# Model Context Protocol (MCP)

> Connect Claude, ChatGPT, Google Gemini, Cursor, VS Code, or CLI agents to Glow over MCP. Build, validate, test, and debug automations on a live canvas.

Connect your AI app or editor to Glow over MCP to build, validate, and test automations live. Supported clients include `Claude`, `ChatGPT`, `Google Gemini`, `Cursor`, `VS Code`, `Zed`, `Windsurf`, `Claude Code`, `Codex CLI`, and `Gemini CLI`.

**At a glance**

- **Use it to:** Connect Claude, ChatGPT, Gemini, Cursor, VS Code, or Zed to build, validate, and test automations over MCP
- **Find it:** Your account → MCP (?tab=mcp) and Team settings → MCP (?section=mcp)
- **Cost:** 0 Step credits and 0 Glow AI tokens for building and validating · Step tests use normal rates
- **Limits:** 16 to 26 tools by permission level (Read only, Build drafts, Full access)

Glow also acts as an MCP client: canvas [AI Agent Steps](https://docs.getglow.ai/developers/mcp/agent-tools) can call reviewed external MCP servers.

## [Connect in 30 Seconds](https://docs.getglow.ai/developers/mcp#connect-in-30-seconds)

### 1. Open Quick Connect

Open **Your account → MCP** (`/app/user/account?tab=mcp`) or click **Connect AI tool** in the **Workflow Assistant** header (**Admin**, **Manager**, or **Member** role).

### 2. Pick your AI client

Copy the **Server URL** (`https://api.getglow.ai/mcp`) and **Workspace ID**, or click **Add to Cursor** / **Add to VS Code**.

### 3. Approve permissions and verify

Sign in with Glow and pick **Read only**, **Build drafts (Recommended)**, or **Full access**. Call `list_workspaces` in a new chat ([setup commands](https://docs.getglow.ai/developers/mcp/client-setup)).

### Which Tools Your Client Sees (by Permission Level)

Glow filters `tools/list` by your connection's preset in **Your account → MCP**:

| Permission Level / Grant                              | Tools        | Included capabilities                                                                                                                                                                                                                                                                   |
| :---------------------------------------------------- | :----------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Read only**                                         | **16 tools** | Discovery, schemas, graph reads, `validate_workflow`, run traces (`get_workflow_execution`), and version diffs.                                                                                                                                                                         |
| **Build drafts (Recommended)** or **Full access**     | **25 tools** | All 16 read tools + 6 draft tools (`create_workflow`, `import_workflow`, `update_workflow`, `delete_workflow_draft`, `save_workflow_version`, `restore_workflow_version`) + 3 test tools (`test_step`, `run_workflow_test`, `cancel_workflow_execution`). Legacy no-test: **22 tools**. |
| **Full access** + **Manage Live automations** (Admin) | **26 tools** | All 25 tools + `set_workflow_status` (publish or pause Live automations).                                                                                                                                                                                                               |

After editing **Your account → MCP → Change permissions**, **start a new chat** to refresh `tools/list` ([troubleshooting](https://docs.getglow.ai/developers/mcp/tool-reference#missing-tools-troubleshooting)).

## [Why Build & Debug Through Glow MCP](https://docs.getglow.ai/developers/mcp#why-build-through-mcp)

**Build Without Copy-Pasting JSON** (Live Canvas & Schema Grounding)

- Queries real app schemas and live picker options (channels, sheets, folders) in batches.
- Commits atomic update_workflow batches while your open Glow tab frames changes live.
- Discovery, building, and validation cost 0 Glow AI tokens and 0 Step credits.

**Debug Without Switching Tabs** (Validation, Testing & Run Traces)

- validate_workflow catches missing fields, broken routes, and invalid references before runs.
- list_workflow_executions and get_workflow_execution inspect step traces with secrets masked.
- test_step and run_workflow_test verify drafts with save_workflow_version checkpoints.

### How a Typical MCP Build Loop Works

[4. Test Steps or Run](https://docs.getglow.ai/developers/mcp/testing)

### [Prompts You Can Try Right Away](https://docs.getglow.ai/developers/mcp#example-prompts)

Ask in plain English or run a slash command (`/build-automation`, `/debug-failed-run`, `/explain-automation`, `/optimize-automation`, `/test-and-verify-draft`):

  

**Build Draft**

    _"Build a draft that receives a webhook lead, classifies it with AI Prompt,
    gates VIP deals with Human Review, and posts to Slack."_ Commits an atomic
    `update_workflow` draft and returns `connectUrl` links.
  

  

**Debug Failure**

    _"Check why 'Stripe to NetSuite Sync' failed today and fix the draft."_
    Calls `list_workflow_executions` and `get_workflow_execution` (secrets
    masked), patches the Step, and runs `validate_workflow`.
  

  

**Test Step**

    _"Test Step 3 on 'Customer Onboarding' with sample data from the last run."_
    Runs `test_step` on Step 3 alone using its connected account.
  

  

**Audit**

    _"/explain-automation 'Quarterly Renewal Alerts'"_ Reads the graph,
    branches, connected account names, and trigger settings, then summarizes the
    logic.
  

## [Choose Your Task](https://docs.getglow.ai/developers/mcp#the-dual-plane-architecture)

| I Want To                                                | Start Here                                                   |
| :------------------------------------------------------- | :----------------------------------------------------------- |
| Connect ChatGPT, Claude, Gemini, Cursor, VS Code, or Zed | [Connecting AI Apps & Editors](https://docs.getglow.ai/developers/mcp/client-setup) |
| Discover components, build drafts, or save checkpoints   | [MCP Tool Reference](https://docs.getglow.ai/developers/mcp/tool-reference)         |
| Run a Step or whole-automation test                      | [Testing Automations with MCP](https://docs.getglow.ai/developers/mcp/testing)      |
| Let a canvas AI Agent call an external MCP server        | [Glow as an MCP Client](https://docs.getglow.ai/developers/mcp/agent-tools)         |
| Set workspace budgets and review permissions             | [MCP Administration](https://docs.getglow.ai/manage/workspace-settings/mcp)         |

> Compare inbound editor requests and outbound AI Agent tool calls in the [MCP bridge walkthrough](https://docs.getglow.ai/developers/mcp) on the web version of this page.

## [Architecture & Governance Controls](https://docs.getglow.ai/developers/mcp#enterprise-architectural-guarantees)

Building, executing, and administering require separate grants.

  

**1. Atomic Batches**

    [Batch edits](https://docs.getglow.ai/developers/mcp/tool-reference#update_workflow-mutations) save atomically while canvas sync reports separately. Read saved state instead of repeating a committed batch. Use `dryRun: true` to check edits without saving.
  

  

**2. Graph Validation**

    [Static graph validation](https://docs.getglow.ai/developers/mcp/tool-reference#static-graph-validation-validate_workflow) flags missing settings, broken routes, missing connections, and uncertain references before runs without calling external apps.
  

  

**3. Credential Isolation**

    Bind accounts by ID (`accountId` or `connectionId`) rather than placing secrets in prompts or parameters. OAuth tokens and API keys stay in Glow's encrypted store ([account discovery](https://docs.getglow.ai/developers/mcp/tool-reference#discover-components-and-account-backed-fields)).
  

  

**4. RBAC & Scoping**

    [Connection grants](https://docs.getglow.ai/manage/workspace-settings/mcp#minimum-token-permissions) scope access per client. Step and workflow tests require separate grants and execute real actions ([testing prerequisites](https://docs.getglow.ai/developers/mcp/testing#before-you-test)).

    MCP stops **only MCP test runs** (`cancel_workflow_execution`); **production runs must be stopped in the Glow UI**. Calling `set_workflow_status` requires an Admin connection with **Manage Live automations** (`Manages Live` badge). Billable runs share balances under [workspace safeguards](https://docs.getglow.ai/manage/workspace-settings/mcp#workspace-safeguards).

  

## [Frequently Asked Questions](https://docs.getglow.ai/developers/mcp#faq)

  

**Live Safety & Undo**

    **Can an AI client change a Live automation?**
    Only when permitted. **Build drafts (Recommended)** reads all automations and runs (Live included) while editing and testing drafts only. Under **Full access**, editing or publishing a Live workflow requires an Admin to enable **Manage Live automations** in **Your account → MCP** and pass `allowActiveWorkflowMutation: true`. Admins can cap workspace permissions in **Team settings → MCP**, and any connection can be scoped to 100 automations.

    **Can I undo AI client edits?**
    Yes. Press `Cmd+Z` / `Ctrl+Z`, or use `save_workflow_version`, `get_workflow_history`, `diff_workflow_versions`, and `restore_workflow_version`. Live edits save an automatic `checkpoint` version first.

  

  

**Debugging & Billing**

    **Can my AI client debug failed runs?**
    Yes. `list_workflow_executions` and `get_workflow_execution` read step traces with secrets masked, then `update_workflow` and `validate_workflow` apply and verify the fix.

    **Does building with MCP use tokens or Step credits?**
    No. Discovery, authoring (`create_workflow`, `update_workflow`), history reads, and validation cost 0 tokens and 0 credits. Balance is used only by `test_step`, `run_workflow_test`, or `generate_workflow`.

  

  

**Server vs. Client**

    **How do Glow MCP Server and Client differ?**
    **Glow as an MCP Server** lets external AI apps (`Claude`, `ChatGPT`, `Gemini`, `Cursor`) build and test automations in your workspace. **[Glow as an MCP Client](https://docs.getglow.ai/developers/mcp/agent-tools)** lets a canvas **AI Agent** Step call tools on an external HTTPS MCP server under **Connections → Agent tools**.
  

**Verify it works**

- **Run:** Open Your account → MCP (?tab=mcp), connect via Quick Connect, and call list_workspaces
- **You should see:** Your AI client lists your Glow workspace and can inspect or build draft automations.

## What's Next?

- 👉 **[Connect your AI app or editor](https://docs.getglow.ai/developers/mcp/client-setup)**, then build and validate before testing.
- **[Glow as an MCP Client](https://docs.getglow.ai/developers/mcp/agent-tools)**: connect external MCP tools for canvas AI Agents.