# Glow as an MCP Client

> Connect AI Agents to reviewed external MCP tools, select permitted actions, and configure human approval policies.

Connect canvas AI Agents to reviewed external MCP servers with scoped tool permissions and human approval gates. Admins register public HTTPS endpoints (Streamable HTTP or SSE) under **Connections → Agent tools → Add external connection**.

**At a glance**

- **Use it to:** Connect canvas AI Agent Steps to external HTTPS MCP servers with Admin-reviewed tool catalogues and per-tool human approval policies
- **Find it:** Workspace Connections → Agent tools → Add external connection, then AI Agent Step → Tools → External tools
- **Limits:** Public HTTPS endpoints only (Streamable HTTP or SSE) · Bearer token or Admin-enabled No authentication · New or changed tools stay withheld until reviewed
- **Instead:** Native App Integration tools in AI Agent → Tools (when the service is already in Glow's built-in app catalogue)

[2. Approve Tool Catalogue](https://docs.getglow.ai/developers/mcp/agent-tools#1-scoped-tool-whitelisting)

To connect external AI apps (`Claude`, `ChatGPT`, `Gemini`, `Cursor`, `VS Code`) to build Automations instead, see [Connecting AI Apps & Editors](https://docs.getglow.ai/developers/mcp/client-setup).

![The Agent tools tab with Add external connection and the workspace switch for allowing external servers without authentication, shown off.](https://docs.getglow.ai/images/docs/workspace/agent-tools.webp)
*Manage external MCP servers under Agent tools. This workspace has no external server configured.*

## [Connection and Tool Governance](https://docs.getglow.ai/developers/mcp/agent-tools#the-enterprise-governance-architecture)

Admins review external connections and tool catalogues first. Builders then select approved tools and configure approval gates on each AI Agent.

> Explore tool-policy checks, read-only calls, and human review paths in the [MCP governance diagram](https://docs.getglow.ai/developers/mcp/agent-tools) on the web version of this page.

### [Choose Between a Native App Tool and an External MCP Server](https://docs.getglow.ai/developers/mcp/agent-tools#choose-between-a-native-app-tool-and-an-external-mcp-server)

An [AI Agent](https://docs.getglow.ai/build/ai-features/ai-agent) can call native Glow integrations and external MCP servers:

**Native App Integration Tool** (Built-In Catalog (kind: integration))

- Use when the service is in Glow's native catalog (Slack, Gmail, Jira, HubSpot).
- Authenticates via workspace app connections under Connections.
- Supports per-account default approval (none or require) plus per-action overrides.

**External MCP Server Tool** (Custom HTTPS Endpoint (kind: mcp))

- Use for internal services, custom tool servers, or specialized APIs.
- Authenticates with a server-level Bearer token or Admin-enabled No authentication.
- Tools are discovered dynamically, approved by an Admin, and governed per tool.

## [1. Select Permitted Tools](https://docs.getglow.ai/developers/mcp/agent-tools#1-scoped-tool-whitelisting)

Saving a connection tests it automatically and opens **Approve current catalogue**. You can also open **View tools** or **Review tools** from the connection row. Builders then select approved tools under **Tools → External tools**.

New or changed tools stay withheld until an Admin approves them. Prompt instructions never bypass tool selection. Over MCP, builders can attach reviewed MCP connections and set `toolApprovalPolicies` via `update_workflow` (see [Configuring `glowAgent` Parameters](https://docs.getglow.ai/developers/mcp/tool-reference#configuring-glowagent-parameters)).

## [2. Human-in-the-Loop Review Gates](https://docs.getglow.ai/developers/mcp/agent-tools#2-human-in-the-loop-review-gates)

Enable **Require approval by default** in the AI Agent's Tools settings. Set per-tool overrides under **Action approval settings**. Any action, including read calls, can require human approval. Approval requests can expire, and review never bypasses permissions.

## [3. Connection Credentials](https://docs.getglow.ai/developers/mcp/agent-tools#3-zero-credential-llm-exposure)

External connections use a **Bearer token** or **No authentication**. Saving **No authentication** also enables unauthenticated external servers for the workspace. Admins can toggle that policy on the **Agent tools** tab. Glow sends bearer tokens server-side so credentials never enter prompts. Use a least-privilege token and require approval for tools that modify data.

**Verify it works**

- **Run:** Attach an MCP server or connected app tool to an AI Agent Step and run a Preview test
- **You should see:** The AI Agent discovers the tool schema and proposes or executes the tool call in the run timeline.

## What's Next?

- **[AI Agent Step Setup Guide](https://docs.getglow.ai/build/ai-features/ai-agent)**
- **[Human Review Step Guide](https://docs.getglow.ai/build/action-steps/user-approval)**
- **[MCP Overview & Architecture](https://docs.getglow.ai/developers/mcp)**