Sorana: Desktop AI Knowledge Workspace & Agent Harness That Never Forgets

Name: Sorana
Author: Tetramatrix

Sorana is a personal desktop AI knowledge workspace and agent harness: it learns from your documents, conversations, and actions and uses that memory to chat, dispatch AI agents, manage email and calendar, browse the web, and run local models all in one private workspace. Give it a goal and the harness plans the steps, dispatches specialized subagents, acts with 90+ tools, and reviews its own work. It runs entirely on your machine and never forgets a project, preference, or decision.

Start for $0, OpenRouter's free tier gives you a free API key (no subscription, no credits) and auto-picks the best free model for every request. Pair it with Sorana and put the agentic AI to work before adding anything else.

What Is Sorana?

Sorana is a personal desktop AI knowledge workspace and agent harness for Windows, Linux, and MacOS. It learns from your documents, conversations, and actions, building a lasting memory of your projects, files, and thinking style. Open it tomorrow and it already knows where you left off. When a task needs it, the harness dispatches specialized subagents to work in parallel then reviews the results before answering. Unlike chatbots that answer one question and stop, Sorana follows through on goals.

Visualize your entire knowledge base on a spatial 2D canvas. Chat with any document in plain language. Let your AI handle repetitive tasks. Organize files, research topics, manage emails and your calendar, all without leaving your workspace.

Everything runs on your machine. Your data stays yours.

Who Is Sorana For?

Researchers & academics, Chat with hundreds of papers, books, and scanned notes in one place.
Writers & journalists, Build a living reference library that remembers your style, sources, and ideas.
Consultants & freelancers, Maintain full AI context across multiple long-running client projects.
Students, Turn notes, PDFs, and lecture slides into a personal knowledge base you can talk to.
Job hunters, Research companies, tailor CVs and cover letters, draft applications, send from Gmail, and add interviews to your calendar.
Household & personal admin, Manage landlords, energy, telecom, repairs, doctors, and banking. Keep sensitive letters, bills, and health records off the cloud by switching to private, local models.
Obsidian & Notion users, Add an AI layer that learns from your vault and can act on it.
Privacy-conscious professionals, Full AI power with zero data leaving your machine.

Why Sorana?

Your AI can't think strategically

Most AI does one thing at a time. Sorana breaks a goal into steps and dispatches specialized subagents, research, code, fact-check, to work parts in parallel, then reviews the results.

Multi-step planning for complex questions
Parallel searches, file reads, and web fetches
Self-review and automatic fallback strategies
Reviews its own output before answering

You re-explain yourself every day

Sorana quietly learns from every conversation, project, and preference. Next time you open it, you pick up exactly where you left off.

Proactive context from recent work
Persistent personal profile across the workspace
No setup, no copy-pasting

Your AI talks but can't act

Tell Sorana what to do in plain language. It organizes files, searches the web, and manages Gmail and Google Calendar, all from one conversation.

Connect Google Drive, GitHub, Gmail, read, write, search, and manage from chat
Read, write, move, search, edit via chat
Web search that shows its sources

Your knowledge lives on someone else's server

Sorana runs AI models locally. Your notes, files, and conversations never leave your machine unless you want them to.

Lemonade, Ollama, Llama.cpp, LM Studio
Portable: runs from a USB stick
Cloud only when you choose

Long conversations get slow and expensive

Auto-compaction compresses old conversations into tight summaries, cutting token use by over 90% while preserving full context.

Dozens of messages condensed into a short paragraph
Smart document search across your archive
Stays fast after months of use

Half your knowledge is locked in PDFs and images

Built-in OCR and vision AI make scanned documents, screenshots, and images searchable and chattable.

Extract text from PDFs and images
Vision results become searchable documents
JPEG, PNG, GIF, WEBP, BMP

Your powerful model is wasted on simple tasks

When your AI plans strategically, reviews its own answers, fetches web results, and routes between tools, every internal step uses a model. Sorana uses a fast model for these internal operations and reserves your primary model for the actual conversation.

Fast model for internal operations like tool calls and routing
Primary model for user conversations
Dedicated embedding model for RAG and semantic search
Hardware acceleration when available

Your AI is a black box

You ask a question and get an answer, but you have no idea how it got there. Did it search your files? Read the right documents? Did it check its work or just guess? Sorana shows you every step it took: what it searched, what it found, why it chose that answer, and when it changed its mind.

Step by step trace of searches, reads, and tool calls
See what the AI considered and what it discarded
Full reasoning shown before every answer
Understand and trust the result, or tweak the approach

Core Features

A workspace that thinks for itself

Sorana breaks complex requests into step-by-step plans, runs searches and file reads in parallel, reviews its own output, and switches tactics when it gets stuck. When a task needs specialized help, the workspace dispatches subagents then assembles and reviews the results.

Projects Canvas

A second canvas beside your Workspace board. Every project gets its own card, its own assistant, its own folder that Sorana keeps and looks after, and its own notes. Right-click the canvas to start a project, right-click a card to open or chat with it. Each canvas keeps its own view, zoom, pan, grid density, and remembered positions, and each one comes back exactly the way you left it.

Agent Budgets & Limits

Set a daily usage limit per agent from the Health Dashboard through a guided wizard, and track it in real time. Leave the field blank to use the app default or set 0 for no per-agent limit. Usage refreshes automatically and spend resets at local midnight, and withheld calls are reported honestly rather than silently truncated.

Local & Global Memory

A living knowledge profile captures facts and project context. Remembers what matters, spots contradictions, and scores importance automatically. Maintain both scoped local memories and a persistent global memory, and sync them anytime. When you decline a proposed action, the agent remembers the pattern and adapts future suggestions to your preferences.

Visual Knowledge Canvas

See your files arranged by topic on a 2D canvas. Drag, drop, resize, and generate mind maps to reveal hidden relationships.

Chat with Documents & OCR

Extract text from PDFs, documents, and images with built-in OCR. Chat with any document to get answers and make your whole archive searchable.

RAG Knowledge Base

Turn your documents, notes, and web pages into a searchable knowledge base. Sorana retrieves the right context automatically so answers stay grounded in your own data. Every answer lists the sources it used, filtered by relevance so you only see what mattered.

Supercharged Chat & MCP Tools

Your AI brain connected to your knowledge and everything you can act on: web, email, calendar, files, and any other tool you connect. You stay in control while your agent works across all of them at once. Swap agent personas on the fly with the Prompt Manager. Every session is archived with full Markdown, reasoning, and tables, so you never lose context.

Calendar & Scheduling

Connect Google Calendar and manage your schedule from chat. Create events, find free slots, and let Sorana plan your day.

Model Manager

Browse, expand, and collapse models from 25+ providers and local sources. Assign any model to five role slots: Primary, Quick, Embed, Vision, and Reviewer then set a Backup or Spare per role for automatic failover. Mark favorites, sort by any column, and probe capabilities before you commit.

Specialized Subagents

When a task needs it, Sorana dispatches specialized subagents, web research, code analysis, and fact-check to work in parallel with their own budgets, then reviews the results before answering.

Five Model Roles + Failover

Assign any model to five role slots: Primary, Quick, Embed, Vision, and Reviewer each optimized for its task. The Reviewer gives answer quality checks an independent second opinion on a different model. Set a Backup model per role and one global Spare so a dead provider never interrupts your work; Sorana hops to healthy alternatives automatically.

Model Probing

One click tests if a model can reason, see images, or handle audio, so you know what it can do before relying on it.

Prompt Manager

Create, organize, and assign prompt pairs to any agent from a central library. 20+ categories, one-click assignment, no file editing.

MCP Manager & Marketplace

Browse, install, and manage MCP servers from a central marketplace. Extend Sorana with new tools and integrations without touching configuration files. See the MCP Servers section below for what ships today.

Health Dashboard

Monitor system health, model status, and AI performance at a glance. Each card opens a mini wizard for guided installation and configuration of every dependency, including backends, OCR, search engines, HuggingFace cache, codebase search, MCP, and more.

Obsidian Export

Use Sorana as your AI workspace, then export spatial groupings and graph relationships straight into your Obsidian vault.

Agent Trace & Debug

Watch your AI think step by step. See what it searched, what it found, and why it chose each answer. Helps you trust the result or tweak the approach.

Code Explorer

Got a folder of code? Sorana reads it, searches it, and answers questions about it, like a teammate who already knows the project.

Tesseract Image OCR

Use Tesseract for image OCR on scanned photos, screenshots, and other images. When a vision-capable model is available, it can read the image directly as a fallback.

Examples in Action

Real things you can say to Sorana. No special syntax, no prompts to craft.

Memory & Knowledge

"Remember that my preferred working hours are 9am to 4pm"
"What did I learn about project budgeting last week?"
"Research electric car trends and save the key facts so I can find them later"
"Read this research paper and save the key findings to my notes"
"Compare what I know about option A versus option B from past research"

Research & Writing

"Summarize this document for me"
"Read these three articles and draft a summary memo, then save it to my project folder"
"Find relevant studies on this topic and add them to my knowledge base"
"Search my notes for the client meeting summary and send a follow up email"
"Organize this folder of research papers into topic groups and summarize each one"

Tasks & Scheduling

"Send an email to the team about the project update"
"What meetings do I have today?"
"Review the project brief, draft a response, and add the deadline to my calendar"
"Find free slots this week for a team review"
"Remind me about the invoice deadline next Friday"

Job Hunting

"Research this company's recent projects and leadership"
"Draft a cover letter using my CV and the job description I saved"
"Review my application and suggest improvements"
"Send a follow up email to the hiring manager"
"Add the interview date to my calendar and set a reminder"

Email & Calendar

"Show me unread emails from today"
"Draft a reply to this email about the project timeline"
"Schedule a meeting with the team next Tuesday at 2pm"
"Find my dentist appointment this month"
"Send meeting notes to everyone who attended"

Household & Personal

"What bills are due this month?"
"Compare internet plans from local providers and save the best option"
"Scan this letter and extract the important details to my knowledge base"
"Search my archives for the insurance policy and summarize the coverage"

Crypto & Market Data

"What is bitcoin's price in USD right now?"
"What is Bitcoin Cash trading at right now?"
"Show ethereum's 24h change and market cap"
"Compare bitcoin and ethereum by market cap"
"List the top 10 coins by market cap"
"Show bitcoin's price over the last 30 days"

Workflows & Pipelines

"Draft a project report, review it for gaps, and polish the final version"
"Research this topic, write a blog post, and have a second agent fact check every claim"
"Read these requirements, draft a technical spec, and review it for clarity"
"Analyze this customer feedback, summarize the key themes, and draft a response plan"
"Review this code change for bugs, suggest improvements, and generate a summary"

How Sorana Compares

What you get	Sorana	Obsidian / Notion	ChatGPT / Claude / others
Runs fully offline & private	Your machine, your data. No internet needed.	Obsidian local / Notion cloud	Cloud only. Your data on their servers.
Model flexibility	25+ providers + capability probing	N/A	Locked to one vendor
Five Model Roles + Failover	Assign any model to five role slots (Primary, Quick, Embed, Vision, Reviewer). Set a Backup or Spare per role for automatic failover.	N/A	One model for everything
Multi-agent orchestration	Specialized subagents work in parallel with their own budgets when the workspace needs them.	Simple automations only	Basic workflows / skills (paid)
Agent Trace & Debug	Watch your AI think step by step. See what it searched, what it found, and why it chose each answer.	No AI agent transparency	No agent tracing available
Sources you can verify	Every answer lists the web pages and documents it used, filtered by relevance.	N/A	Basic citations on paid tier only
Portable	Runs from a USB stick	App install required	Cloud / web only
AI takes action	Built-in automation on files, web, email	View & edit only	Paid connectors with separate setup
Goal-driven autonomy	Plans multi-step work, dispatches subagents when needed, acts with tools, and reviews its own work before answering	No autonomous agent	Agent mode / workflows (paid)
Visual knowledge canvas	AI-organised 2D canvas	Manual (Obsidian Canvas)	Chat only
Chat with documents	Workspace-wide RAG indexing	Plugin required	Project uploads + per-session
Image OCR & Vision AI	Built-in OCR + AI fallback	External tools required	Cloud-only vision
Tool integrations	Built-in MCP servers incl. ListenBrainz music history, plus ToolForest with 30+ toolkits over one connection and keyless CoinGecko crypto market data	No native AI agents	Connector marketplaces (paid)
Calendar & Scheduling	Manage via chat	View-only plugins	Connector (paid)
AI remembers you	Automatic memory across sessions, all categories free	Manual notes only	Basic free notes only; full memory paid
Stays fast over months	Auto-compaction	No AI context	Varies (Claude yes, others no)
Cost	Free app today (planned premium optional) + OpenRouter's free tier or your API keys (or free local)	Free + paid tiers	$8-200/month included

Quick Start

Download for Windows, Linux, or MacOS
Run anywhere, desktop, USB, external drive
Choose a model, use a free cloud model with your own API key (OpenRouter's free tier costs nothing), or run local models through Lemonade, Ollama, or LM Studio
Start working, add your files, chat with documents, and let Sorana remember your projects and preferences

Download: Sorana.exe · Sorana.zip (MacOS)
Microsoft Store: Get from Store

System Requirements

Component	Minimum
OS	Windows 11 (64-bit), modern Linux (Ubuntu 20.04+, glibc-based), or MacOS
RAM	4 GB (8 GB+ recommended for local AI models)
Storage	2 GB (app + model) + `.sorana/` data folder
AI	Built-in offline model included; local or cloud AI optional

AI Model Options

Built-in model (~806 MB), works fully offline out of the box
Recommended: Llama 3.1 8B Instruct or higher for best results (requires 8 GB RAM / 4 GB VRAM)
Cloud: OpenAI, Mistral, Gemini, DeepSeek, NVIDIA, OpenRouter (free tiers available), no local hardware required
Local servers: Lemonade (NPU acceleration), Ollama, LM Studio, Llama.cpp

30+ free cloud models from every major provider, plus unlimited local models. Just add your API key.

Obsidian Integration

Sorana connects directly to your Obsidian vault:

Export to Canvas, bring your AI-organized workspace layout into Obsidian's infinite Canvas
Export to Graph View, generate a knowledge graph from your workspace structure, compatible with Obsidian's Graph View
Full compatibility with Dataview, Templater, Excalidraw, and all other Obsidian plugins
Standard Markdown files, no lock-in

MCP Servers

Sorana reaches tools over MCP on three paths, all managed from the MCP Manager with no config files to hand-edit.

Built-in servers

Seven servers ship with the app. Most are enabled by default, and each can be switched per agent.

Server	What it gives the agent
Memory	Long-term memory storage, search, consolidation, and export
Filesystem	Read, write, search, and move local files and folders
Web search	Search the web and read the pages it finds
Gmail	List, read, send, label, trash, and archive email
Calendar	Read schedules, create and update events, find free slots
Chat sessions	Inspect and manage session messages, context usage, and compaction
ListenBrainz	Your music listening history (new, see below)

ListenBrainz (built-in, new)

Ask your assistant about years of music listening in plain language. Twelve tools cover public charts, your own listening history, and a full local archive of everything you have ever played.

Public charts and top artists — the sitewide chart plus your own most played artists over a week, month, year or all time
Now playing and recent listens — what is playing right now, and your last 100 listens with time bounds and paging
Full history export — ListenBrainz only releases a complete history through its export API, so Sorana queues the export as a background job, downloads it, and imports it locally. The import resumes automatically if it is interrupted, never double counts, and keeps running while you keep working
Queries across the imported archive — top artists over any date range, full history for one artist, and a rediscovery tool that finds artists you played heavily years ago and have not played since

Setup is one step: paste your ListenBrainz user token (found in your ListenBrainz profile settings) into the Configure dialog. Ask things like "my top 50 artists this year", "what am I playing right now", "export my full history", or "which artists did I love years ago but never play anymore".

CoinGecko (marketplace, new)

Ask for crypto prices in plain language. Every price comes straight from CoinGecko's API, so it is today's number rather than one the model remembered. The endpoint needs no account and no API key, so you can add it from the marketplace and start asking straight away.

Live prices — the current price of any coin in any supported currency
Market cap and volume — market capitalisation, 24 hour change and trading volume, plus the top coins by market cap or by volume
Price history — how a coin's price moved over a period such as the last 30 days
Coin metadata — name, symbol, rank and related details

Setup is one step: add CoinGecko from the marketplace and enable it for the agent. Ask things like "What is bitcoin's price in USD right now?", "What is Bitcoin Cash trading at right now?", "Compare bitcoin and ethereum by market cap", or "List the top 10 coins by market cap".

ToolForest (marketplace, new)

ToolForest is a hosted MCP aggregator that gives you 30+ toolkits over a single remote connection, with nothing to install. Add your API key, authorize the toolkits you want at app.toolforest.io, and Sorana discovers the available tools step by step before calling them. ToolForest tools are separate from the built-in Gmail, Calendar and ListenBrainz servers above.

Toolkits grouped in domains:

Google Workspace — Docs, Sheets, Slides, Gmail, Calendar, Tasks, Drive
Music — ListenBrainz, MusicBrainz, Last.fm, Apple Music
Health & Fitness — Fitbit, Oura Ring, Google Health
Prediction Markets — Kalshi, Polymarket, Gemini, Metaculus, Rothera
Social — X (Twitter), Reddit
Learning — Anki, Mochi
Planet Earth — GDELT, Geo, Earth Observation, Google Earth Engine
Other — GitHub, Notion

Highlights:

GDELT — global news coverage and machine-coded events by place, actor, theme and tone, including the headlines and the outlets behind each event. No API key needed.
Kalshi and Polymarket — public prediction market odds and price history. No account needed.
Earth Observation — Sentinel satellite imagery of an area over time, so you can ask what changed on a coastline.

Ask things like "Summarize the most recent emails across all my connected Gmail accounts", "Using GDELT, what is the most-reported event in Iran this week, and which outlets covered it?", "According to my Oura data, am I recovered enough to train hard today?", or "What moved this Kalshi market, and when?".

Other marketplace servers

Marketplace entries for Google Drive, GitHub, Slack, PostgreSQL and movies are under construction and still need manual credential setup. ToolForest already covers most of that ground today.

Using MCP Automation

Open MCP Manager and enable the servers you need
Create an agent in your workspace
Right-click the agent title → Chat
Give instructions in plain language, files, web, email, all accessible

Add external servers from the marketplace: the ToolForest aggregator for 30+ toolkits over one connection, the keyless CoinGecko server for crypto market data, or single-service servers for a specific tool.

OCR for Scanned Documents

To enable OCR on PDFs with embedded images, install Tesseract OCR and ensure it is in your system PATH. Once installed, Sorana processes scanned documents automatically, making your entire paper archive searchable. A vision-capable model can read images directly as a fallback.

Other Tools by Tetramatrix

TabNeuron, AI spatial tab manager & browser research workspace
Aicono, AI intelligent desktop icon organizer
RyzenZPilot, AMD Ryzen power management tool
Spaceship!, Retro 2D arcade game

Chat experience

Sorana is designed to make AI conversations useful, readable, and easy to continue:

Clear, easy-to-scan answers, important details stand out instead of getting lost in a wall of text. Comparisons can be presented as clean, readable tables.
Context that stays with your work, Sorana remembers relevant facts, preferences, and project details, so you spend less time repeating yourself.
Chat with your files, ask questions about PDFs, documents, scans, and images in plain language. Sorana finds the useful information inside your archive.
Goal-driven execution from one conversation, give Sorana a goal and it plans and works through it across your files, email, and calendar while you stay in control.
Long conversations stay on track, older parts are quietly condensed while important details remain available, keeping long-running projects easier to continue.
Workflows for bigger tasks, break complex work into connected steps, using a ready-made template or creating your own.
A clearer view of the work, when available, follow what Sorana searched, found, and used to build an answer, so you can better understand and guide the result.

Formatted responses can include headings, emphasis, links, and comparison tables, making research, planning, writing, and document work easier to follow from the first question to the final result.

Developed by Tetramatrix | Your desktop AI knowledge workspace and agent harness | Portable and private