Cybos

CYBernetic Operating System - AI-powered VC operations assistant. Research companies, process Telegram/email, generate content, write investment memos.

Setup

See SETUP.md for full setup and personalization.

# Install dependencies
curl -fsSL https://bun.sh/install | bash
cp .env.example .env  # Add API keys, DB config, and CYBOS_USER_NAME

# Start database
cd scripts/db && docker compose up -d && cd ../..
bun scripts/db/init.ts
bun scripts/db/index.ts --extract

# Run
claude

First run: Database indexes entities from /deals/, /context/calls/, /context/telegram/. Run /cyber-reindex to rebuild.

Available slash commands:

/cyber-research-company "Company Name" - Company DD research
/cyber-research-tech "Technology" - Technology deep-dive
/cyber-research-market "Market" - Market analysis
/cyber-telegram - Process 1 unread Telegram message (read, draft reply, move to TO ANSWER)
/cyber-telegram --count 3 - Process 3 unread Telegram messages
/cyber-browse - Scan Twitter for trending topics (saves to /content/ideas/)
/cyber-tweet "Topic" or /cyber-tweet @content/ideas/file.md - Draft tweet (with optional source)
/cyber-essay "Topic" or /cyber-essay @content/ideas/file.md "expand" - Write essay (with optional source)
/cyber-image "Concept" - Generate image
/cyber-memo "Company" - Generate investment memo
/cyber-init-deal "Company" - Initialize deal folder
/cyber-log - Show recent activity
/cyber-gtd - Process GTD items autonomously (plan-first by default)
/cyber-gtd --count 3 - Process 3 GTD items
/cyber-reindex - Rebuild entity index from deals/calls/entity files/telegram logs
/cyber-email --sync - Sync emails to /context/emails/ (last 3 days)
/cyber-calendar - Show today + tomorrow calendar events
/cyber-unstuck - Interactive focus ritual for breaking distraction loops
/cyber-init-project "Name" - Initialize a project folder
/cyber-project <slug> - Show project status and tasks
/cyber-projects - List all projects with status
/cyber-gtd --project <slug> - Process GTD tasks for specific project

Architecture

Core Components

CYBOS
├── SKILLS (organized workflows)
│   ├── Research (company, tech, market)
│   ├── Browse (Twitter feed discovery)
│   ├── Telegram (message processing)
│   ├── Content (tweets, essays, images)
│   ├── DDMemo (investment memos)
│   └── GTD (autonomous task execution)
│
├── AGENTS (parallel execution via Task tool)
│   ├── Researchers: company, market, financial, team, tech
│   ├── Content: content-writer, image-prompter
│   ├── Memo: memo-analyst, memo-writer
│   └── Synthesizer: consolidates research
│
├── HOOKS
│   └── SessionStart: load identity & context
│
├── MCP SERVERS (external tools)
│   ├── perplexity, parallel-task, exa (PRIMARY)
│   ├── parallel-search (fallback), playwright
│   ├── claude-in-chrome (browser automation)
│   ├── notion, nano-banana
│
└── CONTEXT (identity, philosophy, brand)

Key Design Principles

File-first: All state is markdown on disk
Scaffolding > prompting: Workflows + tools beat raw prompts
Parallel agents: Multiple Task calls in single response
Context auto-loading: Deal context loads when company mentioned
Single-file logging: One log file per day

Usage Guide

Research Workflows

Company Research

# Initialize a new deal
/cyber-init-deal "Acme Corp"

# Research the company
/cyber-research-company "Acme Corp"

What happens:

Parallel MCP calls gather data (Parallel AI, Perplexity, Exa)
4 parallel agents analyze (company, market, financial, team)
Synthesizer consolidates with investment lens
Report saved to /deals/acme-corp/research/MMDD-<slug>-YY.md
Deal context created/updated
Action logged
Duration: 4-8 minutes

Output:

Comprehensive research report with investment analysis
Deal context with status and key facts
Log entry

Technology Research

/cyber-research-tech "Trusted Execution Environments"

What happens:

Deep technical research via MCP tools
3 parallel agents (tech, market, financial)
Synthesis with focus on investment opportunities
Report saved to /research/trusted-execution-environments/MMDD-<slug>-YY.md
Duration: 5-8 minutes

Focus areas: TEEs/privacy, AI agents, robotics data, compute financialization

Market Research

/cyber-research-market "AI Infrastructure"

What happens:

Market sizing and dynamics research
3 parallel agents analyze market, technology, investment activity
Report includes TAM, key players, white spaces, timing
Saved to /research/ai-infrastructure/MMDD-<slug>-YY.md
Duration: 8-12 minutes

Content Workflows

Tweets

/cyber-tweet "AI agents using crypto wallets"

Workflow: Draft → Review → Polish

Draft: Loads voice-identity.md + writing-style-en.md, creates tweet
Review: User provides feedback or approves
Polish: Revise based on feedback
Output: Save to /content/tweets/MMDD-<slug>-YY.md

Style: Hook-driven, one idea per tweet, specific claims with evidence

Duration: 2-5 minutes

Essays

/cyber-essay "Why compute is becoming a commodity"

Workflow: Draft → Review → Polish

Draft: Loads voice-identity.md + writing-style-en.md, creates essay (500-2500 words)
Review: User feedback on structure/tone
Polish: Finalize with revisions
Output: /content/essays/MMDD-<slug>-YY.md

Structure: Hook → Stakes → Mechanism → Turn (optional) → Landing (forward momentum)

Duration: 15-30 minutes

Images

/cyber-image "Solitary figure in brutalist VC office"

Workflow: Draft Prompt → Review → Generate

Draft Prompt: image-prompter creates detailed prompt following Mural of Souls aesthetic
Review: User approves or adjusts prompt
Generate: Nano Banana MCP (mcp__nano-banana__generate_image) using Gemini 3 Pro Image
Iterate: Use continue_editing for refinements
Output: /content/images/MMDD-<slug>-YY.png (no metadata files)

Aesthetic: Mural of Souls - sacred transformation, particle dissolution, monumental scale

Duration: 3-7 minutes

Source-Driven Content (New Feature)

The Ideas Folder: /content/ideas/ stores raw thoughts, notes, and fragments that can be expanded into polished content.

Quick Capture Workflow

# 1. Save raw idea to /content/ideas/
# Create file: /content/ideas/agent-economy.md
# Content: Bullet points, rough paragraphs, data, quotes - no polish needed

# 2. Expand into essay
/cyber-essay @content/ideas/agent-economy.md "Expand into full essay with examples"

# 3. Or distill into tweet
/cyber-tweet @content/ideas/agent-economy.md "Create tweet from key insight"

How It Works

Source Material Pattern: Commands now support @-prefixed file/folder references:

# Single source
/cyber-essay @content/ideas/tee-compute.md "Expand this"

# Multiple sources (synthesize)
/cyber-essay "TEE market analysis" @content/ideas/tee.md @research/tee-market/

# From research to tweet
/cyber-tweet @deals/acme-corp/research/2025-12-28.md "Distill key insight"

# Thread from folder
/cyber-tweet @research/ai-infrastructure/ "Create thread summarizing findings"

Workflow:

Capture: Save rough ideas to /content/ideas/ (no format required)
Reference: Use @path/to/file.md in /cyber-essay or /cyber-tweet
Expand: Agent reads source, extracts insights, expands with structure/examples
Output: Polished content in /content/essays/ or /content/tweets/

Benefits:

Capture ideas without worrying about polish
Build library of reusable thoughts
Synthesize multiple sources into coherent content
Transform research into public-facing content

See: /content/ideas/README.md for detailed guide and _template.md for starter template.

GTD Runner (Autonomous Task Execution)

# Plan and show what will be done for first GTD item
/cyber-gtd

# Process 3 items
/cyber-gtd --count 3

# Skip plan, execute immediately
/cyber-gtd --execute

# Process specific item
/cyber-gtd "Dan Meissler"

# Rebuild entity index
/cyber-reindex

What happens:

Parse: Read GTD.md, extract items from # Next
Classify: Match patterns to workflows (outreach, call-prep, podcast, research)
Entity Lookup: Query database for people/company context
Plan: Show what will be done for each item
Execute: Run each task sequentially (one at a time)
Output: Save to /content/work/MMDD-<slug>.md with draft + pending actions

Workflows:

Pattern	Workflow	Action
"ask for call", "message", "email"	outreach	Draft message, find contact
"call with", "meeting", "<> X"	call-prep	Prepare agenda, questions
"podcast"	podcast	Research guest, prepare questions
company name, "research"	research	Quick research summary
unknown	best judgment	Log to learnings.md

Entity System:

Entities (people, companies) are indexed in PostgreSQL database from:

/deals/*/ folder names (orgs)
/context/calls/ attendees (people with emails)
/context/entities/*.md (manual enrichment with aliases, notes)
/context/telegram/*.md conversation files (people with telegram usernames)

Query entities: bun scripts/db/query.ts find-entity "<name>"

Index rebuilt automatically every 24 hours or via /cyber-reindex.

Output Format:

Each task produces /content/work/MMDD-<slug>.md with:

Context summary (entity info, previous calls)
Draft content (message, agenda, questions)
Pending actions (checkboxes for Gmail send, etc.)
Execution log

Duration: 1-5 minutes per task

DD Memo Generation

# Prerequisite: Run company research first
/cyber-research-company "Acme Corp"

# Generate investment memo
/cyber-memo "Acme Corp"

What happens:

Gather: Load all research from /deals/acme-corp/research/
Analyze: memo-analyst (Opus) applies investment rubric
Write: memo-writer (Sonnet) fills template
Output: /deals/acme-corp/memo/memo.md (overwrites)
Duration: 8-15 minutes

Output: Comprehensive investment memo with:

Executive summary & investment thesis
Scoring sheet (10 categories)
Product, business model, technology analysis
Team assessment, financials, projections
Risks, exit scenarios, IC Q&A
Clear recommendation (INVEST/PASS/MORE DILIGENCE)

Project System

Projects are multi-week initiatives that don't fit into /deals/ (external companies) or single GTD tasks. Examples: organizing events, running accelerator, building internal products.

# Initialize a new project
/cyber-init-project "Cyber Accelerator Q1"

# View project status
/cyber-project cyber-accelerator-q1

# List all projects
/cyber-projects

# Work on project-specific GTD tasks
/cyber-gtd --project cyber-accelerator-q1

What happens:

Initialize: Creates /projects/cyber-accelerator-q1/.cybos/context.md from template
GTD Integration: Add # cyber-accelerator-q1 heading to GTD.md, tasks underneath belong to project
Auto-loading: Mention project name → context loaded automatically
Progress tracking: Context file tracks status, goal, key results, milestones

GTD.md Integration:

# cyber-accelerator-q1
- [ ] Finalize speaker lineup
- [ ] Create participant application form
- [x] Book venue

# scheduler
- [ ] Implement calendar sync
- [ ] Add recurring event support

Projects are identified by # heading (level-1). Reserved headings (# Next, # Someday, # IC, # Skip) are not projects.

Project Types:

Event: Conference, demo day, meetup
Accelerator: Cohort program
Product: Internal tool development
Initiative: Multi-step strategic effort

Duration: Initialization takes seconds, projects span weeks/months

Investment Philosophy

Cybos applies your organization's investment rubric to all research:

Decision-Making Criteria

Criteria	Green Flag	Red Flag
Market Size	Path to $1B+ revenue	Niche $50M ARR cap
Moat	Data/network/hard tech	Wrapper (6-week rule)
Business Model	Clear revenue model	Token speculation
Founders	High energy, sales DNA, deep expertise	Low energy, lab mindset
Why Now?	Clear catalyst/unlock	Red ocean, no timing
Valuation	Reasonable for stage	>$100M FDV pre-revenue

Auto-Pass Triggers

Can Big Tech build this in 6 weeks?
Pure "wrapper" with no moat
Media/entertainment robotics plays
Regional stablecoins vs global giants
Generic devtools without massive pain point

Focus Areas

AI Infrastructure: Training, inference, privacy (TEEs), agents
Crypto/Blockchain: Self-custodial finance, programmable markets, novel consensus
Robotics: Data moats (teleoperation, simulation), embodied AI

File Structure

cybos/
├── .mcp.json                          # MCP server configuration (uses ${VAR} for env vars)
├── .claude/
│   ├── settings.json              # Hook wiring, permissions
│   ├── skills/                    # Organized workflows
│   │   ├── CORE/                  # System overview
│   │   ├── Research/workflows/    # company, tech, market
│   │   ├── Browse/workflows/      # twitter-feed
│   │   ├── Telegram/workflows/    # answer-messages
│   │   ├── Content/workflows/     # tweet, essay, image
│   │   ├── DDMemo/workflows/      # generate
│   │   └── GTD/                   # autonomous task execution
│   │       ├── SKILL.md           # main skill entry point
│   │       ├── workflows/         # outreach, call-prep, podcast, research
│   │       └── learnings.md       # action log
│   ├── agents/                    # 10 agent profiles
│   ├── hooks/
│   │   └── load-context.ts        # SessionStart hook
│   └── commands/                  # 11 slash commands
│
├── .cybos/logs/                # Daily log files (MMDD-YY.md)
│
├── context/                       # Core context
│   ├── who-am-i.md
│   ├── organization.md
│   ├── investment-philosophy.md
│   ├── style/                     # Writing style guides
│   │   ├── voice-identity.md      # Shared persona, tone (all content)
│   │   ├── writing-style-en.md    # English style (essays, tweets)
│   │   └── writing-style-ru.md    # Russian style (Telegram)
│   ├── img-style.md
│   ├── MEMO_template.md           # DD memo template
│   ├── telegram/                  # Per-person conversation logs
│   │   ├── README.md
│   │   └── <person-slug>.md       # Persistent conversation file
│   └── entities/                  # Entity context (indexed in PostgreSQL)
│       ├── README.md              # Usage documentation
│       ├── people/                # Manual entity files (optional)
│       └── orgs/                  # Manual entity files (optional)
│
├── deals/                         # Deal folders
│   └── <company-slug>/
│       ├── .cybos/
│       │   ├── context.md         # Deal metadata, status
│       │   └── scratchpad/        # Agent working files
│       ├── research/              # Research reports
│       └── memo/                  # Investment memo
│
├── research/                      # Topic/market research
│   └── <topic-slug>/
│
├── projects/                      # Multi-week initiatives
│   └── <project-slug>/
│       └── .cybos/
│           └── context.md         # Project context, status, goals
│
├── content/                       # Generated content
│   ├── ideas/                     # Raw ideas, notes (source material)
│   │   ├── README.md              # Usage guide
│   │   └── _template.md           # Idea template
│   ├── tweets/
│   ├── essays/
│   ├── images/
│   └── work/                      # GTD task outputs (MMDD-<slug>.md)
│
├── .env.example                   # API key template
└── .gitignore                     # Excludes sensitive data

MCP Integration

Tiered Research Strategy

1. FAST SEARCH (seconds)
   └─ perplexity search OR exa search
   └─ Good for: quick facts, validation

2. DEEP RESEARCH (1-5 minutes)
   └─ parallel-task createTask → poll getTask
   └─ Good for: comprehensive reports

3. TARGETED EXTRACTION (as needed)
   └─ exa getContents (PRIMARY)
   └─ Fallback: parallel-search web_fetch
   └─ Last resort: firecrawl scrape

4. HARD SCRAPING (last resort)
   └─ playwright for JS-heavy sites

MCP Tool Reference

Perplexity:

mcp__perplexity__search - Fast search with citations
mcp__perplexity__research - Deep research (basic/deep)

Parallel Task:

mcp__parallel-task__createTask - Start deep research
mcp__parallel-task__getTask - Poll for results

Exa (PRIMARY for URL content):

mcp__exa__search - Web search
mcp__exa__findSimilar - Similar pages
mcp__exa__getContents - Extract content from URLs (PRIMARY)
mcp__exa__linkedinSearch - LinkedIn profiles

Parallel Search (FALLBACK):

mcp__parallel-search__web_search_preview - Web search fallback
mcp__parallel-search__web_fetch - URL content fallback

Firecrawl (LAST RESORT ONLY - use when exa and parallel-search fail):

mcp__firecrawl__scrape - Scrape URL
mcp__firecrawl__extract - Structured extraction

Nano Banana:

mcp__nano-banana__generate_image - Generate images (Gemini 3 Pro Image)
mcp__nano-banana__edit_image - Edit existing images
mcp__nano-banana__continue_editing - Iterative refinement

Agent System

Available Agents

Agent	Model	Purpose
company-researcher	Haiku	Company business model, product, traction
market-researcher	Haiku	TAM, dynamics, competitive landscape
financial-researcher	Haiku	Funding history, metrics, valuation
team-researcher	Haiku	Founder backgrounds, assessment
tech-researcher	Haiku	Technology deep-dives, moat analysis
content-researcher	Haiku	Topic research for content (academic, social, first-principles)
investment-researcher	Haiku	Topic research for investment (market dynamics, opportunities)
quality-reviewer	Sonnet	Gap analysis, quality assurance (deep mode only)
content-writer	Sonnet	Tweets, essays following brand
image-prompter	Sonnet	Image prompt generation
memo-analyst	Opus	Strategic investment analysis
memo-writer	Sonnet	Memo generation from template
synthesizer	Sonnet	Consolidate parallel research

Research agent selection: Agents are selected dynamically based on research type (Company, Tech, Market, Topic) and intensity (Quick, Standard, Deep). See .claude/skills/Research/shared/agent-selection-matrix.md for the complete selection matrix.

Parallel Execution

To run agents in parallel, issue multiple Task calls in the same response:

Task 1: company-researcher
Task 2: market-researcher
Task 3: financial-researcher
Task 4: team-researcher

Claude Code executes these simultaneously for maximum speed.

Logging System

All workflows log to /.cybos/logs/MMDD-<slug>-YY.md (single file per day).

Format:

## HH:MM | category | type | subject
- Workflow: name
- Duration: Xm Ys
- Output: /path/to/output.md
- Agents: (if used)
- Sources: (if used)

---

View logs:

/cyber-log           # Today's activity
/cyber-log week      # Last 7 days
/cyber-log 2025-12-19   # Specific date

Troubleshooting

Hook Not Loading

Symptom: Context not injected at session start

Solution:

# Verify hook is executable
chmod +x .claude/hooks/load-context.ts

# Test hook manually
bun .claude/hooks/load-context.ts < /dev/null

MCP Server Not Found

Symptom: mcp__[server]__[tool] is not available

Solution:

Check API key in .env
Verify .mcp.json configuration (uses ${VAR} syntax for env vars)
Test MCP server independently:

npx -y @perplexity-ai/mcp-server --help

Agent Times Out

Symptom: Task agent doesn't complete

Solution:

Reduce scope of agent task
Use Haiku for simpler tasks (faster)
Opus is slower but provides strategic depth

Research Returns Limited Data

Symptom: Research report is sparse

Solution:

Run research workflow again (MCP services may have been slow)
Try different MCP tools (Exa vs Perplexity)
Check API quotas/limits

Image Generation Fails

Symptom: Nano Banana returns error

Solution:

Check GEMINI_API_KEY is set in environment
Simplify prompt if safety checker triggered
Use continue_editing to iterate on partial results

Best Practices

1. Research Before Memo

Always run company research before generating a memo:

/cyber-research-company "Company"
# Wait for completion
/cyber-memo "Company"

2. Use Parallel Agents

Workflows automatically use parallel agents for speed. Don't try to serialize them.

3. Iterative Content

Content workflows expect iteration. Provide specific feedback:

"Good" or "ship it" = approved
"More technical" = adjust tone
"Different hook" = revise opening

4. Update Deal Context

When new information emerges, update /deals/<company>/.cybos/context.md manually and regenerate memo.

5. Monitor Logs

Check /cyber-log regularly to track what's been done and what's pending.

6. Use Ideas Folder

Capture rough thoughts in /content/ideas/ and expand later:

# Quick capture
echo "# Agent economy\n\n- AI agents need wallets\n- Crypto enables M2M payments" > content/ideas/agent-econ.md

# Expand when ready
/cyber-essay @content/ideas/agent-econ.md "Full essay with examples"

7. Version Control

Commit regularly, especially after research completion:

git add .
git commit -m "Research: Acme Corp"

Development

Adding a New Workflow

Create workflow file: .claude/skills/<Skill>/workflows/<workflow>.md
Create slash command: .claude/commands/cyber-<command>.md
Reference workflow with @.claude/skills/... in command
Test with /cyber-<command>

Adding a New Agent

Create agent profile: .claude/agents/<agent-name>.md
Define purpose, output format, guidelines
Reference in workflow as Task: <agent-name>

Modifying Investment Rubric

Edit context/investment-philosophy.md - all workflows automatically apply updated criteria.

Customizing Brand Voice

Content follows a layered style system:

Shared persona: context/style/voice-identity.md - archetype, tone, anti-patterns (all languages)
English style: context/style/writing-style-en.md - essay/tweet structure, rhythm, banned words
Russian style: context/style/writing-style-ru.md - Telegram posts, Russian-English mix

Edit the appropriate file based on what you want to change. The content-writer agent and all workflows reference these files.

Updating Visual Aesthetic

Edit context/img-style.md - image-prompter agent uses this aesthetic system.

Roadmap

MVP (Current)

Future Enhancements

CLI wrapper (Bun/TypeScript) for headless execution
Scheduled automation (cron jobs)
Voice notifications (ElevenLabs)
Call prep skill (meeting briefings)
Portfolio monitoring (ongoing tracking)
Team collaboration features

License

MIT License with Attribution Requirement

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
.claude		.claude
.cybos/logs		.cybos/logs
config		config
content		content
context		context
deals		deals
docs		docs
projects		projects
research		research
scripts		scripts
.env.example		.env.example
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
GTD.md		GTD.md
README.md		README.md
SETUP.md		SETUP.md
bun.lock		bun.lock
package-lock.json		package-lock.json
package.json		package.json

Folders and files

Latest commit

History

Repository files navigation

Cybos

Setup

Architecture

Core Components

Key Design Principles

Usage Guide

Research Workflows

Company Research

Technology Research

Market Research

Content Workflows

Tweets

Essays

Images

Source-Driven Content (New Feature)

Quick Capture Workflow

How It Works

GTD Runner (Autonomous Task Execution)

DD Memo Generation

Project System

Investment Philosophy

Decision-Making Criteria

Auto-Pass Triggers

Focus Areas

File Structure

MCP Integration

Tiered Research Strategy

MCP Tool Reference

Agent System

Parallel Execution

Logging System

Troubleshooting

Hook Not Loading

MCP Server Not Found

Agent Times Out

Research Returns Limited Data

Image Generation Fails

Best Practices

1. Research Before Memo

2. Use Parallel Agents

3. Iterative Content

4. Update Deal Context

5. Monitor Logs

6. Use Ideas Folder

7. Version Control

Development

Adding a New Workflow

Adding a New Agent

Modifying Investment Rubric

Customizing Brand Voice

Updating Visual Aesthetic

Roadmap

MVP (Current)

Future Enhancements

License

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages