AI Work Reference Guide
AI

Your personal reference for AI work tools — Claude Code, Codex Desktop, Claude Design, AI agent setups, and other AI research tools, with comparisons, workflows, and positioning in one place.

Claude Code Codex Desktop Claude Design AI Agents Other AI Tools Tool Links Cowork
Last updated: May 18, 2026
Change language
Personal Reference · Claude Code Setup
Claude Code Setup — Tools & Stack
I use Claude Code as a technical automation workspace for memory, web research, APIs, Google Workspace, Notion, delivery alerts, deployment, and repeatable data-output workflows.
Coding Agent Research Pipelines Google Workspace Automation Deployable Outputs

Claude Code personal setup v3 — with Google APIs and Notion

Memory & instructions
CLAUDE.md
Stores your SOP, naming rules, data source preferences, and output formats. Loads automatically every session without you repeating yourself.
Committed to GitHub for backup and team sharing. Supports importing other files with @path syntax.
GitHub backedRepo-native
Memory function
Claude stores and recalls project context, preferences, and recurring instructions across sessions. Works alongside CLAUDE.md to build a persistent profile of how you work.
CLAUDE.md = explicit written rules. Memory = things Claude learns and saves automatically during sessions.
PersistentCross-session
Web research & data extraction
Firecrawl
Converts web pages into clean Markdown or structured data. Best for extracting precise numbers, tables, and citations when you already have the URLs.
If a site blocks Firecrawl's crawler, fall back to Playwright — it runs a real browser and is harder to block.
Clean extractionBatch URLs
Playwright
Full browser automation — clicks, logins, filter dropdowns, scrolls, downloads. Use when the page is dynamic and content only appears after interaction.
Primary tool for social media and competitor sites. Also acts as Firecrawl fallback when a site blocks crawlers.
Dynamic pagesSocial media
Direct website APIs
Connect official APIs for maximum accuracy — CRM, ecommerce, finance, academic sources. Structured JSON responses eliminate messy HTML parsing.
Most accurate option when available. Every response is dated, structured, and auditable.
High accuracyVerifiable
Google workspace & Notion
Google Sheets API
Claude Code writes research results, price data, and reports directly into a Google Sheet. The sheet updates live — anyone with the link sees fresh data and you can view it on iPhone instantly.
Use alongside Excel output. Google Sheets = live sharing with team. Excel = offline archive or formatted report for sending.
Live sharingMobile view
Notion API
Claude Code reads specs and tasks from Notion, then writes research results, reports, and summaries back into the right Notion page or database automatically — no copy-paste.
Connect via Notion MCP in Claude Code. Best for knowledge base updates, weekly summaries, and keeping Notion as the single source of truth.
Auto-updateKnowledge base
Google Drive API
Claude Code reads files from Drive (PDFs, Docs, Sheets, Slides), processes them, and saves outputs back — Excel reports, Word docs, or summaries — directly into the right Drive folder.
Useful for reading source documents without downloading them first, and for delivering final outputs to a shared Drive folder the team can access.
Read + writeTeam folders
Automation & integrations
Zapier
Bridges Claude Code to apps without MCP — Gemini API, image generation, Todoist, and others. Claude Code calls a Zapier webhook when a task finishes to trigger the next step.
Bidirectional — Zapier can also trigger Claude Code via webhook if an external app needs to kick off a workflow.
App bridgeBoth directions
Gmail API
Claude Code sends formatted email reports when a workflow finishes — HTML body, subject line, and attached Excel or PDF. Better than Telegram for longer reports or sharing with people outside your chat.
Use Telegram for quick personal mobile alerts. Use Gmail for full reports with attachments sent to colleagues or clients.
Formatted reportsAttachments
Telegram
Claude Code calls the Telegram Bot API to push messages to your personal chat or a group after any workflow completes. Review results on iPhone without opening a laptop.
Async delivery — not a live chat with Claude. Claude pushes results; it doesn't listen for replies unless you build a two-way bot separately.
Async deliveryMobile review
Apple Reminders & Calendar
Uses shortcuts run CLI or AppleScript on macOS to create reminders and calendar events at the end of a workflow.
Mac-only — won't work on a Linux server. Requires macOS with Shortcuts or AppleScript enabled.
macOS onlyShortcuts / AppleScript
Build & deploy
HTML & vibe coding
Create, edit, preview, and debug HTML/CSS/JS files directly. Iterate with natural language — Claude manages all files, fixes console errors, and adjusts layout across the whole project.
Multi-filePreview & debug
Cloudflare deployment
Deploy via Cloudflare Pages using Wrangler CLI. Claude handles the build, deploy command, and any errors. End result is a public .pages.dev link you can open anywhere.
"Upload HTML via link" = Cloudflare Pages giving you a shareable public URL — no manual dashboard steps needed.
Public .pages.dev linkWrangler CLI
GitHub backup
CLAUDE.md, skills, scripts, and HTML projects are all version-controlled. Push directly from Claude Code. Supports rollback, team sharing, and Cowork plugin marketplace sync.
Version controlTeam sharing
Data output workflows
Competitive price research → Google Sheets + Excel
Playwright or Firecrawl visits competitor pages, extracts prices, validates source URLs and dates, then writes to both a local Excel file and a live Google Sheet the team can view instantly.
Playwright / Firecrawl Extract + verify Write .xlsx Google Sheets Telegram alert
Market research → Notion + Gmail report
APIs, Google Drive source files, and web pages feed into a structured report. Claude writes the summary into Notion and sends a formatted email with the full document attached.
Google Drive (source) API / Firecrawl Validate Notion update Gmail report

Your Stack — Claude Code vs Cowork

Your Claude Code stack compared to Cowork

Your tools — which handles it better
Based on personal workflows as of May 2026. Tools evolve quickly, so connector depth, limits, and product behavior should be checked against official docs before making long-term decisions.
Your tool
Claude Code
Cowork
CLAUDE.md + Memory
Claude Code
More repo-native because CLAUDE.md can live in Git, travel with the codebase, and load with project instructions.
Cowork has projects with files, context, instructions, and memory, but it is less repo-native than a Git-tracked CLAUDE.md workflow.
Firecrawl
Claude Code
Has a Firecrawl plugin path and works well for scripted extraction, batch URLs, fallback handling, and source saving.
Cowork does not appear to have Firecrawl as a standard connector. It would require a custom plugin, MCP, or other custom setup.
Playwright
Claude Code
Better for arbitrary Playwright scripts inside a repo, with repeatable login, click, filter, download, screenshot, and verification logic.
Cowork has browser/computer-use tools for navigating, clicking, typing, scrolling, screenshots, and extraction, but is less suited to repo-based Playwright scripts.
Direct website APIs
Claude Code
Better for repeatable API scripts with auth, pagination, retries, rate limits, JSON parsing, and Git-tracked code.
Cowork can run code/shell and make custom API calls, so it is not limited to built-in connectors. Code still wins for durable scripted pipelines.
Google Sheets API
Claude Code
Stronger for complex scripted Sheets API workflows, batch updates, formatting logic, validation, and larger pipelines.
Cowork can use Google Drive/Workspace connectors and local file workflows, but is less suited to deeply custom Sheets automation.
Notion API
Claude Code
Better for complex code-driven Notion pipelines: database updates, structured blocks, validation, sync logic, and multi-tool workflows.
Cowork's Notion connector supports create, edit, search, and update workflows, not just simple read/write.
Google Drive API
Claude Code
Stronger for scripted Drive workflows that read files, process PDFs/Docs/Sheets, generate outputs, and write back to team folders.
Cowork can use Google Drive/Workspace connectors and local file workflows, but complex batch processing is better handled in code.
Gmail API
Claude Code
Stronger for custom email pipelines if proper API credentials are configured: formatted bodies, attachments, recipients, and workflow triggers.
Cowork can search/read Gmail and create drafts, but the Gmail connector cannot send emails directly on the user's behalf.
Zapier
Claude Code
Better for deeply custom multi-step logic, webhook handling, branching, validation, retries, and code-controlled workflows.
Cowork can use Zapier to run actions across many apps, especially for standard business automations.
Telegram
Claude Code
Clearer path through a Telegram plugin or Bot API for sending text, files, formatted messages, and workflow alerts.
Cowork does not appear to have a standard Telegram connector. It may interact indirectly through Telegram Web/Desktop, Zapier, or custom setup.
Apple Reminders & Calendar
Claude Code
Stronger for macOS automation through Shortcuts CLI, AppleScript, or local tools at the end of workflows.
Cowork desktop does not appear to have native Apple Reminders or Apple Calendar connectors.
HTML & vibe coding
Claude Code
Better for multi-file apps, local dev servers, debugging, Git history, and deployment workflows.
Cowork can preview and render interactive HTML through Live Artifacts, but is less suited to full app/project lifecycle work.
Cloudflare deploy
Claude Code
Stronger for repeatable Wrangler CLI deploy workflows inside a repo: build, deploy, debug errors, and redeploy after code changes.
Claude has a Cloudflare connector that can manage/deploy Workers when available and enabled. Cowork can also run code/shell, but is less repo-deploy focused.
GitHub backup
Claude Code
Better for commit, push, branch, PR, version history, rollback, and repo-centered engineering workflows.
Cowork/GitHub support is useful for context and some connected workflows, but not general repo management in the same way.
Excel / Word output
Can generate .xlsx and .docx files programmatically from collected data. Best when code-driven data processing matters most.
Cowork
Stronger for polished Word docs, Excel reports, formatting, layout, and presentation-ready business outputs.
PowerPoint / slides
Can generate .pptx programmatically, especially from structured data, but visual polish may need more direction.
Cowork
Stronger for polished slide deck creation and visual output from existing content.
Summary
Claude Code is strongest for scripted automation
Your stack is heavily built on technical automation — APIs, scraping, browser scripting, pipelines, deployment, and delivery. Claude Code is strongest when the work needs repeatable scripts, Git-tracked code, debugging, and custom workflow logic.
Cowork is useful for the final presentation layer
Once Claude Code has collected and structured data, Cowork is better at turning that into polished Word documents, formatted Excel reports, and slide decks that are ready to present or send to clients.
Best combined workflow
Claude Code collects data (Firecrawl, Playwright, APIs), processes it (Google Sheets, Notion, Drive), delivers alerts (Telegram, Gmail), builds and deploys HTML tools (Cloudflare, GitHub). Then Cowork takes that structured data and produces polished Word reports, Excel summaries, and PowerPoint decks for meetings or clients.

Claude Design app description card with feature summary

Personal Reference · Claude Design
Claude Design — Visual Workflows

Claude Design is an Anthropic Labs research-preview workspace powered by Claude Opus 4.7. It lets paid Claude subscribers create and refine visual work such as designs, prototypes, slides, one-pagers, pitch decks, marketing collateral, and visual reports through natural language, inline edits, direct text changes, and live adjustment controls.

Designs & Prototypes Slides & One-pagers Brand Consistency Exportable Outputs Claude Code Handoff

Design workspace

Create designs, prototypes, slides, and one-pagers
Start from prompts, images, documents, PPTX, XLSX, or codebase context
Edit text, layout, visual hierarchy, and brand styling
Apply team design systems when connected

Data reporting

Create pitch decks, visual reports, and presentation flows
Export to PPTX, PDF, Canva, HTML, or shared internal URLs
Useful for sales, marketing, product, and founder workflows
May need iteration for complex dashboards or strict brand details

Dynamic visuals

Turn static mockups into shareable interactive prototypes
Use inline comments, direct edits, and custom sliders
Handoff build-ready design bundles to Claude Code

Designed for

Claude Pro, Max, Team, and Enterprise subscribers
Designers exploring multiple directions quickly
Non-designers who need polished visual drafts
Claude Design is available as a research preview for qualifying Claude plans. Features, limits, rollout access, and exports may change; check Anthropic's official documentation before relying on it for production workflows. If publishing publicly, consider whether the name "Claude Design" requires Anthropic's permission, since "Claude" is an Anthropic brand name.

Codex Desktop Usage Positioning

Personal Reference · Codex Desktop
Codex Desktop — Usage Positioning
I use Codex Desktop as a flexible research and execution workspace. It helps me search information, verify sources, analyze data, create outputs, automate workflows, generate visuals, manage project instructions, and turn ideas into working artifacts.
Research Agent Execution Agent Business + Technical Plan-first Workflow Plugin-extensible

Codex sits between a coding agent and a business work agent. It is not only a coding assistant — it can handle technical work like code, websites, data pipelines, debugging, and automation, but also business tasks such as research reports, spreadsheets, presentations, HTML dashboards, and image assets. The biggest advantage is that Codex can move from thinking to doing: search for information, compare sources, analyze files, create structured outputs, and continue improving the result until a finished artifact is delivered.

What Codex Can Do
Technical Work
Code & Automation
Codex can write, debug, and refactor code, build websites, design data pipelines, set up automations, and run technical checks. It can also delegate work to parallel subagents — one to research, one to build, one to verify — making it feel like a small work team.
Code & DebugWebsitesData PipelinesAutomationSubagentsMCP Servers
Business Work
📊Outputs & Deliverables
Beyond code, Codex produces finished business deliverables. The output is not just a summary — it is a working artifact: a polished document, a styled spreadsheet, an HTML dashboard, or a visual asset ready to use or present.
Research ReportsSpreadsheetsPresentationsHTML DashboardsDocumentsImage Assets
Research Capability
🔭Comprehensive & Source-based
Research in Codex feels more comprehensive and accurate because it can combine web search, local files, project context, connected tools, code repositories, and structured data — all in one session. Results are traceable and connected to action.
The percentages below are personal observed indicators, not benchmark scores.
Web Search + Source verificationObserved 95%
Combines multiple sources and cross-checks findings
Local Files + Project contextObserved 88%
Reads and reasons over uploaded or project files
Structured data analysisObserved 85%
Processes spreadsheets, CSVs, logs programmatically
Source-linked actionable outputObserved 90%
Research connects directly into deliverables
Customization & Memory
🧠Project-level Control
Codex supports project instruction files like AGENTS.md and memory features. This means I can shape how it works over time — preferred writing style, project rules, research method, design expectations, coding standards, and recurring workflows — without repeating instructions each session.
AGENTS.md instructionsProject memoryWriting style rulesCoding standardsDesign preferencesRecurring workflows
Plugins & Connectors
🔌Extensible by Design
Codex can be extended with plugins, skills, MCP servers, and connectors. Firecrawl via MCP is particularly useful for web scraping, crawling websites, and extracting structured data to improve research workflows.
Google Drive
GitHub
Slack
Linear
Notion
Gmail
Browser Tools
Firecrawl / MCP
Image Generation
Skills & Subagents
Codex vs Cowork
Cowork is strong for non-technical office tasks — organizing files, preparing documents, making spreadsheets, and synthesizing research for business users who want finished deliverables without managing technical details. Codex is better when the work requires deeper research, technical execution, customization, automation, project memory, or repeatable workflows.
CapabilityCodex DesktopClaude Cowork
Research & source-checkable info
Web search, file reading, cross-referencing
✓ Strong◐ Moderate
Technical execution
Code, pipelines, automation, debugging
✓ Strong— Limited
Office deliverables
Docs, slides, spreadsheets, formatted outputs
◐ Good✓ Strong
Plan-first, multi-step projects
Analyze goal → break into steps → execute
✓ Strong◐ Moderate
Plugin / MCP / connector ecosystem
Third-party service integrations
✓ Rich◐ Growing
Project memory & AGENTS.md
Persistent instructions across sessions
✓ Yes— No
Parallel subagents
Multiple agents working simultaneously
✓ Yes— No
Non-technical user accessibility
Easy for business users without technical setup
◐ Moderate✓ Strong
Default visual polish
Output design quality without extra direction
◐ Needs direction◐ Good

Three AI agent setup options

Personal Reference · AI Agent Options
AI Agent Options — Setup Paths
I use this section to compare practical ways to run AI agents: local Mac setups, always-on cloud servers, and advanced custom API agents connected to tools, memory, business systems, and messaging platforms.
Local Machine Cloud Server Custom Agent Phone Control Approval-first Workflow
Option 1 · Local machine

Mac mini / MacBook + Claude Code + Telegram, Discord, or iMessage

Runs the agent directly on your Mac. A Telegram, Discord, or iMessage bridge lets you send tasks from your phone into the agent running on the machine.

Always online?Only if Mac stays awake
Agent toolsClaude Code, local scripts, MCP tools
Phone controlTelegram bot, Discord bot, iMessage, Slack bot
Can message back?Yes, through the bridge you set up
DifficultyMedium
CostLow — use hardware you already own
Local company filesMac appsiMessageLocal documentsIn-office coding
Local document & report generationInternal file cleanupCoding assistant for company projectsDaily summaries from local dataWebsite / app updatesAdmin file preparationLight automation on the company machine
  • Write and fix code
  • Create and edit files
  • Generate reports from local files
  • Read CSV / Excel exports
  • Work with local folders
  • Run tests and scripts
  • Send progress updates to Telegram / Discord
  • Ask for approval before risky actions
  • Use local MCP tools
  • Use Mac-specific workflows including iMessage
  • Mac must not sleep
  • Power or internet outage stops the agent
  • Remote access must be secured
  • Not ideal for 24/7 cloud automation
  • iMessage control is possible but more fragile than Telegram / Discord
Option 2 · Cloud server

VPS + Claude Code CLI or Codex CLI + Telegram / Discord bot

A VPS acts as your always-online agent computer. You message it from Telegram or Discord, and it runs Claude Code CLI or Codex CLI on your behalf around the clock.

Always online?Yes, if VPS is healthy
Agent toolsClaude Code CLI, Codex CLI, scripts, MCP tools
Phone controlTelegram bot or Discord bot
Can message back?Yes
DifficultyMedium
CostVPS cost + LLM / API / subscription cost
24/7 codingReportsMonitoringAPI workServer automation
Daily sales reportsServer error summariesGitHub progress reportsCustomer support summariesWebsite uptime monitoringDeployment assistantDatabase backup checksCRM / export analysisMarketing performance reportsInternal automation tasks
  • Receive work from Telegram / Discord
  • Send progress updates back to your phone
  • Write code and fix bugs
  • Pull / push GitHub or GitLab repos
  • Run tests
  • Generate reports
  • Monitor server logs
  • Check database / API data
  • Create CSV, Excel, Markdown, or PDF reports
  • Deploy websites / apps
  • Restart services
  • Create GitHub issues / PRs
  • Use MCP tools
  • Run scheduled jobs
  • Cannot access local Mac files unless synced or uploaded
  • Needs secure server setup
  • API keys must be protected
  • Needs monitoring and restart logic
  • Needs approval rules for sensitive tasks
Option 3 · Advanced

VPS + LLM API + Hermes / Custom Agent

Instead of mainly using Claude Code or Codex CLI, you build or run a custom agent framework that connects to LLM APIs, tools, memory, business systems, and messaging platforms.

Always online?Yes
Agent toolsHermes, custom agent, OpenAI / Anthropic API, MCP, business APIs
Phone controlTelegram, Discord, Slack, web dashboard
Can message back?Yes
DifficultyHigh
CostVPS + API usage + ongoing maintenance
Repeated business workflowsAdmin automationSemi-autonomous operationsBusiness API integrations
Customer support triageLead qualificationCRM updatesInvoice processingOrder-status checksInventory monitoringDaily BI summariesDrafting replies / proposalsInternal knowledge assistantRepetitive admin automation
  • Receive tasks from Telegram / Discord
  • Remember repeated workflows
  • Use business APIs
  • Create tasks / tickets
  • Generate reports
  • Update CRM
  • Summarize emails and messages
  • Monitor systems
  • Retry failed jobs
  • Ask for approval
  • Run minor tasks while you are away
  • Build reusable workflows and skills over time
  • More complex to build safely
  • Higher chance of mistakes if too autonomous
  • Needs strong permission controls
  • Needs audit logs
  • Needs cost limits
  • Needs human approval for all external actions

Other AI tools for research and workflow support

Personal Reference · Other AI Tools
Other AI Tools

A supporting tool stack for research, browser tasks, and office-native work. Use these tools around Claude Code, Codex Desktop, and agent workflows when they fit the job better.

Other AI Research Tools

Perplexity

Useful for source-based research and fact-checking. Deep Research can search many sources, summarize findings, and create a structured report with citations.

Best for validating claims with web sources, academic references, forums, and public information. Free access is limited, and important citations should still be opened and checked manually.

Citation backed Deep Research

Grok

Useful for fast, updated information, especially current news, trends, product updates, public reactions, and market sentiment.

Strong for quick first-pass research because it can use real-time web search and public X posts. It may still miss context, so final claims should be verified elsewhere.

Real-time search X trends
Research Workflow

Combined research flow

Use Grok first to identify the newest updates quickly, then use Perplexity to organize, compare, and validate the information with citations.

This pairing works well when you need both speed and source checking. For high-stakes work, confirm key points with primary sources.

Fast discovery Verified sources

Claude in Chrome

Claude's Chrome extension can read webpages, click buttons, type into forms, navigate websites, and work across browser tabs.

Useful for browser automation, repetitive website tasks, online research, and webpage testing. Use carefully on banking, legal, medical, or confidential work systems.

Browser control Automation

Copilot

Copilot is a Claude agent add-on and integration tool for Word, Excel, and PowerPoint. It helps bring AI support into the office apps where document, spreadsheet, and slide work already happens.

Word: draft, rewrite, and summarize. Excel: support formulas, insights, and charts. PowerPoint: create slides, summarize decks, improve structure, and adjust formatting.

Claude agent add-on Word · Excel · PowerPoint

Useful official tool links

Reference Links · Official Tools
Tool Links

A quick launch board for the core tools mentioned in this guide: coding agents, web extraction, hosting, source control, VPS providers, and LLM API routing.

Claude Code & Codex
VPS (Virtual Private Server)
LLM API
Note: This is a personal reference guide, last updated May 18, 2026. AI product features, availability, connectors, limits, and subscription requirements can change quickly, so check official documentation before relying on any tool for production, legal, financial, medical, or client-facing work.