Lito Graph

Compile Markdown docs into a knowledge & workflow graph for AI agents. Expose it via MCP so IDEs and assistants can reason, plan, and act over your documentation — not just search it.

Part of the Lito documentation ecosystem.

Why

Traditional docs are flat text optimized for human reading. AI agents need structured, traversable knowledge to:

• Discover capabilities — "What APIs act on the Workspace entity?"
• Plan multi-step tasks — "Walk me through the onboarding workflow, step by step"
• Stay safe — "Does this workflow require human approval? What are the guardrails?"

Lito Graph compiles your Markdown into a typed knowledge graph with nodes (concepts, APIs, workflows, steps) and edges (ACTS_ON, USES_API, NEXT_STEP_OF, etc.), then serves it over MCP.

Quick Start

Use a hosted graph (zero-install)

If your docs site already serves graph.json, any MCP client can connect with one command:

npx -y @litodocs/graph https://docs.example.com/graph.json

Build locally

# Installcd lito-graph
bun install
# Build a graph from your docs
lito-graph build -i ./docs -o ./graph.json
# Inspect it
lito-graph inspect -g ./graph.json --stats
# Start MCP server
lito-graph serve -g ./graph.json

Document Types

Add type to your Markdown frontmatter to create rich graph nodes:

Concept Docs

---
type: conceptcanonical_name: "Workspace"entity_type: "resource"aliases: ["Org Workspace"]related_entities: ["User", "Project"]
---

API Docs

---
type: apioperation_id: "create_workspace"method: "POST"path: "/v1/workspaces"resource: "Workspace"capabilities: ["create"]preconditions: ["user_authenticated"]permissions: ["workspace:write"]
---

Workflow Docs

---
type: workflowworkflow_id: "onboard_new_workspace"goal: "Onboard a new customer workspace."risk_level: "medium"requires_human_approval: true
---

Existing Lito API docs (api: "GET /users") are auto-converted — no migration needed.

CLI Commands

Command	Description
lito-graph build -i ./docs -o ./graph.json	Compile docs into graph
lito-graph serve -g ./graph.json	Start MCP server on stdio
lito-graph serve -i ./docs	Build + serve in one step
lito-graph inspect -g ./graph.json --stats	Graph overview
lito-graph inspect -g ./graph.json --nodes	List all nodes
lito-graph inspect -g ./graph.json --orphans	Find disconnected nodes
lito-graph inspect -g ./graph.json --unresolved	Show broken references

MCP Tools

The server exposes 7 tools for agents:

Tool	Description
list_nodes	Filter nodes by type or tag
get_node	Full details by ID or slug
traverse	Walk edges from a node
search	Keyword search across the graph
get_workflow	Complete workflow with steps + linked APIs
find_apis_for_entity	All APIs acting on an entity
get_graph_stats	Overview statistics

IDE Setup

Claude Code

Add to .mcp.json in your project root:

{
"mcpServers": {
"lito-graph": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@litodocs/graph", "https://docs.example.com/graph.json"]
}
}
}

Claude Desktop / Cursor

Add to your MCP settings:

{
"mcpServers": {
"lito-graph": {
"command": "npx",
"args": ["-y", "@litodocs/graph", "https://docs.example.com/graph.json"]
}
}
}

Public Graph Hosting

Build with --base-url so every node gets a full public URL:

lito-graph build -i ./docs -o ./graph.json --base-url https://docs.example.com

Each node gets a url field agents can navigate to:

{
"title": "Create Workspace",
"slug": "/api/create-workspace",
"url": "https://docs.example.com/api/create-workspace"
}

Host graph.json alongside your docs (e.g., in public/), then any agent can connect remotely:

lito-graph serve --url https://docs.example.com/graph.json

Works with any docs framework — Lito, Mintlify, Docusaurus, Fern, VitePress, GitBook. If it has Markdown files, Lito Graph can compile them.

Lito Graph vs Context7 MCP

Lito Graph and Context7 solve different problems. Use both.

	Context7	Lito Graph
Purpose	Look up third-party library docs	Operate on your own product/domain
Data	Flat text chunks (RAG)	Typed nodes + 14 edge types (graph)
Relationships	None — isolated text	ACTS_ON, USES_API, RELATED_TO, etc.
Planning	Single page at a time	Full workflows with ordered steps + linked APIs
Safety	Hope the agent reads the warning	requires_human_approval, risk_level, guardrails as machine-readable fields
Completeness	Top-k search (might miss things)	Graph traversal (guaranteed exhaustive)
Setup	Zero — works instantly	One npx command + optional frontmatter
Privacy	Cloud service	Fully local, self-hosted

When to use what

Scenario	Best tool
"How do I use Express middleware?"	Context7
"What's the full onboarding flow for our platform?"	Lito Graph
"What's the React useState API?"	Context7
"Which of our APIs require admin permissions?"	Lito Graph
"How do I configure Tailwind?"	Context7
"What happens if workspace creation fails?"	Lito Graph

Context7 prevents hallucinating API signatures (external knowledge). Lito Graph prevents unsafe actions on your own systems (internal knowledge).

Graph Data Model

Node Types

• DocNode — standard documentation pages
• ConceptNode — domain entities (Workspace, User, Project)
• ApiNode — callable capabilities (HTTP endpoints, functions)
• WorkflowNode — multi-step procedures with guardrails
• StepNode — individual actions within workflows

Edge Types

• Structural: PARENT_OF, CHILD_OF, NEXT_SECTION
• Semantic: RELATED_TO, DEPENDS_ON, CONTAINS, DEPRECATED_BY
• Capability: ACTS_ON, REQUIRES, EMITS, USES_API
• Procedural: NEXT_STEP_OF, ON_FAILURE_TRIGGER, ESCALATES_TO

License

MIT