---
url: /handbook/connect.md
description: >-
  The ways AI tools and outside services connect to DutyGraph, and which are
  built or planned.
---

# Connect

DutyGraph connects in two directions. AI tools call DutyGraph over the Model Context Protocol (MCP). DutyGraph calls outside services for AI, email, payments, the graph view and phone notifications.

## Three MCP surfaces

DutyGraph has three MCP surfaces. Each has its own credential and its own job. They never share a token.

| Surface | Who uses it | Endpoint | Credential | What it can do |
| --- | --- | --- | --- | --- |
| **Advisor MCP** | An advisor's AI assistant, such as Claude Code | `POST /api/mcp` | Personal access token `dgp_…` with `read`, `write` and optionally `send_email` | Work with accounts, research, the Company Snapshot, Company Repo files and outreach. See [MCP for advisors](mcp.md). |
| **Employee request tools** | An employee's own AI chat | `POST /api/mcp` | Personal token with only `request_agents`, bound to one person | Ask for an agent inside your own access and follow the request. See [Request from your AI](../governance/request-from-your-ai.md). |
| **Agent gateway** | An issued agent | `POST /api/mcp/agent` | One `dga_…` credential per issued agent | Call the agent's approved tools. Every call is checked. See [Agent gateway](../governance/agent-gateway.md). |

::: warning
**No MCP surface can decide.** No tool approves, changes or denies an agent request. The governance decide, issue, revoke and end-event routes refuse calls that come from MCP. Decisions happen in DutyGraph.
:::

Every token is shown once. DutyGraph stores only a hash. A person can revoke a token at any time.

## Integrations

DutyGraph uses a small set of outside services. Apart from hosting, each one is optional and stays off until a workspace, a company or the operator sets it up.

| Area | Service | Status |
| --- | --- | --- |
| AI | The company's approved Claude (Anthropic) or OpenAI connection | Built |
| Email | Resend | Built |
| Payments | Stripe | Built, off unless keys are set |
| Graph view | Neo4j Aura, a copy of PostgreSQL | Built, optional |
| Phone notifications | Web push | Built |
| Hosting | Vercel, with a PostgreSQL database | Built |

See [Integrations](integrations.md) for how each one connects and what it receives.

## What is planned

| Item | Status |
| --- | --- |
| OAuth sign-in for advisor tools from web AI connectors (the CLI and employees' Claude and ChatGPT connectors already sign in with OAuth) | Planned |
| Read-only connections to Okta, Entra ID and Saviynt for holdings and rules | Planned |
| Creating the access request in ServiceNow | Planned |
| Single sign-on (SSO) for reviewers and employees | Planned |
| Run tools in the employee's chat, and run tracking from OpenTelemetry data | Planned |

Today, holdings come from an uploaded access-export CSV, and tokens are pasted into the AI tool. In the governance sample, Okta, Entra ID, Saviynt and ServiceNow are **Simulated**.

## In this section

| Page | What it covers |
| --- | --- |
| [MCP for advisors](mcp.md) | Tokens, scopes, the advisor tools and how to connect Claude Code and Claude Desktop |
| [MCP and CLI for partners](partners.md) | Every advisor tool and CLI command, sign-in, and which clients a partner firm sees |
| [Integrations](integrations.md) | AI providers, email, payments, the graph, phone notifications, hosting, and what is planned |
| [Request from your AI](../governance/request-from-your-ai.md) | The four employee request tools |
| [Agent gateway](../governance/agent-gateway.md) | The check on every call an issued agent makes |

::: info
**Status.** **Built:** the advisor MCP server, the employee request tools, the agent gateway, token sign-in, and the integrations listed above. **Simulated:** the business systems behind the gateway and the identity systems in the governance sample. **Planned:** OAuth sign-in, live identity and IT service management connections, SSO and run tracking.
:::
