Morphis

An open-source agentic OS. Voice-first. JARVIS-like. Forkable. Yours.

Morphis is an agent kernel — a system that sits above your existing OS (macOS, Windows, Linux, Android) and turns your computer into a voice-driven AI assistant that can install software, build apps, manage your digital life, and remember everything. Like JARVIS from Iron Man, but real, open source, and customizable.

The four-line pitch:

• Voice-first. Talk to your computer. Get answers, actions, and built software back. Sub-300 ms acknowledgement.
• Cloud + Local. Uses Claude / GPT when online. Falls back to local Qwen / Gemma when offline. Auto-routes.
• Owns its memory. Every conversation is a markdown file in an Obsidian-compatible vault you own. It evolves nightly like sleeping human memory.
• Forkable like Linux. Apache 2.0. Stewarded by The Nerfed Lab. Make your own Morphis variant in a YAML recipe.

---

Status — V0.1.0 alpha

This is the alpha cut. Some things work end-to-end today; others are deliberately stubbed so the architecture stays honest as the rest lands. The full V0→V3 timeline is in DEVELOPMENT_PLAN.md.

Works today:

• Voice loop: wake word → STT → LLM (Claude cloud + Ollama local fallback) → episodic memory write → TTS. Try morphis voice loop.
• Living Vault: markdown notes under ~/.morphis/vault/, hybrid dense+recency retrieval, cold-start re-index.
• Sandbox tier 1 (Wasmtime / WebAssembly) — runs code-as-actions today.
• Permission Broker with JSONL audit log — every shell:exec goes through one HITL gate.
• Skill loader (morphis skills install --local <path> / morphis skills list). Three reference skills in examples/skills/.
• morphis doctor — full subsystem health check.
• morphis log — query the audit log with --tail / --grep / --since.

Surface ships, backend lands later:

• Sandbox tier 2 (Firecracker microVM). Skills declaring sandbox: firecracker are refused loudly in V0.1.0 — never silently downgraded. Backends land in V0.5 (ADR 0007).
• morphis-shell (Tauri desktop window). Today the CLI is the only entry point.
• L0–L5 Permission Broker tier ladder. V0.1.0 has one gate; the full ladder lands in V1.
• LanceDB vector index. V0.1.0 ships an in-memory linear-scan cosine impl; the trait surface is the contract (ADR 0005).
• bge-micro-v2 embeddings via fastembed. V0.1.0 uses a deterministic hash embedder.

Not in V0.1.0:

• App marketplace (V1)
• Linux distro variant (V2)
• Android launcher / iOS companion (V3)

Phase	Timeline	What ships
V0 (current)	0–6 months	Voice loop + LLM router + Living Vault + Wasmtime sandbox + Permission Broker V0 + skill loader. CLI-only.
V0.5 — Hardening	months 2–4	Firecracker backends, LanceDB, bge-micro-v2, L0–L5 broker, sub-agents.
V1 — Skills marketplace	months 4–9	SKILL.md adoption + MCP-native + LiteLLM + Sigstore registry + multi-channel inbox.
V2 — Morphis OS	months 9–15	BlueBuild / Fedora Atomic distro variant.
V3 — Mobile	months 15–24	Android launcher + iOS companion + libp2p mesh.

---

What Morphis can do (today)

These are the V0.1.0 demos. Each one actually runs against the code in this repo.

morphis voice loop — wake on "Morphis", STT your turn, route to Claude or local Ollama, write the exchange to ~/.morphis/vault/episodic/, speak the reply.

morphis skills install --local examples/skills/launch-app — installs a skill that opens a host application by name. Permission Broker gates the shell:exec; the decision is logged to ~/.morphis/logs/approvals.jsonl.

morphis doctor — verifies every subsystem reports ✓: platform, vault, sandbox, LLM connectivity, skill store.

morphis log --since 1h --grep approval — query the audit log of every permission decision Morphis has made.

For the full vision (build-a-website via Claude Code sub-agent, multi-channel inbox, etc.), see docs/users/what-morphis-can-do.md.

---

Install (V0.1.0 — from source)

Pre-built installers (brew install, winget install, flatpak install) land in V0.5 alongside Sigstore-signed release artifacts. For V0.1.0, build from source:

# Requirements:
#   - Rust 1.78+ (rustup is fine)
#   - Python 3.12+ (voice loop + Claude provider use it as a shim)
#   - cmake + pkg-config (for cpal / Rust audio crate)
#   - On Linux: libasound2-dev

git clone https://github.com/Nerfed-Lab/morphis
cd morphis
cargo build --release --bin morphis

./target/release/morphis doctor          # health check
./target/release/morphis skills list     # (empty by default)
./target/release/morphis skills install --local examples/skills/launch-app

The voice loop additionally needs:

# Python venv for the voice helpers (STT, TTS, wake word)
python3 -m venv morphis-voice/python/.venv
source morphis-voice/python/.venv/bin/activate
pip install -r morphis-voice/python/requirements.txt

# Optional: cloud LLM. Skip for local-only.
export ANTHROPIC_API_KEY=sk-ant-...

# Optional: local LLM via Ollama
ollama serve &
ollama pull qwen3.5:9b

./target/release/morphis voice loop

Full instructions, troubleshooting, and hardware notes: docs/users/install.md.

---

Pick your entry point

Morphis serves four very different audiences. The docs branch by what you want to do:

You are…	Start here
A user — want to install Morphis and talk to it	docs/users/install.md
A skill author — want to write a skill in 10 minutes	docs/builders/quickstart-skill.md
A developer — want to contribute to Morphis core	docs/developers/architecture.md
A researcher — want to cite or build on Morphis	docs/researchers/theoretical-foundations.md
A fork maker — want to ship your own Morphis variant	docs/forks/creating-a-variant.md
Browsing — want the big picture	docs/start-here.md

For the authoritative architecture, threat model, and research grounding: MORPHIS_RESEARCH.md. For the executional plan: DEVELOPMENT_PLAN.md. For the architectural decisions: docs/decisions/.

---

What makes Morphis different

Each row is a load-bearing design choice — read MORPHIS_RESEARCH.md for the rationale.

Feature	Morphis	OpenClaw	Claude Code / Cursor
Voice-first JARVIS personality	✅ default	bolted on	❌
Sandbox by default (Wasmtime + Firecracker microVM)	✅ from line one (ADR 0006)	retrofitting Q1 2027	❌
Permissioned self-modification (L0–L5)	✅ unique	❌	❌
Per-variant security postures	✅ Reference/Hacker/Caregiver/Sterile/Yolo	❌	❌
Linux distro variant (Morphis OS)	✅ V2 via BlueBuild	third-party experiment	❌
Android launcher (full shell replacement)	✅ V3	companion app only	❌
"Build me X" via Claude Code in microVM	✅ V0.5 anchor demo	not first-class	❌
Self-evolving Obsidian-vault memory ("Living Vault")	✅ unique — cd into your memory (ADR 0005)	opaque vector DB	hidden context window
Code-as-actions (smolagents pattern)	✅ V0 default (ADR 0008)	JSON tool-calls	JSON tool-calls
Apache 2.0 + The Nerfed Lab stewardship	✅ from day 1	MIT + independent foundation	proprietary
20+ chat channels	✅ V1 via Matterbridge sidecar	✅ native	❌
MCP-native	✅ V1	✅	✅
140+ LLM providers	✅ V1 via LiteLLM	✅ via gateway	partial
SKILL.md compatibility (OpenClaw / Claude / Cursor / Codex)	✅ V1	✅	✅

---

Contributing

Morphis is built in the open. Every kind of contribution welcome:

• Bug reports → .github/ISSUE_TEMPLATE/bug_report.md
• Feature requests → .github/ISSUE_TEMPLATE/feature_request.md
• Skill proposals → .github/ISSUE_TEMPLATE/skill_proposal.md
• Security issues → see SECURITY.md — disclose privately first.
• Code contributions → start with CONTRIBUTING.md, then docs/developers/contributing-code.md.
• Doc improvements → PRs welcome; small fixes don't need an issue first.

Governance, RFC process, and TSC structure: GOVERNANCE.md.

---

License

Apache 2.0. Patent grant included. Trademark "Morphis" held by The Nerfed Lab (in formation). See docs/governance/trademark.md for trademark policy.

---

Acknowledgements

Morphis stands on the shoulders of: AIOS (Rutgers, COLM 2025) for the agent-kernel framing; Anthropic Claude Auto Dream and A-MEM (NeurIPS 2025) for memory consolidation; OpenClaw for proving the agent-OS-as-substrate pattern; Universal Blue / BlueBuild for forkable-Linux-by-recipe; Obsidian and Smart Connections for the vault substrate; smolagents and Open Interpreter for the code-as-actions pattern; Wasmtime + Firecracker for the sandbox boundary; LiteLLM, Matterbridge, and mautrix for absorbing huge integration moats so we don't have to.