Auset Brain

cli-anything-strategy

4 min read

CLI-Anything — Self-Expanding Agent Toolbox (March 12, 2026)

The Vision: 1,000,000 CLI Tools

  • Goal: 1 million CLI tools built by our agents, used by our agents
  • Why: This is how we beat the competition. More tools = more capable agents = faster delivery = lower cost
  • How: Agents autonomously build CLIs for any open source tool they need using CLI-Anything
  • Source: https://github.com/HKUDS/CLI-Anything (HKUDS — Hong Kong University, makers of LightRAG)

Why CLI > MCP

  • CLIs live in the terminal — no middleman, no overhead, no config
  • Claude Code / build farm agents already live in the terminal
  • MCP servers consume memory, need health checks, require config
  • CLIs are just executables — zero runtime cost when not in use
  • On EC2 farms (8GB RAM): CLI approach = more concurrent agents

The Self-Expanding Pattern (NON-NEGOTIABLE for all agents)

Agent encounters task requiring tool X
→ Check if CLI for X exists in /tools/cli/
→ If NO: run CLI-Anything against X's source code (~20 min)
→ CLI generated, tested, documented automatically
→ Store in /tools/cli/ — synced to ALL farms via git
→ Agent proceeds with task using new CLI
→ Every agent on every farm now has that tool forever

Rule for ALL agents: If you need a tool to make you more productive, use CLI-Anything to build it. Don’t wait for a human to install it. Build it yourself.

Priority CLIs to Generate First

  1. draw.io — Architecture diagrams, data flow, ERDs (auto-generated from text)
  2. Blender — 3D renderings for QuikAction construction projects (Quik’s #1 want)
  3. Audacity — Audio processing for Heru Feedback voice, Heru Discovery recordings, in-app messaging
  4. Inkscape — SVG/vector graphics for logos, marketing assets
  5. FFmpeg (if not already CLI-wrapped) — Video processing for Heru Discovery, Remotion output
  6. ImageMagick — Image processing, thumbnails, format conversion
  7. Pandoc — Document conversion (markdown → PDF/DOCX/PPTX)
  8. Puppeteer/Playwright — Browser automation (already have some, but CLI wrapper adds flexibility)

Blender + QuikAction LLC (Quik’s Construction Company)

  • QuikAction LLC = Quik’s construction company, another Heru
  • Website: quikactionllc.com (in GitHub org: imaginationeverywhere)
  • Quik’s dream: 3D renderings of construction projects without waiting weeks for graphic artists
  • Solution: Blender CLI via CLI-Anything → agents generate 3D renderings from descriptions/blueprints
  • Typical cost: 5,000 per rendering from a graphic artist, days/weeks turnaround
  • Our cost: Near-zero, same-day turnaround
  • Surprise Quik: Generate a 3D rendering of one of his construction projects before he asks
  • Maat Agents angle: Every contractor Quik knows = potential Maat Agent client

Audacity + Audio Pipeline (Shared Across All Herus)

  • Heru Feedback: Voice narration with screenshots/recordings — noise reduction, normalization
  • Heru Discovery: Voice/video product intake — transcription prep, audio cleanup
  • QuikCarRental: In-app voice messaging between renters and owners
  • Vapi: Post-call audio processing, archival
  • One audio pipeline, shared across every Heru (Ausar Engine pattern)

Daily CLI Harvesting Strategy (NON-NEGOTIABLE)

Target: 500 CLIs per day minimum. Every day. Starting day 1.

One dedicated agent runs 24/7 on the build farm with this mission:

  1. Discover — Crawl GitHub trending, awesome-lists, topic searches, dependency trees. Find repos with CLI potential (has installable software, has open source code, has useful functionality).
  2. Filter — Skip repos that are: too small (<100 stars or no real functionality), already have a good CLI, duplicates of existing CLIs, malicious/abandoned.
  3. Generate — Run CLI-Anything against each qualifying repo. ~20 min per CLI.
  4. Test — CLI-Anything auto-tests. If tests fail, skip and move on. Don’t waste time.
  5. Catalog — Add to CLI registry with: name, description, source repo, test count, capabilities, category.
  6. Sync — Push to all 3 farms automatically.

Math:

  • 500 CLIs/day x 365 = 182,500/year (Phase 1)
  • Ramp to 1,000/day by month 3 = 365,000/year
  • Ramp to 2,740/day by month 6 = 1M/year pace
  • Parallel agents: As farm capacity grows, run 2-3 harvester agents simultaneously

Scaling plan:

  • Month 1-2: 500/day (1 agent, learning, refining process)
  • Month 3-4: 1,000/day (2 agents, categories prioritized)
  • Month 5-6: 2,000/day (3 agents, automated discovery pipelines)
  • Month 7-12: 2,740+/day (full capacity, hit 1M by end of year)

Categories to prioritize (in order):

  1. Developer tools (IDEs, linters, formatters, build tools)
  2. Design/creative (Blender, Inkscape, GIMP, Figma plugins)
  3. Audio/video (Audacity, FFmpeg, OBS, HandBrake)
  4. Data/analytics (databases, visualization, ETL tools)
  5. Business tools (CRM, invoicing, scheduling, project management)
  6. DevOps/infrastructure (Docker, K8s, monitoring, CI/CD)
  7. Communication (email, chat, notification tools)
  8. Finance/payments (accounting, reporting, reconciliation)
  9. Education/learning (LMS, quiz, assessment tools)
  10. Industry-specific (construction, healthcare, legal, real estate)

Sales differentiator: When we sell Maat Agents to businesses, we say: “Our agents have access to 1 million tools. Your competitor’s agents have access to maybe 50. That’s why ours deliver in hours, not weeks.”

Where CLIs Live

  • Dedicated repo: imaginationeverywhere/cli-arsenal (or similar)
  • S3 mirror: s3://quiknation-cli-arsenal/ for fast farm access
  • Registry: JSON catalog file with search/filter capabilities
  • Agent discovery: check registry before building (avoid duplicates)
  • Version tracked: CLIs can be updated when source software updates
  • Daily count published to maat-agents — transparency on progress toward 1M

Graph View

Backlinks