repetition penalty, mirostat, and more.
• Web search toggle — turn DuckDuckGo search on or off.

Privacy & data