Canonia

A git-backed, MCP-served knowledge graph for AI coding agents — one canonical source of truth that many repositories reference instead of copying.

Status: pre-alpha (v0.1 feature-complete). Working today: the concept schema, the graph gates (schema + dangling-reference), the importer (canonia import), the MCP server (canonia serve), the optional local semantic index (canonia index), and the static site (canonia build). Governance/access control is a deliberate future module — until then, serve a canon privately (see docs/deploying.md).

The idea

Knowledge that AI agents (and humans) rely on tends to get copied into every repository's CLAUDE.md / AGENTS.md / docs — and then drifts out of sync. Canonia inverts that:

• One canon — a git repository of small, single-concept Markdown files (the source of truth). Versioned, mergeable, authored by humans and agents.
• Reference, not copy — each consuming repo holds a concept id, not a duplicated copy.
• Fetch on demand over MCP — agents query the Canonia MCP server for a concept when they need it, and write updates back so every other session and repo inherits them.
• A graph, not a tree — concepts link to concepts; backlinks and a browsable web view come for free.
• Versioning & authorship from git — every change is a commit, so history and "who (human or agent) changed what" are built in.

 producers            serving layer                consumers
 ─────────            ─────────────                ─────────
 Claude / agents ─▶  Canonia MCP server  ─▶  reference a concept id
 you (git commit) ─▶  (search·get·write)  ─▶  read-only web view (graph)
                          │
                    your canon (git repo = source of truth) ─▶ GitHub

Quickstart

The end-to-end path: install → create a canon → seed it from docs you already have → make it searchable → point your agent at it → browse it. Each step links to its full guide.

# 1. Install (the [semantic] extra adds local embedding search; omit for a lean install)
pip install 'canonia[semantic]'

# 2. Create a canon (a git repo of concepts) and version it
canonia init my-canon --domains process,infra,ops
cd my-canon && git init

# 3. Seed it from markdown you already have — dry-run first, then --commit
canonia import --zero-config ../my-existing-notes --domain process        # preview
canonia import --zero-config ../my-existing-notes --domain process --commit
canonia validate            # schema + dangling-reference gates must pass

# 4. Build the semantic index (enables hybrid keyword+semantic search)
canonia index build

# 5. Point your LLM/agent at the canon over MCP (see the guide for your client)
canonia serve --canon .     # stdio MCP server: search / get / create / update / …

# 6. Browse the graph as a static site (open offline, or serve privately)
canonia build && open site/index.html

• installing — requirements, install options, canonia init.
• importing — zero-config vs curated mapping.yml, dry-run, duplicate detection, and --prune reconciliation.
• using with agents — connect and instruct your LLM (Claude Code, Claude Desktop, Cursor, and other MCP clients) to actually use the canon.
• indexing — the offline semantic index and hybrid search.
• deploying — serve the canon privately (no built-in auth yet).

What each command does

Configuration for all of them lives in one file, canonia.yml (configuring). Concepts retire without breaking links (Wikipedia-style deprecate / merge / archive) — see lifecycle.

Design notes

• Reference, not copy — two link layers: references: frontmatter (the authoritative graph the gate enforces) + [[id]] inline links in prose.
• No locking — git merge / optimistic concurrency instead, so async agent sessions never block each other.
• Permissive writes, strict gate — agents can create a concept that references one that doesn't exist yet (a warning); canonia validate is the hard gate for CI.
• Fully offline & private — the semantic index runs a local ONNX model; a private canon never leaves your machine. The site makes zero external requests.
• Governance is a future module — a no-op access seam (access.py) is wired now and will scope humans and LLM identities later.

Contributing / from source

git clone https://github.com/canonia/canonia && cd canonia
pip install -e ".[dev]"     # editable install + test/lint deps
pytest -q                   # test suite
ruff check src tests && mypy # lint + type gate (also enforced in CI)

License

Apache-2.0 © 2026 André Lopes