The best chief of staff doesn't wait to be asked.

Build an AI chief of staff on the model you already run. It preps your day, briefs you before every call, and answers like it works here. Because it does.

Get Started See the philosophy
Built for Even G2 Claude Code + Codex compatible
140+ COS Glasses users
In production at agencies and growth companies
5 min to a useful foundation

If AI is electricity, most people still haven't wired the house.

Open a chat. Type a prompt. Get a generic answer. Close the tab. Repeat tomorrow with zero memory of yesterday. That is not a system. It is a match struck in the dark.

The executives getting real leverage from AI are not writing better prompts. They are wiring the house, giving the model the context a world-class chief of staff would have on day 90:

  • Who reports to you, and how each person operates
  • What you are working on this quarter, and why it matters
  • Which meetings need prep, and which are noise
  • The patterns that keep showing up that you keep missing

COS gives you the wiring.

You stop briefing the AI. It starts briefing you.

Walk in already knowing.

A morning brief built from your calendar, your threads, and your people. Prep for every meeting without opening six tabs.

Daily brief · /prep
Ask like it works there.

Instant answers from everything your company knows: clients, decisions, docs, calls. Not generic advice. Your context.

Instant recall · your graph
Add capacity, not headcount.

Skills and automations that run without you. The follow-ups, the recaps, the tracking. Hours back every week.

Skills · scheduled work

“Add revenue without adding headcount.”The brief one agency gave us for their COS. It stuck.

One context graph. Every source you already use.

Not another app to move into. COS turns the tools you already run, your meetings, email, Slack, calls, and calendar, into one connected context graph. Structured, searchable, and yours.

Any source
Bring your own everything

Slack, Gmail, Granola, Fireflies, your CRM, your calendar. Connect through MCP or local ingestion. You choose what to copy, index, and keep.

Structured for you
Organized with review

Connected sources can be organized into people, decisions, and next steps. Low-confidence items stay visible for review.

Model your world
Your schema, not a template

Domains, deals, clients, personas, competitors. Shape the graph around how you actually work.

One graph, yours
Breadth and depth

Vector search to find it fast, a knowledge graph to explain how it connects. Your keys, your server.

Most AI apps keep what they learn inside the app. COS gives you the files.

Everything it knows about your work. Out in the open.

The context you choose to ingest becomes a living knowledge graph: people, companies, meetings, and decisions, extracted and connected. Search it, watch it compound, and keep the working corpus in files and indexes you control. Cloud models and source systems still follow their providers' terms.

Those counts are real, pulled from one running COS. The names in this demo are not.

Run it on your own work. Keep it portable.Open-source and self-hostable: your keys and storage you control, local or your cloud. Fork the stack and carry the working corpus with you.
Build your ownSee the philosophy

Five minutes to a useful foundation.

No setup gymnastics. Open Claude Code or Codex in an empty folder, paste the starter kit, and answer a few questions. It builds the foundation, then deepens as you use it.

View the kit Then paste it into Claude Code or Codex and hit enter.

What it does for you

01Interviews you

A few quick questions, then it writes your CLAUDE.md, the brain your AI loads every session.

02Builds your structure

Profiles your people and scaffolds your directory: operations, intelligence, weekly work.

03Installs your skills

/start, /learn, /prep. A daily briefing, correction memory, and meeting prep.

04Wires a connection

Slack, meetings, or any MCP. Optional, then the before/after that makes it click.

Rather have it done with you?Skip the DIY. We stand up your COS in a working session, then your team runs it.
Book a working session

A directory that is the system.

What you get

my-cos
my-cos/
├── CLAUDE.md                 # your COS brain (auto-generated)
├── context/
│   └── people.md             # your team profiles
├── .claude/commands/
│   ├── start.md              # /start: daily briefing
│   ├── learn.md              # /learn: capture corrections
│   └── prep.md               # /prep: meeting preparation
├── operations/
│   └── <domain>/             # tasks.md · intelligence/ · wk<NN>_<year>/
└── corrections.md            # patterns (auto-populated)

Every Claude Code session in this directory starts with your full context loaded. No re-explaining. Just: here is what matters today.

Not a setup. A system that deepens.

The four tiers

Tier 1 · 5 minutes

Foundation

Run the setup. Get your CLAUDE.md with role, team, and priorities. Consistent, role-aware responses.

Tier 2 · 30 minutes

Operational

Add people profiles. Configure your first skill. Connect one integration. A morning routine that surfaces what matters.

Tier 3 · 2 weeks

Compounding

Use /learn to capture corrections. Track tasks. Add meeting notes. The system starts connecting dots you did not ask about.

Tier 4 · 1+ month

Agentic

Scheduled tasks run automatically. Semantic search across every meeting. A system that works while you sleep.

Most people never get past Tier 1 with AI tools. COS pulls you forward.

Score yourself across 14 domains.

Not sure where to focus? Run /assess to grade your COS, end to end. Your lowest scores are your highest-leverage next steps.

Meeting IntelligenceCapture, classify, extract, search by meaning
Task OrchestrationExtract from conversations, dedup, stage, track
Email & Comms TriagePrioritize inbox, draft responses, follow-ups
Calendar AwarenessMeeting prep, conflict detection, time blocking
People IntelligenceProfiles, comms styles, relationship health
Cross-Source SynthesisConnect signals across meetings, email, tasks, Slack
Strategic AlignmentGoal tracking, resource allocation, priority scoring
Proactive AlertingEarly warnings, signal detection, anomaly flagging
Institutional MemorySession persistence, correction compounding
Delegation & Follow-upTrack who owes what, escalation cadence
Content & ResearchMarket intel, competitor tracking, documents
Reporting & DashboardsStatus updates, briefings, visual dashboards
Knowledge ManagementFile organization, templates, SOPs
Ambient I/OMulti-surface delivery: terminal, mobile, voice
See exactly where you stand.See your team's biggest gaps and what to fix first. 12 questions, about 3 minutes, right in your browser.
Get your Readiness Score

Wired to the tools you already live in.

COS connects through MCP servers. Bring your own.

SlackChannel monitoring, message context10 min
Google WorkspaceEmail, calendar, docs access15 min
GFMeeting ToolsGranola, Fireflies, Fathom sync10 min
Notion & DocsPages, wikis, project docs10 min
HubSpot & CRMDeals, contacts, page analytics20 min
Claude Code / CodexThe engine. Your existing plan0 min
QVector StoreSemantic search across all content30 min
∞Anything with an MCPThousands of servers. Bring your ownvaries
Rolling this out to a team?We'll help you get the whole team using it, not just the power users.
Book a working session

The same COS. Now on your glasses.

Every green screen on this page lives here: the Even G2 smart glasses. A chief of staff on the lens. The same context you've built, now it follows you off the screen.

Get Started Get it on the Even Hub View on npm
COS Glasses lens: an answer streaming in green on the heads-up display.
On the lens

Talk & listen

On-lens chat across Opus, Sonnet, Haiku, or Codex. Voice dictation with send / edit / re-record, answers read to your earbuds.

COS Glasses lens: meeting takeaways and action items in green.
Meeting-aware

It works while you talk

Live meeting capture transcribed locally, with speaker labels. Tasks, calendar, and people on the HUD.

COS Glasses lens: time, weather, and next meeting at a glance in green.
Yours to run

MIT, self-hosted

Your keys, your server, your context. An OpenAI-compatible endpoint drops it into Even's agent slot.

Buying AI is easy. Getting a team to use it is the real work.

For teams

COS gets each person to a useful answer from real work. Shared company context keeps the team aligned while personal setups stay personal. Start with one workflow, prove the value, then expand.

Production proof: COS runs in production at agencies and growth companies. One agency deployment uses role-based context, shared knowledge, scheduled indexing, and integrations with its operating stack, with a formal target of moving utilization from a 51% baseline to 70-80%.

Get your Readiness Score Book a working session

Keep the human as the head.

Frequently asked questions

COS is built on the centaur idea: you keep the direction and judgment. The machine handles the repetition, retrieval, and grunt compute.

Go deeper: context engineering vs prompt engineering explains why the prompt was never the problem, and AI chief of staff vs human chief of staff draws the line between what gets automated and what stays yours.

What is a centaur?

The human is the head and the machine is the body. You bring the domain expertise, the taste, and the judgment; the machine brings the grunt compute. The person using the system is the one who gets amplified, and the main benefactor is you. That is the bar any AI deployment should clear.

What is a reverse centaur?

The flip. The machine is the head and the human is demoted to peripherals: the hands and the eyes, plus the legal liability for whatever the algorithm decided. Warehouse pace-setting and rideshare routing are the classic examples. Throughput first, quality and people second. We think that is the wrong way to deploy AI.

Which one is COS?

A centaur by architecture, not by promise. COS keeps its working context in storage you control, local or your cloud, and answers to you. It preps, drafts, tracks, and connects; you decide. Cloud models and source systems still follow their providers' terms. It exists to 10x the person running it, never the other way around.

What happens when the AI platforms raise prices?

AI vendors sell tokens, not guaranteed outcomes. COS is portable across Claude Code and Codex today and designed for additional model adapters. The context you choose to ingest can live in files and indexes you control, so a pricing or model change does not force a rebuild. Swap the engine. Keep the system.

What is a context graph?

The context you choose to ingest from meetings, email, Slack, calls, docs, and people, extracted into one connected, searchable graph. Its working corpus can live in files and indexes you control.

Why does context beat prompts?

Same model, same prompt, completely different answer once the model actually knows your job. The prompt was never the problem; the missing context was. A model with your context stops being a chatbot and starts being a chief of staff.

Do I need to be technical?

If you can paste one file into Claude Code or Codex, you are technical enough. The starter kit interviews you and builds a useful foundation in about five minutes. Want it done with you instead? Book a working session and we will stand it up together.

What does it cost to run?

The kit is free. Chat runs on the Claude or Codex plan you already pay for. Local voice on the glasses has no per-minute fee. There is no COS subscription, and your files stay yours.

The philosophy

Context compounds. Every correction, every meeting, every person you profile accumulates into a system that knows your work better than any prompt could describe.

The interactive philosophy map traces the four tiers:

  • Foundation: identity is the starting point
  • Operational: routines create rhythm
  • Compounding: corrections accumulate into intelligence
  • Agentic: the system acts without being asked
Start building