Drowning in articles? Let AI agents crawl, organize & analyze for you.
Personal knowledge system that builds a searchable graph from RSS, papers & GitHub — updated daily, driven by one YAML config.
🎬 Full-quality 1080p MP4 · 60fps · 30s · Origin story · Why it’s defensible (moat)
Collect → Structure → Analyze → Act. Each agent deploys independently, but data flows seamlessly between them. Swap one YAML config to retarget the entire system to a new domain.
Extracts entities from articles, papers, and repos. Manages a Neo4j ontology and provides GraphRAG-powered Q&A.
core | Entity extraction, dedup detection, claim tracking, MetaEdge |
crawler | HN/arXiv/GitHub crawlers, ingest pipeline |
graphrag | Dual-Retriever, subgraph exploration, answer synthesis |
ingester | ear archive → graph importer |
mcp-server | MCP protocol server (21 tools) |
Key papers indexed: Attention Is All You Need, GPT-3/4, BERT, LLaMA 1~4, DeepSeek-R1/V3, Gemini, Chain-of-Thought, LoRA/QLoRA, FlashAttention, Mamba, DPO, GraphRAG, LightRAG, SAM, CLIP, ReAct, MemGPT, and more.
34 sources (default AI/ML preset): 31 RSS feeds (OpenAI/Anthropic/Google/Meta/xAI/DeepSeek blogs + researcher feeds) + HN API + arXiv API + GitHub API
Recall 83% → 93%, Latency 20.7s → 13.8s. Grounding Rate is a hallucination-resistant metric: the fraction of entities cited in the answer that actually exist in the graph.
"What is MCP and which companies are using it in production?"
Extracts entities from text, runs propagation simulations, generates comprehensive reports through 10 analysis frameworks.
Input: 2 sentences → Output: 810-line, 15-section report
Key finding: "MCP serves as bridge node across 3 communities (confidence 70%)"
/analysis and /report are server components that inline the full analysis and report markdown into the initial HTML (hidden for sighted users, read by AI crawlers).
Per-page generateMetadata generates title/description from actual findings. OpenGraph, Twitter Card, and JSON-LD (SoftwareApplication) emitted from the root layout.
robots.txt explicitly allows GPTBot, ChatGPT-User, ClaudeBot, Claude-Web, PerplexityBot, Google-Extended.
E2E measurement: /report?job=… initial HTML body 18 B → 32,297 B (×1,794) — paste a URL into any AI and it reads the full report.
Each focused on its role, all connected to the whole system.
Monitors feeds and auto-archives relevant content
Image analysis → correction plan → Photoshop MCP non-destructive editing
Self-learning memory graph: kb_facts + embeddings + auto-consolidation loop
Natural-language workflow automation for Claude Code
Token-efficient headless browser for AI agents
find, batch multi-step flows in one round-trip.--session flag. Multi-tab, advanced waits.Self-evolution harness — 8 LaunchAgents close the harvest→learn→decide→apply loop
Comad World ships with ready-made presets. Swap one file to retarget the entire knowledge pipeline to your field.
Requires: Node.js 20+, Python 3.13+, Neo4j 5, Bun. The install script handles dependency setup and Neo4j schema init. For scheduled jobs, run brain/scripts/schedule-install.sh — it auto-detects macOS (LaunchAgents), Linux/WSL (cron), or Windows (Task Scheduler). See the README.
Under the hood: scripts/upgrade.sh aborts on dirty trees, snapshots comad.config.yaml / .env / ~/.claude/agents/ to .comad/backups/<ts>/, fast-forwards main + 6 module repos, reinstalls deps, and prints a per-module timed summary.
VERSION + comad.lock pin every module SHA like package-lock.json.
📂 Clone anywhere — the repo is path-agnostic. scripts/comad follows its symlink to derive the repo root, scripts/render-templates.sh rewrites {{COMAD_ROOT}} in *.example files at install time, so any folder/any name just works (~/Desktop/foo, /opt/comad, …). No /Users/<author> hardcodes.
/analysis & /report (initial HTML 18 B → 32 KB) · comad upgrade/status/rollback global command · VERSION + comad.lock · clone the repo to any folder, any namecomad-shared.css · black background, blue + warm orange palette · glassmorphic top bar across bothcomad_eye/ · fetchWithTimeout + 45m job deadline unblock a 4h ear-ingest hang · CI now runs the full 1,422-test suite.git archived, single mono-repo restored · ADR 0007–0010 promote each gap packcom.comad.ear-poll, 15 min) replaces the crash-looping ccd KeepAlive wrapper that was burning Discord’s 1,000/day IDENTIFY quota in ~8h · per-run claude -p --strict-mcp-config ensures zero MCP spawncom.comad.cron-catchup replays any LaunchAgent that would have fired while the laptop was asleep, so Monday analysis still runs even if you opened the lid at noon · docs/cron-catalog.md documents all 11 agents with dependencies and missing-run recoverycomad-parallel with five comad-native gates: handoff (auto session doc), qa-gate (.qa-evidence.json verdict), second-opinion-gate (.second-opinion.md verdict), destroy-check (13 destructive patterns post-scan), ear-notify (Discord webhook). T6 self-evolve coupling — parallel worker fix:/feat: commits auto-captured by existing Stop hookcomad-image (7-mode Codex /imagen), comad-motion (HTML→Playwright→ffmpeg, BGM/SFX bundled CC-BY 4.0), comad-pptx (LAYOUT_WIDE deck), comad-infographic (4 patterns), comad-app-prototype (iOS/Android/macOS/Browser frames) · one ./install.sh + ./doctor.sh · 47 MB w/ assetscomad-ci-healer (GH Actions failure → classify → headless claude -p fix → auto PR → Discord notify, dry_run+allowlist guards) and comad-pr-review (4-axis correctness/security/performance/convention scoring → inline+summary comments, codex+claude, headSha dedup) · both live E2E-verified · comad-world-extensions skills 5→8코마드월드는 자가진화 가드와 산출물 스튜디오까지 합쳐 한 시스템을 이룹니다. 어느 레포만 깔아도 동작하고, 셋 다 깔면 전체 워크플로우가 닫힙니다.
8개 핵심 모듈 — brain · ear · eye · photo · sleep · voice · search · browse
조용한 가드 — 9 hooks + 12 skills (자가진화 / QA 증거 / 메모리 / 병렬 외주 / CI 자가복구 / PR 리뷰 / 측정 / brain 활용)
fix:/feat: 커밋 자동 포착 → /comad-learn 승격) · Approval-Gated Destruction (rm -rf, git push --force sha256 승인) · Proof-Driven QA (.qa-evidence.json + .second-opinion.md) · Codex 병렬 외주 (Claude=PM, Codex×N=구현) + 5종 통합 게이트 · CI 자가복구 상시 에이전트 (GH Actions 실패 → headless 수정 → 자동 PR) · 4축 Autonomous PR Reviewer · comad-sdd (Spec-Driven Development 닫힌 루프 — 완료기준→구현→대조) · harness-report (5축 하네스 점수 + 비용/efficiency 측정, 품질 composite와 분리) · comad-recall·comad-foresight (⚠️ comad-brain 필요 — 출처인덱스 +3.12 / 10렌즈 전략 foresight +1.375, 측정).산출물 스튜디오 — 5 visual/media 생성 스킬 (image · motion · pptx · infographic · app-prototype)
./doctor.sh 의존성 매트릭스 진단.
세 레포 모두 독립 설치 — 각자 ./install.sh. 의존성 공유 없이 기능별로 골라 설치 가능.
설치 순서는 world → extensions → studio 권장 (extensions 가 world 의 디렉터리 구조를 가정).