neurokeeper

One portable core, every runtime.
Deterministic hygiene engines for your notes vault and agent memory, run as a Claude Code plugin, a standalone CLI, or a CI gate. One codebase, no fork.

Install · How it works · Engines · CI gate · Docs site · Contributing

---

A portable-core harness for agentic coding/knowledge work. The load-bearing logic lives in harness-agnostic engines (deterministic scripts) + prompt templates + versioned contracts; a thin adapter binds them to a specific runtime. The Claude Code plugin is one adapter: the same core works from an MCP server, a plain CLI, or any other LLM harness.

Status: 0.3.2 (alpha). Licensed MIT (see LICENSE).

Why

"Skills" (and plugins, hooks) are runtime-specific and not portable. Putting the real logic in **engines

• prompts** makes a capability reusable across runtimes, vendor-neutral, and safe to open-source - and it lets you point the same engine at a self-hosted open model to save tokens. See docs/pattern-portable-core-and-adapters.md (the keystone).

The shape of a capability

portable core (this repo)                      adapters (bind the core to a runtime)
  scripts/<engine>.py   deterministic, --json    .claude-plugin + skills/  -> Claude Code
  prompts/<engine>.md   harness-neutral judgment  (future) mcp/            -> MCP server
  (contract = the engine's --json/CLI)            scripts run directly     -> plain CLI / CI / any LLM

Each engine carries a metadata header (@capability/@compute/@effect/@portability/...); scripts/registry-generate.py scans those and emits a registry (anti-rot, never hand-maintained).

Layout

• scripts/ - portable engines (deterministic). _vault_guard.py is a shared safety preflight.
• prompts/ - harness-neutral judgment templates.
• skills/ - Claude Code adapter skills (thin; defer logic to engine+prompt).
• hooks/hooks.json - Claude Code hooks (e.g. SessionStart memory-health).
• .claude-plugin/plugin.json - the Claude Code plugin manifest.
• config.example/ - example consumer config (L2). This repo ships schemas + examples only - never real config or content. Consumers supply their own private config + content.
• docs/ - the pattern doc + decisions.
• tools.lock - runtime tool deps (classified headless / gui / manual).

Install

As a Claude Code plugin (engines + skills + hooks; bin/neurokeeper on PATH):

/plugin marketplace add https://github.com/Wombat164/neurokeeper
/plugin install neurokeeper

As a standalone CLI (the engines, runnable anywhere / in CI - no Claude Code needed):

pipx install neurokeeper                                    # from PyPI
uvx neurokeeper --list                                      # zero-install one-shot (uv)
pipx install git+https://github.com/Wombat164/neurokeeper   # or track main from GitHub
pip install -e ".[dev]"                                     # or from a checkout (editable, with test/lint deps)

Then run any engine via the dispatcher:

neurokeeper <engine> [args]     # e.g.  neurokeeper taxonomy-inventory --json
neurokeeper --list              # list engines

Either way, set CLAUDE_MEMORY_DIR (and optional VAULT_*) per config.example/. The same engines back all three faces (plugin / CLI / direct python scripts/<engine>.py) - one codebase, no fork.

First capability: memory-audit

A file-based-memory health check + consolidation: deterministic engine (scripts/memory-consolidate.py - 5-metric score, orphans, broken-links, importance curve, --check/--terse/--json) + judgment prompt (prompts/memory-audit.md) + the memory-audit skill (the Claude Code adapter). The reference shape every other capability copies.

Vault taxonomy engine suite (L1a)

Deterministic, link-aware, Obsidian-guarded engines for Obsidian-vault KOS hygiene (proven on a ~1,400-note vault). All read VAULT_ROOT + config from the environment (see config.example/); dry-run/report by default, mutating only with --apply behind the Obsidian-running guard; git is the audit.

• vault-taxonomy-inventory.py - naming/tags/frontmatter inventory (read-only)
• vault-ref-audit.py - reference-integrity audit: broken links / orphans / dead-ends / broken .canvas+.base refs / orphan media / name-stem collisions (read-only; --check/--strict)
• vault-doctor.py - aggregate health: runs every applicable engine, one report + an honest tri-state roll-up (ok/fail/skipped); --check/--strict (see ADR-0002)
• vault-frontmatter-lint.py - validate notes against the schema (off-vocab / missing-axes / unknown-fields); --check
• vault-frontmatter-fix.py - apply the schema reconciliation maps (status/maturity/horizon; surgical line-edits)
• vault-set-note-type.py - additive note_type from a folder-derive map
• vault-tag-reconcile.py - detect morphological tag-merge groups (apply via Tag Wrangler)
• vault-name-reconcile.py - de-dash + kebab filenames, link-aware rename
• _vault_lib.py - shared core (walk / frontmatter / folder-suffix / kebab); _vault_guard.py - Obsidian-running guard

Schema-as-code lives in your config; config.example/frontmatter-schema.example.yaml is a worked example.

Gate a vault in CI (pre-commit + GitHub Action)

The same engines run as a commit/CI gate. neurokeeper ships a .pre-commit-hooks.yaml and a composite action.yml; both run the vault-graph-aware checks it uniquely owns and compose with the ecosystem (markdownlint / lychee / check-jsonschema) for the commoditized ones.

# .pre-commit-config.yaml (in your vault repo)
repos:
  - repo: https://github.com/Wombat164/neurokeeper
    rev: v0.3.2
    hooks: [{ id: neurokeeper-doctor }]

# a GitHub workflow step
- uses: Wombat164/neurokeeper@v0.3.2
  with: { vault-path: ".", engine: "doctor", strict: "false" }

Exit semantics follow ADR-0002 (fails on broken .canvas/.base refs or engine errors; advisory + skipped engines never fail). Full guide + the example vault: docs/ci-adapters.md / examples/vault/.

Fresh-machine reinstall (resilience) - Linux + macOS + Windows, co-equal

bootstrap/ rebuilds the whole multi-env setup on a clean Linux, macOS, or Windows box (the engines are cross-platform Python; only the bootstrap kit is OS-specific):

• tools-manifest.md - toolchain catalog with per-OS install source (Linux apt/dnf/brew + Windows winget) + class
• bootstrap.sh (Linux/macOS) / bootstrap.ps1 (Windows) - idempotent installers: headless toolchain + Claude Code + clone repos from your repos.txt
• RUNBOOK.md - end-to-end procedure (both OSes) incl. auth, memory restore, Obsidian + plugins, plugin install + a from-zero VM/container test
• repos.example.txt - template for your private repo list (repos.txt is gitignored)

Quickstart - Linux: bash bootstrap/bootstrap.sh bootstrap/repos.txt ~/Projects / Windows: pwsh -File bootstrap/bootstrap.ps1 -Root ~/Projects - then follow RUNBOOK.md.

Two-lane model handoff (optional, token-saving)

Keep hard agentic work on your normal Claude lane; route mechanical, high-volume work (commit messages, summaries, extraction, classification, formatting) to a self-hosted open model via the claude-cheap wrapper, so cheap work costs ~nothing. It points ANTHROPIC_BASE_URL at your /v1/messages-compatible endpoint for that invocation only - your default claude is untouched.

claude-cheap -p "write a conventional-commit message for the staged diff"

Config via config.example/cheap-lane.env.example. The design, the billing trap to avoid (a credentialed gateway flips you off your subscription onto per-token), the task-class whitelist, the serving recipes (vLLM-native / LiteLLM / claude-code-router), the data-egress/sovereignty warning, and a measure-before-you-trust plan are in docs/two-lane-model-handoff.md.

Related projects

neurokeeper builds on a healthy ecosystem; if it doesn't fit, one of these might:

• claude-mnemo - Claude Code plugin: Obsidian-vault memory + a lean MEMORY.md index + a /health audit (the closest neighbour on the memory side).
• claude-memory-health / claude-memory-compiler - markdown-memory audit + consolidation.
• obsidian-mcp-server - MCP tools for Obsidian (tag / frontmatter primitives).
• Native Obsidian plugins - Tag Wrangler, Linter, Smart Rename - the GUI "apply" layer this project deliberately defers to (the engines detect + propose; these apply).

What's different here: the deterministic vault-hygiene engine suite (tag / frontmatter / taxonomy / link-aware rename) + memory-audit behind a portable core - one codebase running as a Claude Code plugin and a standalone CLI/CI tool - rather than any single one of the above. A source-verified survey of the field (and exactly where neurokeeper is differentiated) is in docs/competitive-landscape.md.

Where this is heading (repository-agnostic backends, editor-state preflight, Bases/dashboard output adapters, and what is deliberately NOT planned) lives in docs/roadmap.md; the backend-seam contract it builds on is docs/adr-0001-backend-contract.md.

Security & license

• License: MIT - see LICENSE.
• Secret scanning: CI runs gitleaks with its default rules (secret-only) on every push, blocking. Maintainers additionally run a private keyword OPSEC scan before publishing; that config enumerates the terms it looks for, so it is kept out of the public tree (gitignored, local-only) and is not needed to build, test, or use the project. See SECURITY.md.