Typed Graph

Give your coding agent a memory it can query.

Stop re-explaining your architecture every session. Typed Graph gives Claude Code, Cursor and any MCP client a typed graph of your design docs and code, and a check that stops it from finishing with broken references.

Set up Claude Code → All install options Read the guide

See it work

Real tg output on an example checkout API with requirements, @tg: annotations and an OpenSpec change folder. A query finds the requirement with no implementation; one typed comment closes it; a typo fails the check.

Claude Code in one command

tg is a CLI your agent can call directly. tg init prints a diff of every file it would touch and writes only with --write.

npm install -g @typedgraph/cli
tg init --write      # CLAUDE.md block, hooks, .mcp.json, two skills
tg check             # does the documentation still match the code?

For Cursor use tg init --agent cursor; for a plain AGENTS.md use --agent agents.

What the agent gets

  1. Search first: a prompt hook reminds it to search, expands [[refs]] in your prompt and adds the top lexical hits. It needs no model, key or network.
  2. Query the graph: tg cypher over sections and code symbols, or the eight tools of tg mcp.
  3. Finish honestly: a stop hook blocks completion while tg check fails, or when several code lines changed with little change in the docs.

Memory engineering, not a pile of notes

Typed edges are memory an agent can traverse. Everything stays plain Markdown a human can review.

// @tg: implements:: [[auth#Token expiry]] {since: 2}
// @tg: -contradicts:: [[design#Stateless sessions]]
function legacyCheck() {}

$ tg cypher "MATCH (c:CodeSymbol)-[:implements]->(s:Section)
             RETURN s.title, count(c)"
⇄

One space with the humans

Your team edits the same files in Obsidian or an editor. What the agent learns is a diff you can read and review.

✓

Checked, not hoped for

Links, symbols, section rules and test-spec coverage are validated in CI and by the agent hook.

≈

Search that runs anywhere

Lexical search works in a fresh clone. Add Ollama or an OpenAI-compatible endpoint for hybrid search and cited GraphRAG.

↗

Compatible with lat.md

It reads the same lat.md/ folders and agrees with lat check, then adds typed edges, Cypher and a visual graph. Compared in detail.

In your AI editor

The extension shows typed backlinks and the graph for Markdown and annotated code, and offers to run tg init for you.

code --install-extension pavlyshyn.typegraph-vscode

VS Code Marketplace · Open VSX (Cursor, Windsurf, VSCodium, Gitpod)

Use an Obsidian vault as memory

Run the sidecar next to your vault to serve Cypher, vector search and cited GraphRAG over MCP.

claude mcp add typed-graph \
  -e OBSIGRAPH_VAULT=/path/to/vault \
  -e OBSIGRAPH_DATA=/path/to/data \
  -- node packages/sidecar/dist/server.mjs --stdio

Sidecar docs · Vault as RAG storage