A personal AI orchestration backend for Apple Silicon. Private, local, modular — Claude-equivalent capabilities for scientific and software work, fully under your control. Named after Linus Pauling and Linus Torvalds.

Status: v0.5.0 — the local orchestration substrate, tagged 2026-06-02 (first revealed at the Agora hackathon, 2026-05-25). Debuting alongside KnowledgeBase (the data substrate) and Archimedes (q-fin strategy engine — Linus specialized to one domain, made externally verifiable, and shipped on-chain).

Logo placeholder: a carbon atom — sp³ orbitals, the substrate of biochemistry. Asset to land before reveal.

OpenAI- and Anthropic-compatible HTTP backend that runs on a MacBook, exposes domain-specific tools backed by a personal knowledge base, coordinates local models in parallel, enforces sandbox policy regardless of front-end, and is designed to evolve toward fine-tuned models trained on its owner's data. Not a hosted-Claude replacement — Linus is the Worker that executes under hosted Claude + you (the Maestro), and serves you directly when private or offline operation is needed. The orchestration layer is what accrues value across every harness, every paper read, every synthesis written.

git clone --recurse-submodules git@github.com:dbrowneup/Linus.git && cd Linus
conda env create -f environment.yml && conda activate linus && pip install -e .
brew install ollama && brew services start ollama && ollama pull qwen3:8b
linus-serve                                     # backend on http://localhost:8000
streamlit run src/linus/app/main.py             # optional KB-consuming UI

Designed for Apple Silicon (M1/M2/M3 with Metal acceleration). Reference machine: 32 GB M1 Max. Full env notes, Ollama performance env vars, and known library quirks live in CLAUDE.md.

The local orchestration substrate: OpenAI + Anthropic chat-completions endpoints (streaming + tool calls, multi-iteration tool loop) · server-side session persistence (SQLite WAL) · fastmcp-based MCP tool registry (DEC-0045) · KnowledgeBase read-only adapter feeding Streamlit pages (paper graph, cluster explorer, KG render, search) · a grounding/rigor gate (DEC-0059), loud /healthz degradation (DEC-0060), and per-tool network policy (DEC-0061) · five-layer memory pillar (DEC-0028; intra-step latent → within-session scratchpad → cross-session episodic → investigation memory → KnowledgeBase semantic) · path-validating sandbox enforcing the SAFETY.md Tier 0/1 contract · hermetic test suite, 704 tests in ~6s. paper-qa ships as an optional "few named papers" deep-dive tool — the corpus-scale, citation-grounded RAG marquee is the v0.6.0 deliverable, a Linus-native full-text-chunk substrate (see the evaluation). Component-by-component breakdown in ARCHITECTURE.md; per-phase status in ROADMAP.md.

Linus/
├── src/linus/              # The product (FastAPI backend, KB integration, memory, sandbox, tools, Streamlit UI, tests)
├── modules/KnowledgeBase/  # Paper-analysis pipeline (tracked submodule)
├── repos/                  # Reference clones (gitignored, read-only study material)
├── docs/                   # ADRs, specs, syntheses, paper-notes, repo-notes, session summaries
├── benchmarks/             # Dan-tasks v0 + dated result JSON
├── experiments/            # Throwaway scripts, ablations
└── environment.yml · pyproject.toml

MIT, per LICENSE. The bundled modules/KnowledgeBase/ submodule is itself MIT but pulls in PyMuPDF (AGPL-3.0-or-later); if you redistribute Linus with KB, AGPL governs the combined distribution. Linus core alone is MIT-only. See KB's LICENSE for the AGPL-inheritance details.

Named after Linus Pauling (1901–1994; scientist, humanitarian, two-time unshared Nobel laureate) and Linus Torvalds (1969–; engineer, open-source advocate; creator of Linux and Git) — both Oregonians. Symbol: a carbon atom (sp³ orbitals, the foundation of biochemistry; a nice hook for what happens when carbon-based life meets silicon-based intelligence). Extended rationale in VISION.md. Per-repo credit for everything in repos/ lives in docs/repo-notes/. Thanks to Anthropic for Claude, Apple's MLX team for making Apple Silicon a real ML platform, and the authors of every paper and project that informs this work.