Avanai Claude Cowork · Deep Dive
Intro 01 / 00
A Deep Dive · 2026

Claude Cowork

The native desktop agent that turns Claude from a chatbot you talk to into a colleague that works alongside you - on your files, your apps, and your schedule.

Local files Persistent memory Connectors & MCP Skills & plugins Scheduled tasks Live artifacts
Press → to begin · ? for help
Why this deck

Most teams are using 10% of what Cowork can do.

Cowork looks like another chat window. It isn't. Under the hood it's the same agent that powers Claude Code, wrapped in a UI that any non-technical operator can use. Once you understand the primitives - workspace, memory, connectors, skills, scheduled tasks - Cowork stops feeling like a tool and starts feeling like a teammate.

01

What Cowork actually is

How it differs from Claude Chat and Claude Code, and why that matters.

02

Every feature, one slide each

From workspace folders to dispatch - the full primitive set in plain English.

03

10 enterprise-grade use cases

Real workflows finance, sales, legal, and ops teams are running today.

02
The Claude product surface

One agent, three doors.

Anthropic ships the same underlying Claude agent across three modes inside one desktop app. They share a model, a runtime, and a connector layer - but each is tuned to a different person.

Chat

Claude Chat

The familiar chatbot. Cloud-hosted. Great for shower-thoughts and one-off Q&A.

20 files / 30 MB per upload. Output stays in the chat window.

Cowork

Claude Cowork

The native desktop agent. Reads & writes your local files. Remembers across sessions. Connects to your stack.

For operators, founders, knowledge workers - non-technical or otherwise.

Code

Claude Code

The terminal-native coding agent. Maximum customization, hooks, sub-agents.

For engineers shipping production code; powers Cowork under the hood.

03
Cowork vs Chat vs Code

The capability gap, one row at a time.

Capability
Chat
Cowork
Code
Reads files on your computer
No
Yes - folder-scoped
Yes
Writes & edits real files
No (download only)
Yes
Yes
Persistent memory across sessions
Limited (cloud)
Yes - local CLAUDE.md / memory.md
Yes
Context window
Standard
~1M tokens
~1M tokens
Connectors (Gmail, Notion, Slack, Drive, …)
Some
Full + custom MCP
Full + custom MCP
Scheduled / unattended runs
No
Yes - daily / hourly / cron-like
Yes
Browser & computer control
No
Yes (Chrome ext + computer use)
Via tools
Mobile control of the desktop
No
Yes - Dispatch
Via Code app
Primary user
Anyone with a question
Operators & teams
Engineers
04
The mindset shift

Stop prompting. Start delegating.

Chat rewards task-first language: "review my photos and recommend a naming convention." Cowork rewards outcome-first language: "I have 15 photos in this folder, I need them organized into subfolders by topic with descriptive filenames." You define the result and the constraints. Cowork executes.

Chat-style prompt
  • "Can you look at my receipts and tell me what to do?"
  • You read the answer, then do the work yourself.
  • Output lives in a chat window you'll lose tomorrow.
Cowork-style prompt
  • "In the receipts folder, build an Excel with date, vendor, category, amount + totals row. Flag anything blurry."
  • Cowork reads every receipt, extracts data, writes the file, flags exceptions.
  • Output is a file in your folder - and the workflow can become a reusable skill.
05
For enterprise teams

A general-purpose agent with your institutional knowledge baked in.

95%
Token reduction reported when moving from raw search to a Cowork knowledge wiki
5-10×
Output multiplier on operator-heavy roles in early rollouts
Sessions remembered when memory + projects are configured

Cowork isn't a productivity hack. It's a substrate: every conversation, every skill, and every connector compounds into a system that knows your business better month over month.

06
02

The Deep Dive

Every primitive in Cowork. Workspace through Dispatch.

07
Feature 01 · Workspace

Folder-scoped access to your computer.

Cowork starts every session by mounting a folder you choose. It can read, write, rename, and organize anything inside that folder - and nothing outside it. Files never leave your machine.

📁

Opt-in by folder

You grant access per folder. "Always allow" remembers it for next time.

🔒

Hard sandbox

Files outside the mounted folder are invisible - even drag-drops from Downloads.

🧹

Deletion protection

Destructive actions prompt for confirmation. Guardrails are configurable.

08
Feature 02 · Outcome-first prompting

Describe the deliverable, not the steps.

Define the end state, the constraints, and the quality bar - Cowork plans the rest.

✗ Step-by-step (avoid)
# Micromanaging the agent 1. Open /q4-marketing folder 2. Read update-alice.doc 3. Read update-bob.doc 4. Read update-carol.doc 5. Merge them into report.doc 6. Add a metrics section 7. Convert to PDF → OK... (but misses your intent, asks at every step)
✓ Outcome-first (do this)
# Describe the deliverable In /q4-marketing, combine the three team update docs into one weekly report I can send to leadership. Lead with three top-line metrics, then 3 highlights and 3 lowlights. Cap at 300 words. Output as PDF. → Plan: read 3 docs, extract metrics, draft, render to PDF, save.
09
Feature 03 · Persistent memory

A notebook your agent checks every morning.

Two simple markdown files do the heavy lifting. CLAUDE.md holds standing instructions for a folder or project. memory.md is the running log Cowork writes to as it learns your preferences, decisions, and corrections.

CLAUDE.md - the rules

Loaded at the start of every session. Who you are, your folder map, naming conventions, safety rules, output preferences. Plain text - no special format.

memory.md - the log

Append-only. Captures corrections you give Cowork ("never sign off with 'best'"), context from past sessions, and project decisions. Survives restarts and machine moves.

10
Feature 04 · Global instructions

One identity, every session.

Settings → Cowork → Global instructions is a permanent system prompt that applies to every Cowork session, regardless of folder. Use it for your role, your tone, your safety rules, and the universal "always do this, never do that" guardrails.

Use globals for…

Tone of voice · British/US spelling · "ask before deleting" · "always plan first, then execute" · pointer to your About-Me folder.

Use CLAUDE.md for…

Project-specific rules · folder map · client-specific tone · approval workflows · which connectors to use here.

11
Feature 05 · Projects

A separate brain per workstream.

A Project is a folder + its own memory + its own instructions + its own scheduled tasks + its own connectors. Stop blending your YouTube planning with your Q4 close. Spin up a Project per client, per workstream, or per team.

Per-client workspace

Client A, Client B, Client C - each with their own brand voice, deliverables, contracts.

Per-team workspace

Finance, RevOps, Legal, Marketing - each with their own connectors, skills, and reporting cadence.

Per-program workspace

Quarterly planning, board reporting, an active M&A diligence - short-lived, high-context Projects.

12
Feature 06 · Context window

~1 million tokens of working memory.

Roughly the equivalent of a full novel of context per conversation. Long meeting transcripts, multi-doc reviews, sprawling codebases - all stay loaded. When the desk fills up, Cowork compacts the conversation rather than dropping it.

1M
tokens per session
13
Feature 07 · Multimodal input

Everything Claude can see.

Cowork ingests text, PDFs, images, screenshots, and charts natively. Drop a folder of receipt photos, a stack of contract scans, or a competitor screenshot - Cowork reads, extracts, and acts.

PDFs

Multi-hundred-page docs, contracts, statements.

Images

Receipts, whiteboards, UI screenshots, infographics.

Spreadsheets

XLSX / CSV - read, transform, write back.

Office docs

Word, PowerPoint, and Markdown round-trip cleanly.

14
Feature 08 · Web search

Live information, in-line with the workflow.

Cowork can search the web mid-task - pulling pricing, regulatory updates, news, or reference docs and folding them into the artifact it's producing. Combine with the Chrome extension to actually navigate, fill forms, and lift content from authenticated sources.

  • Knowledge cutoff is May 2025 - web search keeps responses current.
  • Pulls into the same context window as your local files, so research lands next to your draft.
  • Cite-able: surface URLs in the output so reviewers can verify.
15
Feature 09 · Extended thinking

A "take your time" mode for the hard ones.

Toggle extended thinking when the task involves planning, multi-step reasoning, or trade-offs. Cowork takes longer but produces dramatically better plans - and on the right model, the smarter run uses fewer tokens overall, not more.

Use it for

Strategic plans · contract analysis · architecting a new system · diagnosing a complex bug · synthesizing many sources.

Skip it for

Quick rewrites · simple file moves · single-file edits · summarizing a single email.

16
Feature 10 · Artifacts

Outputs that aren't just text.

Cowork can render structured outputs - interactive HTML dashboards, charts, mini-apps, comparison tables, slide decks - directly in the conversation. The deck you're reading now is an artifact-style output: HTML, CSS, and JS produced from a single prompt.

Dashboards

KPIs, pipelines, project status - interactive HTML.

Documents

Word, PDF, Markdown - fully formatted, ready to send.

Decks & reports

Branded HTML or PPTX from a script or transcript.

17
Feature 11 · Live Artifacts (new)

Dashboards that refresh themselves.

A Live Artifact is an HTML page that calls your connectors every time it opens. Build a Stripe revenue dashboard once; it pulls fresh numbers every time you click it. Build a meeting prep view once; it shows your next meeting, every time.

Personalized software, on demand

Each artifact is a tiny app you didn't have to build, deploy, or maintain.

Pinned in the sidebar

Surfaces alongside your tasks. One click to a fresh view.

Cheaper than a scheduled task

Refreshes on open instead of burning tokens hourly.

18
Feature 12 · Connectors

Reach into the tools your team already uses.

First-class connectors expose Gmail, Calendar, Drive, Notion, Slack, Stripe, HubSpot, GitHub, Linear and more - with granular per-action permissions. Cowork composes them: pull from one, cross-reference with another, write into a third.

Gmail
Calendar
Drive
Notion
Slack
Stripe
HubSpot
GitHub
Linear
Asana
Salesforce
+ 60 more
n8n
Automate everything

n8n connects 400+ apps and bridges directly into Cowork via MCP.

19
Feature 13 · MCP servers

When the connector you need doesn't exist yet.

Model Context Protocol is the open standard for plugging any system into the agent. Cowork treats MCP tools as first-class - same permissions, same UI.

Internal MCPs

Wrap your data warehouse, ticketing system, or homegrown CRM.

Vendor MCPs

Notion, Browserbase, and dozens of partners ship official servers.

Bridge MCPs · The long-tail unlock

One MCP connection → thousands of apps callable from Cowork.

n8n 400+ app workflows
Zapier 8,000+ app actions
20
Feature 14 · Skills

Procedural knowledge, packaged in a folder.

A Skill is a folder with a SKILL.md and (optionally) scripts, templates, and assets. It teaches Cowork how your organization does a specific thing - close the books, review an NDA, triage a support ticket - exactly the way you want it done. Skills load on demand to keep the context window lean.

  • Foundational skills: Anthropic-built capabilities (Excel, PowerPoint, PDF, contract review).
  • Vendor skills: shipped by partners - Notion deep research, Browserbase Stagehand, etc.
  • Org skills: where the moat is. Built by your team, for your team's bespoke workflows.
21
Feature 15 · The skill creator

Reverse-engineer a skill from the workflow you just did.

Best practice: do the workflow once, end-to-end, with feedback and corrections. Then ask Cowork to "turn what we just did into a skill." It interviews you on the edge cases, generates the SKILL.md, and tests against scenarios. The result is a one-click capability your whole team can use.

Don't
  • Build skills from scratch in the abstract.
  • Skip the test-cases step.
Do
  • Run the workflow live first, give corrections.
  • Ask Cowork to codify the conversation into a skill.
  • Back the skill up to Drive - it doesn't auto-sync across machines.
22
Feature 16 · Slash commands

One keystroke triggers a workflow.

Slash commands are the keyboard shortcut to your skills. Type /morning-briefing and Cowork loads the skill, pulls inbox + calendar + news, drafts replies, and renders a dashboard. For ops-heavy roles, slash commands turn a 30-minute morning into a 30-second one.

/morning-briefing → Loading skill: morning-briefing.md → Calendar: 4 meetings today → Inbox: 18 unread, 3 marked urgent → Drafts: 3 prepared in Gmail → Rendering dashboard…
23
Feature 17 · Plugins

Skills, distributable as a bundle.

A plugin is a zipped collection of skills (and optionally MCP configs) you install with one click. Anthropic ships first-party plugins for Legal, Marketing, Engineering and more; teams ship their own internally. The Avanai use case: package the full workflow for an entire role, not a single task.

Anthropic plugins

Legal triage, customer support, design, engineering review.

Vendor plugins

From your tooling vendors - Apollo for sales, etc.

Org plugins

"Onboard a new BDR." "Run a Q-end close." "Open a new project office." Whole roles, packaged.

24
Feature 18 · Sub-agents

Parallel workers for complex briefs.

For a brief like "prep my investor update," Cowork can fan out: one agent pulls financials, one runs market research, one drafts the narrative, one builds the deck. They run concurrently and their outputs are merged. The user sees one task; under the hood, four ran in parallel.

Researcher

Web + competitive landscape.

Analyst

Pulls data from Stripe / Notion / Sheet.

Writer

Drafts in your tone of voice.

Designer

Renders the slide deck.

25
Feature 19 · Bash & code execution

When the answer requires real computation.

Cowork can write and run code in a sandboxed environment - Python for data analysis, shell for file ops, Node for web scraping. You don't see the code unless you want to. You see the result.

  • Resize 200 images, batch-rename a folder, transform a CSV - these become one-shot operations.
  • Real numerical analysis on real data, not summaries of data.
  • Sandboxed: code runs in a VM, isolated from your broader system.
26
Feature 20 · Claude in Chrome

When there's no API, there's a browser.

The Chrome extension lets Cowork drive a browser tab - log into legacy portals, scrape an admin panel, fill a vendor form, pull a report no one's bothered to API-ify. Granular permissions per site, with a visible red ring while it's controlling the page.

Use it for

Legacy SaaS without APIs · government & insurance portals · authenticated dashboards · multi-step forms.

Watch out for

Slower than a connector - burns more tokens. Always prefer MCP/connector if one exists.

27
Feature 21 · Computer use

The full desktop, when you really need it.

Beyond the browser, Cowork can drive native apps - open Cap Cut and import a video, drag a file into iMessage, click through Excel. It's the option of last resort: slowest, most token-hungry, but unblocks anything else.

Connectors first Then MCP Then browser Computer use as fallback
28
Feature 22 · Scheduled tasks

An employee that works the graveyard shift.

Schedule any prompt - daily, hourly, on a cron-like cadence, or "every weekday at 6 AM." The moment that turns Cowork from "tool" to "teammate." Pair with a connector and a skill, and you have an unattended job that produces a finished artifact before you wake up.

Morning command center

Inbox + calendar + AI news → pinned dashboard at 6 AM.

End-of-day wrap

Cowork tells you what it did while you were in meetings.

Weekly & monthly

Roll-ups, reports, anomaly scans - same day, same time, every week.

29
Feature 23 · Dispatch

Drive your desktop from your phone.

Dispatch links the Claude mobile app to your always-on desktop. From the gym, the commute, or a coffee shop, send a task and Cowork executes it on your real machine - with your files, your connectors, your permissions. Optionally pair with iMessage so a watch tap kicks off a workflow.

Voice-friendly

"Hey Siri, send Frank a message" → kicks off a research run before you're back.

Two-way conversation

If Cowork needs clarification, it asks - you answer from the phone.

Always-on Mac pattern

Many heavy users keep a Mac mini awake for unattended runs.

30
Feature 24 · Safety, permissions & reverse elicitation

A teammate that asks before it does anything dangerous.

🛡️

Folder sandboxing

It only sees folders you mount. Drag-drops from outside don't grant access.

↩️

Deletion protection

Destructive ops require explicit confirmation. Globals can require it for moves & renames too.

Reverse elicitation

When uncertain, Cowork asks - instead of guessing. Trained behavior, not a setting.

🧬

Alignment work

Mechanistic interpretability + adversarial training to harden against prompt injection.

📨

Drafts, not sends

Default behavior for email is "draft for review." You stay the approver.

🔐

Per-action permissions

Allow once, allow always, deny - at the level of individual connector actions.

31
Feature 25 · The recommended folder pattern

A second-brain layout that scales.

Most heavy users converge on the same structure inside their workspace folder. Building this once gives Cowork a stable mental model for everything you do.

# /Cowork (your workspace root) ├── CLAUDE.md # global rules for this workspace ├── memory.md # running learning log ├── about-me/ # identity, tone, context maps │ ├── about.md │ ├── writing-rules.md │ └── context-map.md ├── projects/ # per-project briefs & memory ├── outputs/ # where Cowork drops finished artifacts └── inputs/ # raw material in (receipts, transcripts…)
32
03

10 Enterprise Use Cases

From the morning command center to autonomous compliance - workflows real teams ship today.

33
Use case 01 · Executive Operations

The Morning Command Center.

A scheduled task runs at 6 AM: pulls every overnight email, scans the calendar, scrapes three industry sources, and drafts replies for the inbox-zero pile. The exec opens a single live-artifact dashboard with their day already triaged. Replaces the first 60 minutes of every workday.

~45 min
Reclaimed daily
3 sources
Auto-monitored
100%
Drafts, never sends
Workflow
Scheduled · 6:00 AM
1
Read Gmail · classify by urgency & topic
2
Pull today's calendar · note conflicts
3
Web-search 3 industry feeds for the exec's domain
4
Draft replies for the inbox-zero pile (in Gmail)
5
Render live-artifact dashboard, pin to sidebar
Stack: Gmail · Calendar · web search · Live Artifacts
34
Use case 02 · Finance & Accounting

Receipts → Categorized Expense Report.

Drop the month's receipts (PDFs and JPEGs) into a folder. Cowork OCRs every one, normalizes vendor names, classifies into your chart of accounts, flags anything below the audit threshold, and writes a finalized XLSX with a totals row and an exception sheet. Pair with a Stripe / Mercury connector and you have continuous reconciliation, not month-end panic.

100s
Receipts/run
~95%
Auto-categorized
Weekly
Cadence (was monthly)
Workflow
Skill · receipt-scanner
1
OCR every receipt in /receipts
2
Normalize vendor + map to chart of accounts
3
Cross-reference with Stripe / Mercury for matches
4
Flag anomalies (duplicate, >threshold, missing date)
5
Output: report.xlsx + exceptions.md, send to controller via Slack
Stack: file system · Stripe · Mercury · Slack · custom skill
35
Use case 03 · Cross-Functional Reporting

The 300-Word Weekly Leadership Update.

Friday at 4 PM, Cowork ingests three (or thirteen) team updates in incompatible formats - one in Notion, one in Google Doc, one as a Slack thread - extracts metrics, reconciles them to the OKRs in a tracking sheet, and produces a 300-word leadership PDF with three highlights, three lowlights, and the top-line numbers. Sent for review by 4:30.

3+
Source systems
300w
Hard cap
Friday
Auto, every week
Workflow
Scheduled · weekly
1
Pull updates: Notion DB · Google Doc · Slack canvas
2
Extract metrics, reconcile against OKR sheet
3
Draft: 3 top-line · 3 highlights · 3 lowlights
4
Render PDF in brand template
5
Email draft to chief-of-staff for review
Stack: Notion · Drive · Slack · Sheets · Gmail
36
Use case 04 · Revenue Operations

A Live Revenue Dashboard, Wired to Stripe.

A Live Artifact pinned in the sidebar. Each open pulls real-time Stripe data - MRR, ARR, active subs, average payment, revenue concentration. Layered with an outstanding-invoice tracker the team edits in-place. Replaces three separate dashboards and the "wait, what's the number?" Slack thread.

Live
Refresh on open
1 click
From sidebar
0
Tools to log into
Workflow
Live Artifact
1
On open: callMcpTool('stripe.charges.list')
2
Roll up by month, brand-deal vs. recurring
3
Surface anomaly (MoM > ±15%)
4
Render KPI grid + 12-month chart + invoice tracker
Stack: Stripe MCP · HTML/JS artifact
37
Use case 05 · Legal & Compliance

NDA & Contract Red-Flag Review.

Drop a contract PDF. Cowork loads the legal plugin, runs the contract-review skill, and returns: a one-page summary, key terms in plain English, red flags (clauses that hurt you), yellow flags (worth negotiating), missing standard terms, a sign / don't-sign verdict, and a marked-up redline. In-house counsel reviews; doesn't draft from scratch.

~10 min
Per contract (was hours)
100%
Get reviewed (vs. sample)
Audit-ready
Trail per doc
Workflow
Plugin · Legal
1
Classify: NDA · MSA · employment · vendor agreement
2
Match against your standard playbook (Drive)
3
Surface red/yellow flags + missing standards
4
Generate redline + negotiation script
5
File the audit-ready summary in /legal/reviews/
Stack: PDF skill · legal plugin · Drive · file system
38
Use case 06 · Strategy & Product Marketing

A Competitive Intelligence Tracker That Maintains Itself.

Daily scheduled run + Apify (or browser) for sources without APIs. Cowork monitors named competitors' release notes, pricing pages, blog, LinkedIn, and YouTube uploads. Diffs them against the prior snapshot. Surfaces only what changed, with a one-line "so what" and a recommended response action. Renders as a Live Artifact your CMO opens daily.

Daily
Auto-refreshed
N + 1
Add competitors instantly
"So what"
For every change
Workflow
Scheduled · daily
1
Crawl named-competitor sources (Apify + browser)
2
Diff vs. last snapshot stored in /comp-intel/
3
Triage by impact (pricing > product > talent)
4
Generate "so what + action" per change
5
Push to Slack #pmm-intel + render dashboard
Stack: Apify · Chrome ext · file diff · Slack · Live Artifacts
39
Use case 07 · Customer Operations

Tier-1 Support Triage on Autopilot.

Inbound tickets land in Zendesk. Cowork classifies, attaches the relevant runbook from your internal wiki, drafts the first response in your brand voice, and either routes to a human (with everything pre-loaded) or - for the well-known answer types - replies and closes. Pairs with a customized customer-support plugin per product line.

~70%
Auto-resolved at L1
< 2 min
First response time
24 / 7
Coverage
Workflow
Plugin · customer-support
1
New ticket → classify (issue type, severity, product)
2
Attach matching runbook from internal wiki MCP
3
Draft response in brand voice + attach diagnostics
4
If high-confidence + low-risk: send + close
5
Else: route to human, pre-loaded with everything
Stack: Zendesk MCP · internal-wiki MCP · custom plugin
40
Use case 08 · Executive Office

Board-Ready Decks From a One-Line Brief.

"Prep next week's board update." Cowork fans out four sub-agents in parallel: one pulls the quarterly financials from the data warehouse, one runs market-comp research on the web, one drafts the narrative in the CEO's tone (loaded from past memos), one renders the deck in the brand template. The CEO reviews, redlines once, ships.

Parallel sub-agents
~3 hrs
Down from 2 days
On-brand
Every time
Workflow
Sub-agents · parallel
A
Analyst → financials, KPIs, MoM/YoY commentary
B
Researcher → competitor & market context
C
Writer → narrative in CEO's voice from /memos/
D
Designer → renders deck in /templates/board.pptx
Merge → reviewer-ready PDF + editable PPTX
Stack: warehouse MCP · web · skills · sub-agents
41
Use case 09 · Knowledge Management

A Self-Maintaining Institutional Wiki.

Every meeting transcript, every PRD, every customer call gets ingested into a folder. Cowork extracts entities, builds an index, and links them - Karpathy-style. New hires query in plain English: "what do we know about Account X?" - Cowork follows the links and answers with citations. One reported team dropped query token usage by 95% switching from RAG to this pattern.

−95%
Query tokens vs RAG
Day 1
New-hire productivity
Compounds with use
Workflow
Skill · wiki-ingestor
1
Watch /raw → ingest new docs (transcripts, PRDs, calls)
2
Extract entities (people, accounts, products, decisions)
3
Update /wiki MD files + index + bidirectional links
4
Weekly lint pass: dedupe, fix stale refs, fill gaps
5
Query agent answers Qs by following links, with citations
Stack: file system · custom skill · scheduled lint
42
Use case 10 · Sales Engineering & Bid Desk

First-Draft RFP Responses in an Afternoon.

Drop the RFP PDF and the prospect's RFI. Cowork retrieves the matching answers from the answer-bank wiki (security questionnaire, SOC 2, prior wins), drafts the response in the company's voice, flags the gaps that need a human SME, and produces a final Word document in the brand template. Bid desks reclaim the days that used to be copy-paste.

~80%
Auto-answered
SME-flagged
Gaps explicit
2× more
Bids per quarter
Workflow
Project + skill
1
Parse RFP into discrete questions
2
Retrieve from answer-bank (Drive + Notion)
3
Draft answers in voice; cite sources internally
4
Mark gaps requiring a named SME
5
Render final DOCX in brand template
Stack: PDF skill · DOCX skill · Drive · Notion · Slack pings
43
Gallery
04

The Bigger Picture

Why Cowork compounds and how to roll it out without breaking anything.

44
The compounding effect

Cowork on day 30 is dramatically better than Cowork on day 1.

Every correction lands in memory.md. Every workflow can become a skill. Every skill can become part of a plugin. Every plugin shipped to a teammate makes the next teammate faster. This is not a tool you install - it's a system you build, and it gets better the more you use it.

Day 1

Generic agent. Same as anyone else's.

Day 30

Knows your tone, your folders, your tools, your edge cases.

Day 90

Drives entire roles. Onboards new hires in hours, not weeks.

45
A short field manual

What the heaviest users do differently.

Set up the foundation once

Global instructions. About-Me folder. Project per workstream. CLAUDE.md + memory.md per project. The 30 minutes you spend here pay back forever.

Prefer connectors → MCP → browser → computer use

Each step down is slower and more expensive. Always start with the cleanest interface available.

Reverse-engineer skills from real workflows

Run the workflow live with feedback. Then ask Cowork to codify it. Don't write skills in the abstract.

Plan first, then auto-accept

Get the plan good. Switch to auto-accept. Run multiple Cowork sessions in parallel - that's where the multiplier lives.

Always-on machine for scheduled tasks

A Mac mini on the desk turns Cowork from "tool" into "employee." Pair with Dispatch from your phone.

Drafts, not sends

Default Cowork to drafting outbound. You stay the approver. Trust grows; permissions widen on your timeline.

46
Where Avanai sits

We design the systems that turn Cowork into a teammate.

The model is given. The connectors are given. What's not given is the institutional knowledge - the skills, plugins, folder patterns, and scheduled jobs that make a generic agent into your agent. That's the work. That's where Avanai operates.

Cowork rollouts Custom skills & plugins MCP integrations Team enablement
47
Thank you

Let's build it.

Questions, demos, or a Cowork pilot for your team - avanai.io

Press Home to restart · F for fullscreen · ? for shortcuts

End
01 / 00