Download .skills/wiki-export/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does wiki-export live?
The publisher keeps it at https://github.com/Ar9av/obsidian-wiki.
Which agents can use wiki-export?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.
>
name: wiki-export
description: >
Export the Obsidian wiki graph to JSON, GraphML, Neo4j Cypher, Postgres/SQL, HTML, or OKF bundles. Use when transferring or visualizing wiki data in external tools; wiki-import handles the reverse direction.
Wiki Export — Knowledge Graph Export
You are exporting the wiki's wikilink graph to structured formats so it can be used in external tools (Gephi, Neo4j, custom scripts, browser visualization).
Before You Start
Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → global config → prompt setup). This gives OBSIDIAN_VAULT_PATH
Confirm the vault has pages to export — if fewer than 5 pages exist, warn the user and stop
Project Filter (optional)
If the user's invocation includes a project name — e.g. /wiki-export prismor, "export the prismor project", "export project:security" — activate project filter mode:
Extract the project name from the argument or phrase. Normalise: lowercase, strip the word "project".
Keep only pages where either condition holds:
The page id starts with projects/<name>/ (path-based match)
The page's tags array contains <name> (tag-based match)
Drop any edge where either endpoint was excluded.
Note the filter in the summary: (filtered: project:<name> — X of Y pages)
Set graph.graph.filter = "project:<name>" in the JSON output.
If both a project filter and a visibility filter are active, apply both (project filter first, then visibility filter on the remaining set).
Visibility Filter (optional)
By default, all pages are exported regardless of visibility tags. This preserves existing behavior.
If the user requests a filtered export — phrases like "public export", "user-facing export", "exclude internal", "no internal pages" — activate visibility filtered mode:
Build a blocked tag set: {visibility/internal, visibility/pii}
Skip any page whose frontmatter tags contain a blocked tag when building the node list
Skip any edge where either endpoint was excluded
Note the filter in the summary: (filtered: visibility/internal, visibility/pii excluded)
Pages with no visibility/ tag, or tagged visibility/public, are always included.
Step 1: Build the Node and Edge Lists
Glob all .md files in the vault (excluding _archives/, _raw/, _readouts/, .obsidian/, index.md, log.md, _insights.md). Apply any active filters (project and/or visibility) after collecting the full file list.
For each page, extract from frontmatter:
id — relative path from vault root, without .md extension (e.g. concepts/transformers)
label — title field from frontmatter, or filename if missing
For each page, Grep the body for \[\[.*?\]\] to extract all wikilinks:
Parse each [[target]] or [[target|display]] — use the target part only
Resolve the target to a node id (normalize: lowercase, spaces→hyphens, strip .md)
Skip links that point outside the node list (broken links)
Each resolved link becomes an edge: {source: page_id, target: linked_id, relation: "wikilink", confidence: "EXTRACTED"}
If the linking sentence ends with ^[inferred] or ^[ambiguous], override confidence accordingly
Typed edge enrichment: After building the wikilink edge list, read each page's relationships: frontmatter block. For each {target, type} entry:
The target YAML value is a quoted wikilink string such as "[[concepts/lstm]]". Strip the surrounding [[ and ]] characters, then apply the same normalization (lowercase, spaces→hyphens, strip .md) to get the node id.
Skip entries whose resolved target is not in the node list (broken link)
If an edge for this (source, target) pair already exists, override its relation field with the typed value (e.g., "contradicts") and set typed: true
If no edge exists yet for this pair, add one: {source: page_id, target: target_id, relation: <type>, confidence: "EXTRACTED", typed: true}
This means relation: "wikilink" is the default for plain untyped links; a relationships: entry promotes it to a named semantic type. Edges that originated from both a body wikilink and a relationships: entry keep a single record — the typed version wins.
This is your edge list.
Step 2: Assign Community IDs
Group pages into communities by tag clustering:
Pages sharing the same dominant tag belong to the same community
Dominant tag = the first tag in the page's frontmatter tags array
Pages with no tags get community id null
Number communities starting from 0, ordered by size descending (largest community = 0)
This enables community-based coloring in the HTML visualization and tools like Gephi.
Step 3: Write the Output Files
Create wiki-export/ at the vault root if it doesn't exist. Write all five files:
3a. graph.json
NetworkX node_link format — standard for graph tools and scripts:
Write one <node> per page and one <edge> per link. For typed edges (those where typed: true in the edge list), emit both <data key="relation"> with the semantic type value and<data key="type"> with the same value — this keeps relation readable for tools that already consume it while letting type-aware tools filter on the dedicated type key. Untyped wikilinks omit the <data key="type"> element entirely.
3c. cypher.txt
Neo4j Cypher MERGE statements — paste into Neo4j Browser or run with cypher-shell:
// Wiki knowledge graph export — <TIMESTAMP>
// Load with: cypher-shell -u neo4j -p password < cypher.txt
// Nodes
MERGE (n:Page {id: "concepts/transformers"}) SET n.label = "Transformer Architecture", n.category = "concepts", n.tags = ["ml","architecture"], n.community = 0;
MERGE (n:Page {id: "entities/vaswani"}) SET n.label = "Ashish Vaswani", n.category = "entities", n.tags = ["person","ml"], n.community = 0;
MERGE (n:Page {id: "concepts/lstm"}) SET n.label = "LSTM", n.category = "concepts", n.tags = ["ml","rnn"], n.community = 0;
// Relationships
// Untyped wikilinks use [:WIKILINK]
MATCH (a:Page {id: "concepts/transformers"}), (b:Page {id: "entities/vaswani"}) MERGE (a)-[:WIKILINK {relation: "wikilink", confidence: "EXTRACTED"}]->(b);
// Typed edges use the relationship type as the label (UPPERCASE)
MATCH (a:Page {id: "concepts/transformers"}), (b:Page {id: "concepts/lstm"}) MERGE (a)-[:CONTRADICTS {relation: "contradicts", confidence: "EXTRACTED"}]->(b);
Write one MERGE node statement per page, then one MATCH/MERGE relationship statement per edge. For typed edges, use the type value uppercased as the Cypher relationship label (e.g., contradicts → [:CONTRADICTS], derived_from → [:DERIVED_FROM]). Untyped wikilinks always use [:WIKILINK].
3d. postgres.sql
Plain SQL — loadable into any Postgres database (local, Supabase, RDS, Neon, …) with psql -f postgres.sql or a migration runner. Two tables: wiki_pages (nodes) and wiki_edges (links), with ON CONFLICT upserts so re-running the export is safe and idempotent, mirroring the MERGE semantics of cypher.txt.
-- Wiki knowledge graph export — <TIMESTAMP>
-- Load with: psql -d yourdb -f postgres.sql
CREATE TABLE IF NOT EXISTS wiki_pages (
id TEXT PRIMARY KEY,
label TEXT NOT NULL,
category TEXT,
tags JSONB NOT NULL DEFAULT '[]'::jsonb,
summary TEXT,
community INT
);
CREATE TABLE IF NOT EXISTS wiki_edges (
source TEXT NOT NULL REFERENCES wiki_pages(id) ON DELETE CASCADE,
target TEXT NOT NULL REFERENCES wiki_pages(id) ON DELETE CASCADE,
relation TEXT NOT NULL DEFAULT 'wikilink',
confidence TEXT,
typed BOOLEAN NOT NULL DEFAULT false,
PRIMARY KEY (source, target, relation)
);
CREATE INDEX IF NOT EXISTS wiki_edges_source_idx ON wiki_edges(source);
CREATE INDEX IF NOT EXISTS wiki_edges_target_idx ON wiki_edges(target);
-- Nodes
INSERT INTO wiki_pages (id, label, category, tags, summary, community)
VALUES ('concepts/transformers', 'Transformer Architecture', 'concepts', '["ml","architecture"]'::jsonb, 'The attention-based architecture introduced in Attention Is All You Need.', 0)
ON CONFLICT (id) DO UPDATE SET
label = EXCLUDED.label, category = EXCLUDED.category, tags = EXCLUDED.tags,
summary = EXCLUDED.summary, community = EXCLUDED.community;
-- Edges
-- Untyped wikilink
INSERT INTO wiki_edges (source, target, relation, confidence, typed)
VALUES ('concepts/transformers', 'entities/vaswani', 'wikilink', 'EXTRACTED', false)
ON CONFLICT (source, target, relation) DO UPDATE SET confidence = EXCLUDED.confidence, typed = EXCLUDED.typed;
-- Typed edge from relationships: block
INSERT INTO wiki_edges (source, target, relation, confidence, typed)
VALUES ('concepts/transformers', 'concepts/lstm', 'contradicts', 'EXTRACTED', true)
ON CONFLICT (source, target, relation) DO UPDATE SET confidence = EXCLUDED.confidence, typed = EXCLUDED.typed;
Write one INSERT ... ON CONFLICT (id) DO UPDATE statement per page (values escaped: single quotes doubled, tags serialized as a JSON array literal cast to jsonb), then one INSERT ... ON CONFLICT (source, target, relation) DO UPDATE per edge. relation stays lowercase here (unlike the uppercased Cypher relationship label) since it's a plain column value, not a schema identifier — this keeps it directly filterable with WHERE relation = 'contradicts' or joinable without case-folding. typed is true only for edges promoted by a relationships: frontmatter entry; plain wikilinks stay false.
Skip pages whose id collides only after the ON CONFLICT clause fires from a prior run — do not attempt to deduplicate synthetic multi-edges (e.g. same source/target with both a wikilink and a typed relation) since the composite primary key (source, target, relation) already keeps them as distinct rows, matching the "typed version wins" merge behavior of graph.json/graph.graphml at the query layer (SELECT * FROM wiki_edges WHERE source=$1 AND target=$2 ORDER BY typed DESC LIMIT 1).
3e. graph.html
A self-contained interactive visualization using the vis.js CDN (no local dependencies). The user opens this file in any browser — no server needed.
Build the HTML file by:
Generating a JSON array of node objects for vis.js:
Replace /* NODES_JSON */ and /* EDGES_JSON */ with the actual JSON arrays you generated in step 1.
Step 3.5: OKF Bundle Export (optional)
Run this step only when the user asks for OKF / a markdown bundle (phrases like "export to OKF", "OKF bundle", "open knowledge format", "export as markdown bundle"). It is additive — the five graph files above are always produced; this writes an extra full-fidelity markdown bundle.
The five graph files are a lossy projection (graph skeleton only). An OKF bundle is the actual page bodies, so an export→wiki-import round-trip through OKF preserves full content, and the bundle drops straight into MkDocs, Notion, Hugo, GitHub's renderer, or any Open Knowledge Format consumer.
Our one-line summary: is exactly OKF's description (used in index.md entries).
tags
tags
tags
Verbatim list. visibility/* system tags pass through unchanged.
generated
updated → generated.at; by is the producer
updated ← generated.at
Write generated: { by: obsidian-wiki/<version>, at: <datetime> }, taking <version> from obsidian-wiki --version (plain obsidian-wiki if the CLI isn't installed). OKF §5 requires every timestamp to be a datetime with an explicit offset, so a date-only updated: 2026-04-12 becomes 2026-04-12T00:00:00Z. Replaces v0.1's timestamp, which is no longer written.
sources
sources (list of strings)
sources (list of strings)
OKF sources is a list of objects with a required resource, so each native string becomes - resource: <string>. Opaque strings like conversation:2026-04-12 are valid: §5.1 allows scope descriptors a consumer can't follow. Emit nothing else per entry, so import recovers the exact strings.
status
lifecycle
— (native lifecycle is preserved)
draft → draft; reviewed, verified, disputed → stable; archived → deprecated. Omit status when lifecycle is absent (absent means stable in OKF).
verified
_meta/trust-ledger.json entry for the page
— (preserved verbatim)
Only when the page has a ledger entry: verified: { by: human:vault-owner, at: <entry reviewed_at> }. trust-record only writes entries a human approved with --approved, so the human: actor is accurate and gives the page OKF's human-reviewed tier (§5.3). Never derive verified from lifecycle alone.
resource
first sources: entry iff it is an http(s):// URL
—
Optional; omit when no source URL. Most pages describe abstract knowledge and have none.
OKF §4.1 requires consumers to preserve unknown keys. Writing our native keys as OKF extension frontmatter is what makes the round-trip lossless — on import, preserved category/created/updated/lifecycle are preferred over re-deriving from type/generated/status. (updated rides along because generated.at must be a full datetime, so a date-only updated would otherwise come back as midnight UTC.) sources is not an extension any more: it's a spec field (row above).
Steps
Reuse the node list from Step 1 (with any active project/visibility filters already applied). Write a directory tree under wiki-export/okf/:
One file per in-scope page. For each page, parse its frontmatter, apply the mapping table above to build the OKF frontmatter (required type first, then title, description, tags, generated, status, verified, sources, optional resource, then the preserved extension keys), transform the body links (below), and write to wiki-export/okf/<category>/<slug>.md — same relative path the page has in the vault.
Body link transform ([[wikilinks]] → standard markdown links):
[[concepts/transformers]] → [<target title>](<file-relative path>.md), e.g. from entities/foo.md a link to concepts/transformers becomes [Transformer Architecture](../concepts/transformers.md). Link text = the target page's title (fall back to the target id if unknown).
[[target|display]] → [display](<rel path>.md).
Use file-relative paths (../concepts/x.md), never/-absolute — /-rooted links break GitHub rendering. (This matches knowledge-catalog's own production agent.)
Compute that relative path from the target file path, not the bare page id: normalize the wikilink target to its page id, append .md, then compute relpath(<target-file>, <source-file-dir>). Never call relpath() on the id before adding .md.
This is required for the common "folder note" layout where a page id exists both as a file and as a directory prefix, e.g. projects/social-twitter.md plus projects/social-twitter/.... From projects/social-twitter/concepts/mem0-memory-analysis.md, a link to [[projects/social-twitter]] must export to ../../social-twitter.md, not ...md.
Resolve link targets with the same normalization used in Step 1 (lowercase, spaces→hyphens, strip .md). Handle unresolved targets by form, so forward-references survive the round-trip:
Resolves to an in-scope page → relative markdown link to it.
Path-form target (contains a /, e.g. [[concepts/attention-mechanism]]) with no page yet, and not excluded by an active filter → still emit the relative markdown link. OKF §11 forbids rejecting a bundle over a broken cross-link, and keeping the link makes the user's forward-references lossless on re-import. (Verified on st3ve: dropping these silently deleted real [[wikilinks]].)
Excluded by an active project/visibility filter → plain text. Do not emit a path pointing into filtered-out content.
Bare-title target with no match (e.g. [[tractorex]] when no such page exists in scope) → plain text; there is no reliable path to write.
Leave existing external http(s):// links and # Citations sections untouched.
Generate index.md files (OKF §8 progressive disclosure; these contain no per-entry frontmatter):
Bundle root wiki-export/okf/index.md — a # Subdirectories section listing each category folder: * [<category>](<category>/index.md) - <one-line description of the category>. This is the only index permitted frontmatter: add a single key okf_version: "0.2" (OKF §8, §12).
One index.md per category folder listing its pages: * [<title>](<slug>.md) - <description from the page's summary>.
Write log.md from the vault root's log.md, reshaped to OKF §9, which requires ## YYYY-MM-DD date headings, newest first. Our log is a flat, oldest-first list of - [<timestamp>] VERB key=value … lines, so: start with # Wiki Update Log; group lines by the date part of <timestamp>; emit groups newest first under ## <date>; render each line as * **<Verb>**: <rest of line>, with the verb title-cased (INGEST → Ingest). Keep lines that don't match the pattern under the date of the line above them, verbatim after * . wiki-import ignores log.md, so this costs nothing on the round-trip.
Filters. Honor the same project/visibility filters as the graph export — filtered pages are omitted from the bundle and their inbound links degrade to plain text per step 2.
Excluded from the bundle (same as the graph export): _archives/, _raw/, _readouts/, .obsidian/, _insights.md, _meta/*.base, and the vault root index.md (regenerated above).
Step 4: Print Summary
Wiki export complete → wiki-export/
graph.json — N nodes, M edges (NetworkX node_link format)
graph.graphml — N nodes, M edges (Gephi / yEd / Cytoscape)
cypher.txt — N MERGE nodes + M MERGE relationships (Neo4j)
postgres.sql — N upsert rows (wiki_pages) + M upsert rows (wiki_edges) (any Postgres)
graph.html — interactive browser visualization (open in any browser)
Append this line only when the OKF bundle was produced (Step 3.5):
(filtered: project:prismor — 19 of 67 pages)
(filtered: X of Y pages excluded — visibility/internal, visibility/pii)
Only include lines for filters that were actually applied.
Notes
Re-running is safe — all output files (and the okf/ bundle) are overwritten on each run
Broken wikilinks are skipped — only edges to pages that exist in the vault are exported; in the OKF bundle, a wikilink to a missing/filtered page degrades to plain text
OKF is the lossless format — graph.json reconstructs only stubs on import, while the okf/ bundle preserves full page bodies. Use OKF for vault-to-vault transfer and external markdown tools (MkDocs, Notion, GitHub); use the graph files for analysis tools (Gephi, Neo4j)
The wiki-export/ directory should be gitignored if the vault is version-controlled — these are derived artifacts
graph.json is the primary format — the others are derived from it. If a future tool supports graph queries natively, point it at graph.json
How do I install wiki-export?
Download .skills/wiki-export/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does wiki-export live?
The publisher keeps it at https://github.com/Ar9av/obsidian-wiki.
Which agents can use wiki-export?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.