I gave my AI access to my notes. It corrupted them.
I built Matryca Plumber so that never happens again.

Built on Andrej Karpathy's LLM-Wiki vision, Matryca Plumber is a local background assistant that keeps your Logseq notes organized, linked, and safe while you (or your AI) work — no cloud, no Logseq API, no silent overwrites. It has also CLI/MCP for your AI agents to directly read/write on your notes. It's a promising new architecture for agentic memory with graph architecture and markdown based.

Developed by Marco Porcellato · Matryca.ai — the product name is Matryca Plumber (not “Matryca” alone). See docs/BRANDING.md.

• Runs in the background on your laptop and reads your note files locally
• Summarizes pages, suggests tags, and fixes broken [[links]] while you sleep
• Lets Cursor / Claude connect via MCP to query and update notes safely
• Aborts AI edits if you changed the same page — your typing always wins
• Works offline with LM Studio or Ollama (no OpenAI/Anthropic subscription required)
• Ships a local dashboard (matryca plumber status) to start/stop and tune safety modes

Matryca Plumber targets Logseq OG (Markdown on disk); Obsidian comparisons refer to common community plugins, not a single product.

🤖 Are you an AI Agent (Cursor, Windsurf, Claude, Hermes)? To interact with the user's Logseq graph, do not parse Markdown manually. Execute uvx matryca-plumber --help directly in the terminal to read the graph via PyPI, or read llms.txt / .well-known/llms.txt for verified v1.9.14 CLI/MCP commands (LOGSEQ_GRAPH_PATH, stdio MCP, Hermes lazy handshake, AX lenient page resolution, Security & Sandbox hardening, journal-aware Phase 2 clustering, bootstrap_status, no --graph). Spec: docs/openspec/agent-onboarding.md · AX: docs/openspec/agent-ax-robustness.md · Security: docs/openspec/security-sandbox.md · Hermes: docs/integrations/hermes-agent.md · LLM OS: docs/openspec/llm-os-instructions.md.

Current: v1.9.14 — Contributor Readiness & Tech Debt Cleanup: OSS onboarding blueprints, journal-aware Phase 2 clustering, entity-consolidation token savings, and scoped tech-debt fixes (#62, #64) — 710+ passing tests. Upgrade with uvx matryca-plumber; full notes in CHANGELOG.md.

"Logseq is building the best local outliner database. But AI Agent memory is at the very bottom of their roadmap. Matryca Plumber gives you that future today, safely bridging your local agents to your Logseq graph without waiting years." - Marco Porcellato - Matryca.ai chief architect and co-founder

Under the hood it is a headless daemon + CLI that edits your Logseq .md files directly — see Architecture and What does it actually do? below.

Matryca Plumber edits your local .md files directly. While it features safe Optimistic Concurrency Control (OCC) to prevent data loss, we strongly recommend testing it on a clone of your graph first. This allows you to see the AI in action and explore its capabilities without affecting your primary notes.

How to safely clone your graph (crucial if you use Logseq Sync):

1. Make a copy of your entire Logseq graph folder on your computer (e.g., duplicate your MyGraph folder and rename it to MyGraph_Test).
2. Open Logseq, click on your graph name in the top left, and select Add new graph.
3. Choose the new MyGraph_Test directory.
4. If you use Logseq Sync: Do not enable Sync on this test graph. This ensures the AI's test edits remain strictly local and do not propagate to your other devices.
5. Alternatively, for a minimal test graph: a. Create an empty folder and add it as a new graph in Logseq. b. Close Logseq; copy only pages/, journals/, assets/ from production. c. Reopen and Re-index.
6. Point Matryca Plumber's .env configuration (LOGSEQ_GRAPH_PATH) to this test folder.

Once you are comfortable with how Matryca Plumber behaves and have tuned the safety tiers, you can point it to your main graph.

The fastest way to get started is using uv, the blazing-fast Python package manager.

Run the CLI directly without polluting your system. This only opens the Sovereign UI at http://127.0.0.1:8500 — it does not start graph maintenance until you complete the pre-flight checklist and click Start Engine (or run matryca plumber start separately).

uvx --from matryca-plumber matryca-plumber status

(matryca-plumber status is shorthand for matryca plumber status.)

Install the binary to use the matryca command anywhere:

uv tool install matryca-plumber

matryca plumber status
# same as: matryca-plumber status

The browser opens the Sovereign UI. A Pre-flight checklist modal appears when checks are not yet green (it auto-dismisses once they pass). Use Continue to dashboard or reopen Pre-flight from the header anytime, then click Start Engine to launch the maintenance daemon from the UI.

Optional — headless daemon before opening the UI:

matryca plumber start    # background worker only; no browser
matryca plumber status   # UI still shows pre-flight; engine may already show IDLE/RUNNING

matryca-plumber status is shorthand for matryca plumber status (same for start, stop, ui, audit, cluster).

The UI binds port 8500 in seconds (lazy graph index). Graph analytics cards may take longer on the first load while the in-memory AST warms up. The React bundle is served from frontend/dist/ (auto-built on first status/ui when frontend/node_modules/ exists).

Common mistake: matryca plumber start does not open the dashboard — run matryca plumber status in another terminal (or skip start and use Start Engine from the UI).

Install it as a LaunchAgent/systemd service so it wakes up with your OS:

matryca service install

Same headless contract as MCP tools; add --json for structured stdout:

export LOGSEQ_GRAPH_PATH=/path/to/your/graph

matryca --json read page "My Project"
matryca context load "My Project"
matryca context load "My Project|aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa"
matryca --json read subtree '{"page":"My Project","block_uuid":"…","heading":"Implementation"}'

Full spec: docs/openspec/agent-dx.md. Background link checks: docs/openspec/link-verification.md.

Unlike generic scripts, Matryca Plumber is a continuous background engine. When paired with a local LLM (Gemma 4-E4b Instruct via LM Studio or Ollama), it provides:

• Semantic Indexing: Automatically generates summaries, suggested tags, and cross-references for your pages.
• Dangling Link Healing: Finds broken [[WikiLinks]] and creates isolated seed pages for them.
• Entity Consolidation: Suggests alias:: properties for overlapping concepts.
• Auto-Split Dense Blocks: Extracts oversized subtrees into new pages to keep your graph fast and readable.
• Claude Desktop Integration (FastMCP): Seven MCP tools (five mega-tools + store_fact + ingest_document) query and mutate your graph headlessly. Set MATRYCA_MCP_ENABLED=true in .env only on machines where you trust the MCP host (stdio MCP is off by default; the host has full graph read/write with no separate authentication).
• Telos & Identity (in-graph persona): Optional pages/matryca___config.md or pages/matryca-config.md with - # Telos and - # AI Constraints headings — injected into daemon LLM prompts and MCP output; store_fact appends durable preferences under Constraints (docs/openspec/identity-config.md).
• Atomic document ingestion: ingest_document parses external Markdown via an OS temp file (never under pages/), stamps fresh id:: UUIDs, and appends to daily Ingest/YYYY-MM-DD or MATRYCA_INGEST_PAGE, with optional LOG / GLOSSARY ledgers (docs/openspec/ingest.md).
• Structural link verification (v1.9): Passive harvest of URLs and assets/ paths into .matryca_link_registry.json; async HTTP HEAD + filesystem checks; OCC-safe dead-link:: / missing-asset:: block properties (docs/openspec/link-verification.md).
• Agent-centric DX (v1.9): Global CLI --json, matryca context load, read subtree, and Journey Log — one cumulative - 🤖 Matryca Activity bullet in today's journal (daily totals, updated in place) (docs/openspec/agent-dx.md).
• Agent onboarding (v1.9.2+): Machine-readable llms.txt with PyPI/uvx execution contract and anti-patterns (docs/openspec/agent-onboarding.md).
• Security & Sandbox (v1.9.9): Sandbox-validated graph reads (read_graph_file_text), bounded JSON checkpoints, link-registry path validation, CI sandbox-read-check (docs/openspec/security-sandbox.md, SECURITY.md).
• LLM OS contract (v1.9.5): Tier-2 agents check bootstrap_status and [[Matryca Master Index]] before blind vault search; Soft Gate offers Local Daemon / Blind Search / Cloud Indexing (docs/openspec/llm-os-instructions.md, SYSTEM_PROMPT.md § "LLM OS").

Matryca Plumber is a three-surface runtime — background daemon, Sovereign UI, and optional CLI/MCP — that converges on one headless mutation plane. Every surface reads and writes the same Logseq OG vault on disk (LOGSEQ_GRAPH_PATH): no Logseq HTTP API, no cloud database, no split-brain datastore. Humans co-edit the same .md trees while the daemon runs Phase 1 catalog harvest and Phase 2 cognitive lint against a local LLM (LM Studio or Ollama).

flowchart TB
  subgraph operators [Operators and agents]
    Human[Human · Logseq optional]
    UI[Sovereign UI :8500]
    Daemon[Maintenance daemon]
    CLI[matryca CLI --json]
    MCP[FastMCP stdio optional]
  end

  subgraph plane [Shared headless mutation plane]
    GD[graph_dispatch]
    Lock[OCC + page_rmw_lock]
    Parser[logseq-matryca-parser]
    GD --> Lock
    GD --> Parser
  end

  subgraph local [Local inference — 100% offline]
    LLM[LM Studio / Ollama]
  end

  subgraph vault [LOGSEQ_GRAPH_PATH — single source of truth]
    Pages[pages/ · journals/ · templates/]
    Meta[.matryca_* cache & ledgers\nmatryca-l1/ session rules]
  end

  Human <-->|co-edit Markdown| Pages
  UI -->|start · stop · config · telemetry| Daemon
  Daemon -->|Phase 1 harvest · Phase 2 lint| GD
  Daemon <-->|structured JSON| LLM
  CLI --> GD
  MCP --> GD
  Lock -->|atomic UTF-8 writes| Pages
  Meta -.-> GD

Deep dive: docs/ARCHITECTURE.md (Phase 1→2 lifecycle, Trust & Safety tiers, LLM OS contract).

Matryca Plumber is 100% headless, but it ships with a Sovereign UI Cockpit (matryca plumber status or uvx … matryca-plumber status). It's a local React dashboard running on http://127.0.0.1:8500 that provides:

• Pre-flight checklist (modal when checks fail; auto-dismiss when green): operator guidance plus automated readiness checks — only fail blocks Start Engine; warn is advisory (e.g. degraded file locks, model not listed).
• Live Graph Telemetry (v1.9.3): Progress bar, Phase 1/2 pills, and token counters refresh about every 5 seconds while the engine runs — including during long LLM turns (docs/openspec/live-telemetry-ui.md).
• Dynamic Impact: Mathematically separates Organic Human Mind (your notes) from Agent Cognition (AI enhancements).
• Zero-Trust Security: Every REST call requires a Bearer token (X-Matryca-Token). Set MATRYCA_UI_TOKEN on shared hosts; new installs from .env.example template MATRYCA_UI_REQUIRE_EXPLICIT_TOKEN=true. Session bootstrap is loopback-only; split rate limits for authenticated vs anonymous API traffic.
• Trust & Safety Drawer: Visually toggle what the AI is allowed to edit (Safe Mode, Augmented Mode, Surgeon Mode).

See SECURITY.md for the full operator hardening matrix (MATRYCA_MCP_ENABLED, graph path allowlist, bounded JSON, shared LLM SSRF policy, log redaction).

Matryca Plumber provisions missing runtime files automatically where possible (repo .env from .env.example, matryca-l1/, cache dirs, matryca-wiki.yml). The modal still walks you through setup so nothing surprises you on first run. It is developed by Marco Porcellato at Matryca.ai — the same attribution shown in the Sovereign UI pre-flight wizard.

Operator steps (same text as the UI wizard):

1. Control room connection — If you can read this dashboard, matryca plumber status (or ui) has started the local API on port 8500. That command does not start the maintenance daemon by itself. Keep the window open while the engine runs.

2. Logseq graph (test vault first) — Point LOGSEQ_GRAPH_PATH at the root of a Logseq OG vault (the folder that contains pages/). Use a clone for your first run; do not enable Logseq Sync on test graphs. In the UI: Settings (gear) → Logseq Graph Path → absolute path → Save.

3. Local LLM — Start an OpenAI-compatible server (LM Studio, Ollama, etc.). In Settings set the base URL (e.g. http://localhost:1234/v1) and the exact model id, then Refresh models to confirm discovery.

   Matryca Plumber (by Marco Porcellato · Matryca.ai) is built for offline, CPU-only use on a typical 16 GB RAM machine — no cloud subscription or discrete GPU required. The recommended and tested model is Gemma 4-E4b Instruct — set the exact id gemma-4-e4b-it in Settings, then Refresh models. For CPU inference, prefer GGUF weights at Q4_K_M or Q5_K_M. We are actively testing additional open models to improve CPU-only, 16 GB setups; Gemma 4-E4b Instruct is our current default. Avoid large MoE models (e.g. Llama 4 Scout): full weights still require 60GB+ RAM.

4. First-run expectations — Phase 1 catalogs the entire graph (can take a long time on large vaults; v1.8 yields to the OS periodically during harvest; the UI shows catalog pills about every 5 pages). Phase 2 processes roughly one LLM-heavy page per poll interval by default, with live token totals and progress in the cockpit. After Phase 1, the daemon releases heavy in-memory indexes to keep RAM stable for long runs. If you started the daemon from a terminal before opening the UI, the dashboard still polls state when it sees a live daemon_pid — click Start Engine for the full live console (logs + graph analytics).

Live checks (re-run anytime with Re-run checks):

Start Engine stays disabled while any live check is fail (yellow warn rows are OK). If you started the daemon earlier with matryca plumber start, the UI may already show IDLE or RUNNING — use Pre-flight in the header to review settings. On large vaults, settings save and Start Engine return in seconds (lazy bootstrap; no full AST load in the UI process).

• 🤖 100% Local-First & Headless: No Logseq HTTP API required. It edits the .md files directly using atomic file I/O.
• 📐 Exact Logseq AST Compliance: True line-0 page frontmatter, block properties at +2 indent, and exact namespace encoding. Other tools break your graph; Matryca Plumber keeps it pristine.
• 🔐 Optimistic Concurrency Control: It snapshots st_mtime before inference and acquires the page lock only for the write. If you edited in Logseq while the model was thinking, the commit aborts. No silent data loss — and Logseq can still save during long local runs.
• 🪟 Windows, macOS & Linux Support: Runs safely in the background everywhere using a robust cross-platform lock (.matryca_plumber_daemon.lock).
• ⚡ Context Acceleration Shield: Shrinks megabyte-class pages to Phase 1 summaries or semantic skeletons before they reach the local LLM — essential on CPU-only hardware.
• 🛡️ TRIZ-governed LLM resilience: Caps completion tokens, first-delimiter balanced JSON extraction (objects and arrays), string-aware trailing trim, prose sanitization on compression/history paths, stateless ontology reports, and an 8k block-catalog cap on semantic index prompts — see docs/resilience-llm-json-triz.md.
• 🖥️ Edge computing profile (v1.8): KV-cache-aligned prompts (PagePromptSession), bounded RAM (BM25 postings-lite, semantic cache LRU, post-bootstrap teardown), and cooperative bootstrap I/O — tuned for 16 GB laptops and vaults up to ~10,000 pages. See docs/v1.8-OPTIMIZATION-PLAN.md.
• 🔗 Structural hygiene (v1.9): Background link rot and missing-asset detection without LLM tokens; a single Journey Log bullet in today's journal with cumulative indexing, link-check, and flag totals (no per-cycle journal spam).
• 🤖 Agent-native CLI (v1.9): matryca --json … for machine-readable stdout; matryca context load and read subtree to shrink context windows.
• 📋 Agent onboarding (v1.9.2): llms.txt — canonical uvx commands for Cursor, Claude Code, and other hosts; no git clone required.
• 🧠 LLM OS (v1.9.5): Two-tier agent discipline, bootstrap_status semaphore, Master Index Soft Gate — docs/openspec/llm-os-instructions.md.
• ⚡ Live telemetry (v1.9.3): 5s Sovereign UI updates, thread-safe daemon heartbeat, API token overlay from ops log — docs/openspec/live-telemetry-ui.md.
• 🛡️ Enterprise Resilience (v1.9.13): Vault Sandbox traversal blocked; TOCTOU-safe bounded JSON; namespace-aware semantic cache; exact embedding dedup; subtree heading fences; string-aware LLM JSON recovery; self-healing vector store and daemon state; truthful plumber stop exit codes — CHANGELOG.md · docs/resilience-llm-json-triz.md.
• 🧹 Contributor readiness (v1.9.14): good_first_issues_blueprints.md for external contributors; journal pages isolated in Phase 2 clustering; entity-consolidation skips journal/date wikilink pairs; shared NoRedirect HTTP helper and alias-registry compatibility helpers — CHANGELOG.md · ROADMAP.md.
• 🖥️ Sovereign UI reliability (v1.9.11): Settings save and Start Engine stay responsive on large vaults (lazy bootstrap); pre-flight warn is advisory — docs/openspec/runtime-bootstrap.md.

You are in control. Nothing mutates your prose unless you explicitly enable it in the UI.

Copy .env.example to .env. The only required variable is your graph path:

LOGSEQ_GRAPH_PATH=/absolute/path/to/your/Logseq/graph
MATRYCA_LM_BASE_URL=http://localhost:1234/v1   # LM Studio or Ollama endpoint
MATRYCA_LM_MODEL=gemma-4-e4b-it                # Gemma 4-E4b Instruct — tested default

# Optional: Claude Desktop / Cursor MCP (off by default)
MATRYCA_MCP_ENABLED=true

On first start (daemon, CLI, MCP, or UI), Matryca Plumber automatically creates anything missing for a healthy runtime:

• logs/ (or paths from MATRYCA_PLUMBER_LOG_PATH / MATRYCA_LOGURU_LOG_PATH)
• <parent-of-your-vault>/matryca-l1/ — session rules beside the vault (not inside pages/); optional override via MATRYCA_L1_PATH
• <vault>/.matryca_semantic_cache/, templates/, and matryca-wiki.yml (from matryca-wiki.example.yml when absent)
• In-memory graph index at startup (AST cache); identity loaded when a Telos/Constraints config page exists

The identity config page is not auto-created (use Logseq or store_fact). Ingest / LOG / GLOSSARY pages are created on first ingest_document call. Optional MATRYCA_INGEST_PAGE pins a fixed inbox (e.g. AI_Inbox). See docs/openspec/identity-config.md and docs/openspec/ingest.md.

See docs/openspec/runtime-bootstrap.md for rationale (L1 vs L2, idempotency, and what is intentionally not auto-created).

(See docs/ARCHITECTURE.md for advanced thermal pacing, context compression, and deep linter settings. Copy the full template from .env.example — it documents UI auth, rate limits, graph allowlists, and log redaction.)

Copy the v1.8 Edge computing & performance block from .env.example. Highlights:

Install CPU affinity support: uv sync --extra edge or pip install matryca-plumber[edge] (psutil).

Deep dive: docs/v1.8-OPTIMIZATION-PLAN.md · docs/v1.8-SOFTWARE-EDGE-PLAN.md · docs/openspec/llm-performance.md

Load testing: uv run python scripts/gen_synthetic_graph.py /path/to/graph --count 1000 · Slow CI: make perf

Want to contribute or run from source?

git clone [https://github.com/MarcoPorcellato/matryca-plumber.git](https://github.com/MarcoPorcellato/matryca-plumber.git)
cd matryca-plumber
make install

# Build the React frontend
cd frontend && npm install && npm run build && cd ..

# Run tests (710+ passing, Mypy strict)
make check

# Optional: slow memory / harvest soak tests
make perf

Apache-2.0 — see LICENSE.