Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        llamaclaw ecosystem                        │
│                                                                   │
│  ┌───────────────┐   ┌───────────────┐   ┌─────────────────────┐ │
│  │   end user    │──▶│  terminal     │──▶│  esml Python + R    │ │
│  │  (researcher) │   │   (tide/Go or │   │  (stats + causal)   │ │
│  └───────────────┘   │   kronos/Rust)│   │                     │ │
│                      └───────────────┘   │  ┌───────────────┐  │ │
│                              │           │  │ turboquant    │  │ │
│                              ▼           │  │ (KV compress) │  │ │
│                      ┌───────────────┐   │  └───────────────┘  │ │
│                      │    perseus    │   │  ┌───────────────┐  │ │
│                      │  (LLM relay)  │◀──┤  │   autoresearch│  │ │
│                      └───────────────┘   │  │  (pretraining)│  │ │
│                         │     ▲          │  └───────────────┘  │ │
│                         ▼     │          └─────────────────────┘ │
│                      ┌───────────────┐                           │
│                      │    Ollama     │◀──────────────┐           │
│                      │ (local weights│               │           │
│                      └───────────────┘   ┌───────────────────┐   │
│                                          │ ollama.com/       │   │
│  ┌────────────────────────────────────┐  │ llamaclaw         │   │
│  │        Raspberry Pi 5 zeus.local   │  │ (model registry)  │   │
│  │                                    │  └───────────────────┘   │
│  │  luci (ZeroClaw agent) ─────┐      │                          │
│  │  perseus-relay service      ├─▶ systemd (deploy/)             │
│  │  ollama daemon              │      │                          │
│  │  kronos terminal            │      │                          │
│  └────────────────────────────────────┘                          │
└──────────────────────────────────────────────────────────────────┘

Dependency direction

Arrows point downward — higher-level repos depend on lower-level ones.

installer  (Homebrew + curl|sh)
    │
    ▼
 esml  ◀── turboquant ── autoresearch
    │         ▲
    ▼         │
 perseus ── Ollama
    │
    ▼
 tide / kronos  (UI layer)

  • esml depends on: numpy, scipy, statsmodels, pandas, and (as VCS deps) turboquant + perseus.

  • turboquant depends on: numpy, scipy only. Standalone compression.

  • perseus depends on: httpx only. Pure LLM relay; auto-detects Ollama / FreeAPI / Gemini / OpenAI at runtime.

  • autoresearch depends on: PyTorch + MPS (main) or CUDA (upstream) + turboquant for PTQ.

  • tide is a Go binary that shells out to esml CLI + Ollama HTTP.

  • kronos is a Rust ARM binary with no external runtime deps.

  • luci is shell orchestration; queries perseus-relay on the Pi.

  • deploy is systemd/launchd unit files only.

  • yoda is pure Markdown (Claude’s persona + journal).

  • installer is the user-facing install entry point; doesn’t depend on any other llamaclaw repo — it just knows how to install them.

  • llamaclaw (umbrella) pulls all of the above as git submodules.

Hardware matrix

M-series Mac

Raspberry Pi 5

Linux x86_64

esml (Python + R)

✓

✓

✓

turboquant (C bridge)

✓

✓

✓

perseus

✓

✓

✓

autoresearch

✓ (MPS branch)

✗

✓ (upstream branch)

tide

✓

✓ (Go 1.26+)

✓

kronos

✓

✓ (native target)

✓

luci

✗ (needs Pi)

✓

✗

Ollama

✓

✓

✓

Where the data lives

Location

Size

41 built-in public-health datasets

esml/py-package/esml/data/esml_datasets.db (Git LFS)

~500 MB

fn/ registry (8,872 functions)

esml/py-package/esml/fn/

~80 MB

LLM model weights (Perseus)

Ollama on zeus.local + ollama.com/llamaclaw

8–10 GB per model

User analysis outputs

$ESML_OUTPUT_DIR (default: data/manifest/outputs/)

varies

Cross-repo CI patterns

  • Per-repo CI: tests, lint, build — on every push to dev.

  • Secrets scan: org-level workflow runs gitleaks on every push + weekly.

  • Dependabot: org-level config — weekly updates for github-actions/pip/gomod/cargo, monthly for Docker.

  • Docker builds: esml, tide, kronos publish multi-arch images to ghcr.io/llamaclaw/<repo> on push to main (and on semver tags).

  • Release coordination: llamaclaw/scripts/release.sh opens synchronized dev → main PRs across sub-repos and tags them.