MCP Connections

Connect external services through Model Context Protocol, inspect their tools, grant only the needed affordances to each agent, and govern consequential calls with policy and approval.

Manage connections under Integrations. MindGraph supports directory-listed providers and custom HTTPS Streamable HTTP endpoints.

Note:Connecting a server and granting a tool are separate actions. A new connection does not give every agent access to every discovered tool.

MindGraph's MCP server

The public mindgraph-mcp package connects Claude Code, Codex, Claude Desktop, and other MCP clients to the same governed memory used by the dashboard and SDKs.

# Inspect commands and installation targets
npx -y mindgraph-mcp@latest --help

# Explicitly materialize the skills published and granted to this actor
npx -y mindgraph-mcp@latest skills pull --dir .claude/skills

# Import a local SKILL.md as a governed candidate for review
npx -y mindgraph-mcp@latest skills push .claude/skills/reconcile-revenue

mindgraph_capture action="skill" creates a provenance-linked candidate; it never publishes directly.
skills pull and skills push are explicit CLI actions. Hooks and background processes never write skill files.
mindgraph_series_query reads latest values, bounded windows, and aggregates without flooding a model's general retrieval context with raw points.

Connection flow

Choose or inspect. Select a listed provider or inspect a custom endpoint before credentials are stored.
Authorize. Use no auth, a bearer credential, or OAuth as supported by the server.
Synchronize. MindGraph discovers tools and records their schemas in the registry.
Grant. Select the individual tools each agent may invoke.
Govern. Choose gated or automatic connection behavior; organization policies still apply in either mode.
Audit. Review calls, policy decisions, approvals, results, and registry drift.

Credential and endpoint security

Connection management requires a signed-in dashboard user with owner or admin authority; API keys do not manage credential custody.
Credentials are encrypted at rest and are never returned after storage.
Custom endpoints must use HTTPS and pass outbound network validation.
Inspection occurs before persistence so an operator can review the server and tools first.
Disconnecting a connection removes its future authority and cleans up dependent registry state.

Warning:A provider's tool description is not an authorization policy. Review arguments, sensitivity, and side effects before granting a tool or enabling automatic execution.

Approval modes

Mode	Behavior
Gated	Tool calls pause for an authorized human decision before execution.
Automatic	Granted tools may run without a connection-level pause, but published governance policies can still deny or require approval.

Approval records include the agent, connection, tool, arguments, decision, and execution result. If the grant, policy, or tool schema changes while a call is parked, the call must pass the current checks before it can resume.

Availability

Connection custody, registry synchronization, agent grants, and approval resolution are JWT Management API operations used by the dashboard. Connection mutations require an owner or admin user JWT; read and approval routes have their own user-JWT requirements. API keys can read the synchronized MCP registry but cannot manage credentials or approvals.

Note:See the API Reference for the HTTP routes and authentication requirements. Connect and govern the server in Integrations, then the configured agent can use its granted tools at runtime.