# AnythingGraph

> Build AI-ready knowledge capsules. AnythingGraph is an open-source semantic layer that connects your existing systems, lets you define governance once, and reuse it everywhere — ontology playbooks, policy as code, natural language via MCP, query-in-place without moving data.

Tagline: **Reasoning as Code.**

License: Apache-2.0 · Self-hosted · [GitHub](https://github.com/AnythingGraph)

## What AnythingGraph is

AnythingGraph helps teams build **AI-ready knowledge capsules** — governed semantic layers that sit between your agents (Claude, Cursor, OpenAI, LangChain, Copilot, custom copilots) and your systems of record.

Connect Postgres, Salesforce, MongoDB, MySQL, documents, and more **without copying data**. Define **what things mean** (ontology), **who may see what** (policy as code), and **where data lives** (bindings). Agents ask in natural language; AnythingGraph returns scoped, structured answers with proof.

**One governance model, many agents** — define access and vocabulary once in git, reuse across every MCP client.

## Why teams use it

- **Prompts don't enforce access** — policy belongs in code, not chat instructions.
- **RAG isn't row-level governance** — great for PDFs; not for live CRM, warehouse, and API facts scoped to a person or role.
- **Agent-written SQL is still full database access** — one bad query bypasses intent.
- **Data is scattered** — teams need one ontology across sources without a migration project.
- **Every copilot shouldn't reimplement auth** — one semantic layer, many agents.

## Six things to know

### 1. Ontology, semantic layer, reasoning AI

AnythingGraph is a **semantic layer** for AI: business vocabulary (customers, patients, accounts, orders) and how they **relate** (owns, manages, assigned_to, has_result).

That ontology powers **reasoning AI** — agents traverse relationships and compile questions into read-only plans over live data, instead of guessing from schema dumps or hallucinating joins.

**Knowledge capsules** are scoped, governed graph contexts per playbook — each agent or use case gets a bounded world model, not your entire data lake.

### 2. Policy as code, git-native, AI governance

Access rules live in **playbooks** alongside the ontology — who may see which resources, scoped by business relationships (ReBAC).

Playbooks and bindings are **git-native**: review, diff, approve, and roll back like any critical config. **AI governance** is enforced at query time — before answers reach the model.

Every governed answer can include **proof**: which playbook, subject, sources, and relationship path produced the result.

### 3. Connect multiple data sources without moving data

Query **in place**. Data stays in Postgres, MySQL, SQL Server, Salesforce, MongoDB, documents, and REST APIs — no mandatory ETL, no copy-to-vector-store pipeline for operational facts.

**Bindings** map playbook concepts to physical tables, collections, or API paths per source. One ontology federates across systems — joined through relationships, not bulk export.

### 4. Ontology playbook by AI

Building an ontology used to take quarters. AnythingGraph works with the agents you already use.

Connect MCP, **introspect live sources**, and let your agent **propose an ontology playbook** — entities, relationships, and suggested bindings. Humans review, test, and save to git. Often **minutes, not months**.

### 5. Talk to your data via natural language, MCP

Add AnythingGraph as an **MCP server** in Claude, Cursor, LangChain, or any MCP client. Users and agents ask business questions in **natural language**; the stack routes through governed tools (`query_graph` and related APIs).

No schema dumps in every prompt. No unconstrained SQL from chat. One **governed front door** for every copilot.

### 6. Open source

AnythingGraph is **open source** (Apache-2.0). Run locally or self-host. Playbooks and bindings are plain JSON and YAML in your repo — no vendor lock-in on ontology format.

Primary product: [anything-cli](https://github.com/AnythingGraph/AnythingGraph) · install via `npm install -g @anythinggraph/cli`

Community playbooks: [AnythingGraph/Playbooks](https://github.com/AnythingGraph/Playbooks)

## Who it is for

- **Developers and platform teams** — ship governed knowledge capsules to copilots in days.
- **Security and data owners** — define governance once; enforce scope before data reaches the model.
- **Domain and ops teams** — one shared business vocabulary; consistent answers across agents and apps.

## What AnythingGraph is not

- Not a vector database or document RAG product — use RAG for policies and PDFs; use AnythingGraph for governed structured facts across live systems.
- Not an ETL or data lake — query in place; no required sync pipeline.
- Not prompt-stuffed schema — meaning lives in versioned playbooks, not repeated every turn.

## How to get started

```bash
npm install -g @anythinggraph/cli@latest
anythinggraph onboard --install-daemon
anythinggraph start
```

Add MCP to your agent: `http://127.0.0.1:3334/mcp`

Then: connect sources → introspect → propose playbook → define access → save to git → query in natural language.

Technical reference: [anything-cli README](https://github.com/AnythingGraph/AnythingGraph/blob/main/Sources/OSS/ag-cli/README.md) · [AGENTS.md](https://github.com/AnythingGraph/AnythingGraph/blob/main/Sources/OSS/ag-cli/AGENTS.md)

## Learn more

- [Home](https://anythinggraph.com/) — Build AI-ready knowledge capsules; install and overview
- [Why AnythingGraph](https://anythinggraph.com/why.html) — the control problem and how we solve it
- [Ontology playbooks](https://anythinggraph.com/ontology-playbooks.html) — 20 starter playbooks by data source and use case
- [Use cases](https://anythinggraph.com/usecases.html) — healthcare, finance, military, physical AI, drones, CRM, supply chain, platform
- [Connect in action](https://anythinggraph.com/connect_in_action.html) — animated diagram: sources → AnythingGraph → AI agents
- [Interactive demo](https://anythinggraph.com/demo.html) — scripted walkthrough (clinical PHI, ontology, ReBAC, proof)
- [Documentation](https://anythinggraph.github.io/) — playbooks, bindings, adapters
- [Playbooks guide](https://anythinggraph.github.io/playbooks/)

## Blog

- [Reasoning as Code](https://anythinggraph.com/blog/reasoning-as-code.html) — the missing layer between enterprise data and AI agents
- [Stop moving data for AI](https://anythinggraph.com/blog/stop-moving-data-for-ai.html) — query-in-place vs ETL for agents
- [Why document search is not enough](https://anythinggraph.com/blog/why-document-search-is-not-enough.html) — RAG vs structured facts
- [Blog index](https://anythinggraph.com/blog/index.html)

## Key terms

- **Knowledge capsule** — A scoped, governed semantic layer (playbook + bindings) — a bounded world model for agents, not full database access.
- **Playbook** — Versioned JSON: ontology (entities, relationships) plus optional access policy.
- **Binding** — YAML mapping from playbook concepts to a live source (table, collection, API).
- **Semantic layer** — Business meaning between agents and raw systems of record.
- **Reasoning AI** — Structured traversal and planning over ontology, not free-form SQL or chunk retrieval.
- **Proof** — Auditable metadata on governed answers (playbook, subject, sources, path).

## Repositories

- [AnythingGraph/AnythingGraph](https://github.com/AnythingGraph/AnythingGraph) — OSS monorepo: CLI, reasoning service, MCP, playbooks, bindings
- [AnythingGraph/AnythingGraph](https://github.com/AnythingGraph/AnythingGraph) — OSS monorepo (CLI installer, website)
- [AnythingGraph/Playbooks](https://github.com/AnythingGraph/Playbooks) — example ontology playbooks
- [npm @anythinggraph/cli](https://www.npmjs.com/package/@anythinggraph/cli)

## Contact

- Website: https://anythinggraph.com
- GitHub: https://github.com/AnythingGraph
- Discord: https://discord.gg/b778aXsKYe