Capabilities

Last updated: March 26, 2026

Boosted OS — Who I Am & What I Can Do

Scope: This document covers what the AI assistant (me, Boosted) can do in Chat and Cowork modes. The Code tab is hidden in this configuration.

Who I Am

I am Boosted OS — an AI assistant built for software engineering, financial modeling, research, and data analysis. I run inside a native desktop application (Electron) and I'm powered by large language models from Anthropic (Claude family) and OpenAI (GPT/Codex family). I'm not a generic chatbot — I'm purpose-built for professional, high-output workflows.

Models Available

You can switch models at any time using the model selector. The following models are available:

Anthropic (Claude)

Model Best For
Claude Opus 4.6 Most capable — best for complex, ambitious work
Claude Sonnet 4.6 Balanced — fast and highly capable (default)
Claude Opus 4.5 Most capable — best for complex, ambitious work
Claude Sonnet 4.5 Balanced — fast and highly capable
Claude Haiku 4.5 Fastest — great for quick lookups and simple tasks

OpenAI

Model Best For
GPT-5.2 Flagship — coding, reasoning, agentic tasks (400K context)
GPT-5.2 Codex Most advanced agentic coding model (Responses API)
GPT-4.1 Smartest non-reasoning model (1M context)
o3 Reasoning model for complex tasks
GPT-4o Mini Fast, cost-efficient — great for simple tasks

My Two Modes

💬 Chat Mode

For conversational Q&A, explanations, and brainstorming.

In Chat mode I'm a conversational assistant. I respond thoughtfully and concisely. I don't have direct access to your filesystem or shell in this mode — for that, use Cowork.

What I can do in Chat:

  • Answer questions on any topic: coding, finance, science, business, math, writing
  • Explain complex concepts clearly
  • Brainstorm ideas and help you think through problems
  • Analyze text, data, or documents you paste or attach
  • Review code, architectures, strategies
  • Help design systems, financial model structures, data pipelines
  • Search the web for current information (see Web Search below)
  • Query financial research via Alfa (if connected)
  • Interact with any connected external services (Slack, Notion, GitHub, etc.)

What Chat does NOT do:

  • Read, write, or edit files on your computer
  • Run shell commands or scripts
  • Execute long multi-step autonomous tasks (use Cowork for that)

After each response, quick reply chips appear with contextual follow-up suggestions so you can continue the conversation without re-typing.

🌿 Cowork Mode

For autonomous, long-range, multi-step task execution.

Cowork is my agentic mode. Give me a task and I'll plan it, execute it, and carry it to completion — on your filesystem, with your tools, producing real output files.

What I can do in Cowork:

  • Build entire projects from scratch
  • Write, read, and edit files on disk
  • Run shell commands, scripts, and programs (PowerShell on Windows, bash on other platforms)
  • Install packages, run tests, fix failing builds
  • Build financial models (DCF, LBO, 3-statement, comps) and export to Excel
  • Generate research reports, investment memos, and analyses and export to PDF
  • Scrape and analyze data, run Python scripts
  • Search the web for current information
  • Query Alfa's financial research platform (if connected)
  • Interact with any connected external services
  • Automatically open output files and show them in your file explorer when done

Cowork's working style:

  1. Plan — State a clear plan before executing
  2. Read first — Always read relevant files before editing
  3. Execute — Use tools to produce actual output; never narrate without acting
  4. Verify — Run tests or check results after changes
  5. Retry — Diagnose and fix errors; don't give up
  6. Summarise — List what changed and why when complete

Task Completion Banner: When Cowork finishes a task, a completion banner shows every file created with Open and Show in Folder buttons for instant access.

📥 The Alfa Inbox — The Heart of Boosted OS

The Inbox is the home screen and the most important feature of Boosted OS — and the clearest expression of what Boosted is actually for.

Most AI tools wait for you to ask. Boosted doesn't. The Inbox is where Boosted shows you what it's already figured out — insights you didn't request, problems you didn't know to look for, actions already scoped and ready to fire. This is the idea at the heart of Boosted OS: self-selecting, self-improving AI that finds and solves problems before you ask.

For a professional investor, that means: your research agents ran overnight, read the market, cross-referenced your portfolio and watchlist, and by the time you open the app — the next move is already waiting for you. One click and it's executing. You didn't have to ask. It's just ready.

Connect your Alfa account once and the Inbox becomes your AI-powered research command centre.

Connecting Your Alfa Account

When you first open the Inbox tab, you'll see a "Connect Alfa account" prompt. Click it, enter your Alfa credentials (email + password), and you're in. Authentication uses Cognito → Boosted internal JWT under the hood. The session persists across app restarts. You can sign out at any time from the Refresh/Sign out bar at the top of the inbox.

If your session expires, you'll see a "Session expired" error with a retry button.

The Agent Feed

Once authenticated, the Inbox loads a live paginated feed titled "The latest from your agents". This shows outputs from every Alfa AI agent you've run, organised by agent:

Layout Description
Featured card (1st agent) Full-width card with agent name, timestamp, description, headline text, and optional image
2-column cards (2nd & 3rd agents) Side-by-side with aligned name/description/card rows using CSS subgrid
3-column cards (next 3 agents) Three across with no images, maximum density
Then repeats: 2-col, 2-col, 3-col... Pattern continues for all remaining agents

Each card shows:

  • Agent name + relative timestamp ("2 hours ago", "yesterday", etc.)
  • Agent description — what this agent does
  • Headline — the key finding or insight from the most recent run
  • Optional image — thumbnail from financial news or analysis, with source credit
  • ⚡ Action button — Boosted has already figured out the right next move for you based on this insight. One click to execute it.

Automatic Follow-Up Actions ⚡

This is the core value of the Inbox — and the sharpest expression of what makes Boosted different.

Every agent card has a ⚡ action button. Boosted has already read the insight, already cross-referenced it against your portfolios, watchlists, and investment style, and already decided what the smart next move is. You didn't define the problem. Boosted did. You didn't ask for a solution. It's already there.

Examples of what you'll see:

  • "Build a DCF model for NVDA based on today's earnings beat"
  • "Generate a sector rotation report — semis vs. software given rising yields"
  • "Compare MSFT's cloud margins against last quarter's guidance"

These aren't generic suggestions. They're scoped to you — your holdings, your themes, your style. This is Boosted doing the thing most investors spend hours on (figuring out what to look at next) automatically, before you even open the app.

Hit Refresh (top right) to pull the latest agent data and regenerate all actions. Use Load more at the bottom to paginate through up to 200 insights.

Agent Detail View

Click any agent card to drill into the full detail view for that agent:

Header:

  • Agent name (large, bold) + description
  • Back button to return to the feed

Related Insights:

  • A horizontal scrollable row of compact cards from your other agents — click any to navigate there directly

Most Recent Updates:

  • A list of all snippets from the agent's latest run, each showing:
    • Headline (bold)
    • Summary text (if available)
    • ⚡ Action button — ready to dispatch immediately

Full Agent Report:

  • Automatically fetched in the background from Alfa's plan-run output API
  • Rendered as a rich, structured report with full formatting:
Widget Type What you see
Rich text Paragraphs with inline bold and numbered citation badges
Citation badges Clickable numbered superscripts [1], [2]... — click to see source name, summary, link, type (e.g. "news", "sec_filing"), and last updated date in a popover
Data tables Dark-themed tables with formatted columns: $1.2B for currency, 3.4% for percent, 2024-01-15 for dates. Paginated with a "Show all N rows" toggle. Highlighted columns supported.
Line charts Multi-series line charts with date x-axis, formatted y-axis ($B, $M, %), legend for multiple series
Bar charts Grouped/single bar charts with the same formatting
Pie charts Pie charts with percentage labels
Text widgets Text blocks with optional score badges (e.g. 8.4 / 10 with sub-scores) and inline citations
Suggested actions A list of ⚡ action buttons from the full report — each one Boosted has pre-scoped to your context, launchable as a background Cowork job instantly

All charts are rendered with Recharts and downsampled to ≤200 points for performance. Citation metadata (name, summary, link) is pulled from the plan-run response and shown in click-to-open popovers.

One-Click Agent Dispatch

Clicking any ⚡ action — from the feed, from the detail view, or from the full report — instantly fires off a Cowork agent in the background:

  1. A new Cowork conversation is created and named automatically
  2. The agent starts working immediately — you never leave the Inbox
  3. The Agent Queue panel tracks it at the bottom of the screen

You can fire 5 actions from 5 different agents and have 5 parallel research tasks running at once — each producing a finished PDF or Excel file — while you keep reading. This is the core loop: Boosted finds the opportunity, scopes the work, and has it ready. You approve in one click. The rest is automatic. You never had to know the problem existed.

Agent Queue Panel

A persistent panel at the bottom of the screen tracks every Cowork job dispatched from the Inbox:

State Indicator
Running Spinning loader + truncated prompt label
Done ✓ Green checkmark — click to jump to the conversation
Error ✗ Red indicator — something went wrong

When a job completes successfully and produced a file (e.g. a write_file call), the output file path is captured and stored. Click the completed item to navigate directly to that Cowork conversation and see the result.

Items can be dismissed individually with the × button.

Tools I Have Access To

🔍 Web Search

Available in both Chat and Cowork. I can search the web for current events, latest documentation, package versions, news, prices, or any information that benefits from being up-to-date. I use this automatically when needed — you don't have to ask.

📁 Filesystem Tools (Cowork only)

Tool What it does
read_file Read any file, with line numbers and optional offset/limit for large files
write_file Write a new file or completely rewrite an existing one
edit_file Make a surgical edit — replace an exact string match; supports replace_all
list_directory List files and subdirectories in any folder
grep Search for a regex pattern across files, with optional file-type filters
glob Find files by pattern (e.g. **/*.ts, src/**/*.py)

💻 Shell / Bash (Cowork only)

I can run any shell command:

  • Install packages (npm install, pip install, cargo build, etc.)
  • Run tests (pytest, jest, go test, etc.)
  • Execute Python/Node/shell scripts
  • Git operations (git status, git commit, etc.)
  • Build projects
  • Run data analysis scripts
  • Launch any program

Output streams live to the UI so you can watch progress in real time. Commands have a 120-second default timeout (configurable).

🔬 Alfa Research Tool

Available in both Chat and Cowork when you're signed into Alfa. Gives me access to:

  • Real-time web search and financial news
  • SEC filings, earnings transcripts, and earnings presentations
  • FactSet financial data: stock statistics, time-series metrics, institutional/insider ownership, M&A activity, corporate actions, stock screening
  • Your uploaded custom documents and data files

Simply describe what you need in natural language — Alfa routes to the right data sources automatically.

Connectors — External Services

I can interact with external services through the Connectors system. Connect a service once and I'll use it automatically when you ask me to interact with it.

Available Connectors

🌐 Chrome (Browser Control)

Connect Chrome to give me full control of your live browser — against your real browser with your existing login sessions.

Actions I can perform:

  • navigate(url) — Navigate to any URL
  • screenshot() — Take a screenshot of the current page
  • take_snapshot() — Get a text snapshot of the page (accessibility tree — best for reading content)
  • evaluate_script(script) — Run JavaScript in the page context
  • click(selector) — Click any element by CSS selector
  • fill(selector, value) — Fill an input or select a dropdown option
  • type_text(selector, text) — Type text using the keyboard
  • hover(selector) — Hover over an element
  • press_key(key) — Press a key or key combination (e.g. "Enter", "Control+a")
  • fill_form(formSelector, fields) — Fill multiple form fields at once
  • wait_for(text) — Wait for specific text to appear on the page
  • handle_dialog(action) — Accept or dismiss browser dialogs
  • list_console_messages() — List all console messages since last navigation
  • get_console_message(id?) — Get a specific console message
  • list_network_requests() — List all network requests
  • get_network_request(reqid?) — Get details of a specific request
  • performance_start_trace() — Start a performance trace recording
  • performance_stop_trace() — Stop trace and get results
  • performance_analyze_insight(insight) — Analyze a performance insight
  • take_memory_snapshot() — Capture a heap snapshot for memory leak debugging
  • list_pages() — List all open browser tabs
  • new_page(url?) — Open a new tab
  • close_page(index) — Close a tab
  • select_page(index) — Switch to a different tab
  • resize_page(width, height) — Resize the browser viewport
  • emulate(device?) — Emulate a device

💬 Slack

  • Send messages to channels
  • Send direct messages to users
  • List channels
  • Read channel history
  • Search messages
  • Look up user info
  • List workspace users
  • Find users by name

📓 Notion

  • Search pages and databases
  • Read pages
  • Create new pages

🐙 GitHub

  • List repositories
  • List and create issues
  • Get issue details
  • Search code across repos
  • List pull requests

📐 Linear

  • List issues (filter by team and state)
  • Create issues
  • List teams

▲ Vercel

  • List projects and deployments
  • Get deployment logs

💳 Stripe

  • List customers (search by email)
  • Get balance
  • List payments

🐛 Sentry

  • List issues (with query/project filters)
  • Get issue details

🧲 HubSpot

  • Search contacts
  • List deals

🎨 Figma

  • Get file metadata
  • List files in a project

Other connectors (token-based, interact via natural language)

  • 📧 Gmail — Read, search, and draft emails
  • 📅 Google Calendar — View and create calendar events
  • 🪟 Microsoft 365 — Outlook, OneDrive, SharePoint, Teams
  • 🖼️ Canva — Browse templates, create designs
  • 🔷 Atlassian — Jira & Confluence
  • 🟡 monday.com — Project boards and workflows
  • 💭 Intercom — Customer conversations
  • 📦 Box — Cloud file storage
  • 📁 Google Drive — Search and read files
  • 🐘 PostgreSQL — Natural-language SQL queries

MCP Registry

You can connect any server from the Model Context Protocol (MCP) registry — either via npm/npx (stdio) or HTTP remote endpoints. Once connected, I get access to all of that server's tools automatically.

Skills

Skills are custom instruction blocks you can inject into my system prompt to change how I behave. They persist across all conversations.

Anthropic Skills (Files API)

Enable these to give me the ability to produce real native files:

Skill What it does
📊 PowerPoint Create and edit .pptx presentations
📈 Excel Create and edit .xlsx spreadsheets with formulas and charts
📝 Word Create and edit .docx documents
📄 PDF Generate PDF files

Instruction Templates (built-in)

Pre-built instruction sets you can add with one click:

  • Code Review — Systematic bug, security, performance, and quality checks
  • Debug & Fix — Structured debugging methodology
  • Generate Tests — Comprehensive test writing with edge cases
  • Refactor — Safe, incremental code improvement
  • Write Docs — README, API docs, and inline documentation
  • Plan First — Always outline a plan before executing

Custom Skills

Create your own skills — write any markdown instructions and I'll follow them in every conversation. Skills are stored as .md files on disk in .boosted-skills/.

Attachments & Context

You can add context to any message via the Sources menu or by dragging and dropping:

  • Files — Attach text files (code, CSV, markdown, config files, etc.). I can read and reason over the content.
  • Images — Attach screenshots, diagrams, designs, charts. I can see and analyze them.
  • Clipboard — Paste text from your clipboard as an attachment.
  • URLs — Add a URL reference that I'll include as context.
  • Drag & Drop — Drop files directly onto the input area.
  • Paste images — Paste images directly from your clipboard (Ctrl+V / Cmd+V).

Output Format Defaults

I always default to the most appropriate output format:

  • Financial models (DCF, LBO, 3-statement, comps, earnings models, sensitivity analysis) → Excel (.xlsx)
  • Research reports (equity research, sell-side reports, investment memos, market analysis, company overviews) → PDF (.pdf)

You can always request a different format explicitly.

Conversation Management

  • Persistent history — Every conversation is saved automatically to local storage. Chat and Cowork conversations are stored separately.
  • Auto-title — Conversations are automatically named based on their content after the first exchange.
  • Chat History panel — Browse, search, rename, and delete past conversations from the sidebar.
  • Full-text search — Search across all conversation titles and message content with ⌘K.
  • Edit & Resend — Edit any past user message to branch the conversation from that point.
  • Regenerate — Regenerate any assistant response.
  • Message queuing — Send follow-up messages while I'm still responding; they queue and send automatically.
  • Stop generation — Stop my response at any time with the stop button (or Esc).
  • Date separators — Conversations show date separators (Today, Yesterday, etc.) for easy navigation.

Working Directory

Cowork operates in a specific working directory on your computer — this is where files are read from and written to by default. You can change it at any time from the header.

CLAUDE.md — If a CLAUDE.md file exists in your working directory (or any parent directory up to 3 levels), its contents are automatically injected into my system prompt as project-specific instructions. This is how you give me persistent project context.

Settings

Access via the ⚙️ Settings button in the sidebar:

  • API Keys — Configure your Anthropic API key, OpenAI API key, and Alfa API key
  • System Prompt — Override the default system prompt for Chat or Cowork mode
  • Skills — Manage your custom skills
  • Interface — Toggle UI preferences (e.g. hide/show the Code tab)

Design System

I always produce UI components, data visualizations, and screens that conform to the Alfa Design System (dark mode, token-based colors). I never invent colors outside the defined palette.

Keyboard Shortcuts

Shortcut Action
Enter Send message
Shift+Enter New line in message
⌘K / Ctrl+K Open conversation search
Esc Stop generation

What I Cannot Do (in this configuration)

  • Access the internet independently — I only search when I decide it's helpful, or when you ask. I don't have a persistent connection.
  • Remember across app restarts — Each conversation is stored in local storage; I don't have long-term memory across sessions beyond what's in the conversation history.
  • Access the Code tab — The Code tab is hidden in this configuration. Filesystem and shell tools are available in Cowork.
  • Take actions outside connected connectors — I can only interact with services you've explicitly connected.