Connect sources
anythinggraph source add — data stays in place
Access risk · Hallucination · Token burn
Giving AI agents your data shouldn't feel this risky
Connect your existing systems, define governance once, reuse everywhere.
Agent HUD · governed queries
Ai governance in minutes.
zsh
npm install -g @anythinggraph/cli@latest
anythinggraph onboard
anythinggraph start
How it works
Four steps from live sources to governed agent answers — no warehouse dump, no glue per model.
Inspect data schema
Browse data schema via MCP
Define ontology
Ontology + ReBAC in git
Query via MCP
Agents ask in English — get scoped answers + proof
What you get
Governed queries over live systems — less token burn, less rework, answers you can ship to prod.
Lower token spend
Query only what the playbook needs — not millions of warehouse rows or document chunks in every prompt.
Not months of integration
Install, map a playbook to your sources, and connect MCP. Skip the data-lake and ETL project before your first agent answer.
Data copy cost
Records stay in Postgres, Salesforce, CSV, and APIs where they already live — no migration tax or sync pipeline to maintain.
Fewer bad answers
Early teams report cutting agent hallucinations when answers route through playbooks, bindings, and proof instead of guesswork.
Rules, many agents
Write access and vocabulary once in a playbook — reuse the same governed layer for Claude, Cursor, OpenAI, and other MCP clients.
On every reply
Source records and relationship paths come back with the answer — less audit friction and less time re-checking the agent by hand.
Terraform for agent data access — playbooks and bindings in git, query_graph through a governed MCP layer, proof envelope on every answer.
Connect your data sources
Query systems in place. These adapters ship today with the same playbook + binding model.
PostgreSQL
sql
CSV files
csv
Salesforce
soql
MySQL
mysql
SQL Server
mssql
MongoDB
mongodb
REST / OpenAPI
rest
Ready to deploy to your agent?
Open source on GitHub. Boot the stack, paste MCP config, run your first governed query_graph — minutes, not quarters.