Connect your own agent

Connect Cursor, Claude Desktop, or another MCP client to Anything Graph so your agent can query your data through the reasoning layer — not with raw database access.

The MCP server at http://127.0.0.1:3334/mcp is a thin bridge to the Rust reasoning service at http://127.0.0.1:8787. Agents call MCP tools; MCP forwards requests to the reasoning API.

Before you connect

Make sure Anything Graph is running:

anythinggraph start

Check that both services are up:

anythinggraph status

Service	URL
Reasoning API	`http://127.0.0.1:8787`
MCP (Cursor / Claude)	`http://127.0.0.1:3334/mcp`

If you have not connected data sources yet, complete Connect your data first.

Connect in Cursor

Run anythinggraph start.
Open Cursor → Settings → MCP.
Add the MCP server URL: http://127.0.0.1:3334/mcp.
Or run anythinggraph mcp print-config and paste the generated JSON snippet.

Example Cursor MCP config (local dev, no auth):

{
  "mcpServers": {
    "anythinggraph-cli": {
      "url": "http://127.0.0.1:3334/mcp"
    }
  }
}

When auth is enabled, add an Authorization header (admin for authoring, user for query-only):

{
  "mcpServers": {
    "anythinggraph-cli": {
      "url": "http://127.0.0.1:3334/mcp",
      "headers": {
        "Authorization": "Bearer admin-secret-change-me"
      }
    }
  }
}

Print a ready-to-paste config from the CLI:

anythinggraph mcp print-config

Connect in Claude Desktop

Use the Claude-specific config snippet:

anythinggraph mcp print-config --target claude

Paste the output into your Claude Desktop MCP configuration.

Verify the connection

In your agent, ask a simple health or discovery question:

Using anythinggraph-cli MCP: list available playbooks and run a health check.

Typical query flow for everyday use:

list_playbooks — see loaded playbook ids
get_playbook_context — read entities and relationships for a playbook
query_graph — ask a business question (plan + execute + proof)

Example prompt (replace the playbook id and names with yours):

For playbook crm-payroll-access: how many accounts does Alex Anderson own?

!!! note "Read-only by design" Live data sources accept SELECT-style reads only. MCP has no insert, update, or delete tools.

MCP tool reference

The anythinggraph-cli MCP server (http://127.0.0.1:3334/mcp) exposes the tools below. Which tools your agent can call depends on the bearer token role (user or admin).

Tool	Purpose
`health_check`	Ping the Rust reasoning service
`list_playbooks`	List playbook ids loaded from `playbooks/`
`get_playbook_context`	Load playbook schema summary (entities and relationships)
`list_entity`	List rows for a playbook entity (bounded browse; default limit 1000)
`sample_entity`	Return a small sample of rows for a playbook entity (default limit 5)
`plan_query`	Compile a structured federated query into plan IR
`execute_plan`	Execute a compiled plan IR via read-only adapters
`query_graph`	Compile and execute a federated read query in one step (proof envelope)
`list_allowed_rows`	List row identifiers a subject may read under enforced ReBAC rules

Admin tools

Admin tokens include all user tools plus:

Tool	Purpose
`list_sources`	List configured data sources from `profiles/local.yaml` (no secrets)
`get_adapter_guide`	Per-adapter binding authoring guide for a `source_id` — call before `propose_binding`
`list_bindings`	List loaded binding file stems in `bindings/`
`get_binding`	Load one saved binding YAML by stem
`introspect_source`	Read source schema for agent mapping (tables/columns — read-only)
`sample_source`	Read a few raw rows from a source table/collection/object (no playbook required)
`suggest_bindings`	Suggest playbook entity-to-table mappings from introspected schema
`propose_binding`	Validate declarative binding YAML (no SQL)
`test_binding`	Compile a sample query against a proposed or saved binding; optionally execute read-only
`save_binding`	Save declarative binding YAML to `bindings/{playbook_id}.{adapter_suffix}.yaml`
`propose_playbook`	Validate compact playbook JSON
`save_playbook`	Save compact playbook JSON to `playbooks/{playbook_id}.json`

!!! note "Profiles are manual" Profiles are never written via MCP — edit profiles/local.yaml yourself. Admin agents use list_sources → get_adapter_guide → introspect_source before authoring bindings.

Next steps

Follow the playbook authoring path: