The mabl MCP server gives an AI coding assistant access to your organization's mabl workspace. This article explains how the mabl MCP authenticates, where the data goes, and which models see the data.
What the mabl MCP server is
The mabl MCP server lets an AI coding assistant such as Claude, Cursor, or GitHub Copilot work with your organization's mabl tests, including analyzing results, investigating failures, authoring and editing tests, and triggering runs. The MCP server inherits the permissions of the credential it authenticates with: a mabl user's role or an API key with scoped permissions.
mabl hosts the server, so there is nothing to install or run inside your network. Your team enables it by adding the server to the configuration of the AI coding assistant they already use. See Set up the cloud MCP server for the steps your team follows.
Architecture and network
The mabl MCP server is a hosted service in the mabl cloud. It sits between the AI client and the mabl platform API, and it reaches workspace data only through that API, the same path the mabl app, the mabl CLI, and the mabl desktop app use. The following diagram shows where each piece of mabl software runs and how it connects to the mabl cloud.
Reaching private applications
The MCP server never connects to your applications. However, it can trigger mabl test runs, and those runs reach your applications the same way as other mabl cloud runs:
- Public applications can be accessed directly.
- Applications inside your network can be accessed through mabl Link, a small service you run in your own network.
Link opens an outbound tunnel to the mabl cloud, and cloud test runs travel back through that tunnel, so no inbound firewall rule is needed. Tunnels use 4096-bit RSA keys specific to your workspace and agent.
If your environments resolve publicly, allowlisting mabl's static IP addresses is an alternative to Link. Environments with internal-only DNS require Link.
Outbound connections only
Every connection between your environment and the mabl cloud starts on your side, as outbound HTTPS on port 443: the AI client calls the MCP server, the CLI and desktop app call the platform API, and mabl Link opens a tunnel from inside your network. If your organization blocks outbound traffic, see Setting up your network for mabl for the domains to allowlist.
Identity and access
Every piece of mabl software that connects to the mabl cloud, including the MCP server, authenticates as an identity your organization controls and can revoke in the mabl app.
The MCP server accepts two kinds of credentials:
- OAuth sign-in, which authenticates as a mabl user and carries that user's workspace roles and permissions.
- A workspace-scoped mabl API key, which only workspace owners can create. The MCP server accepts Editor keys, which carry the editor role's permissions, and Viewer keys, which are read-only. Access is limited to the workspace the key was generated in and lasts until the key expires or is deleted. See Managing API keys and API key types.
With single sign-on configured, your identity provider governs who can sign in to mabl, and therefore who can authenticate the MCP server with OAuth.
For more details on workspace-level controls such as role alignment, minimally scoped keys, and key expiration, see Recommended security best practices.
Data protection
mabl is SOC 2 Type II certified. All customer data is stored in the United States. mabl encrypts data in transit with TLS 1.2 or later and at rest with AES-256-GCM, using workspace-specific keys managed by Google Cloud KMS. For the full list of standards, the retention policy, and key management details, see Data security and privacy.
AI models and your data
Two sets of AI models are involved when your organization uses the mabl MCP server: the model behind your AI client, and the models that power mabl's own capabilities. They run independently of each other, even if they come from the same provider.
Your AI client's model
The AI client sends its context to a model endpoint your organization already chose: a vendor API, or models hosted through Amazon Bedrock, Google Vertex AI, or Microsoft Azure. mabl doesn't select, host, or contract with that model, and the mabl MCP server never connects to it.
Enabling the MCP server means only that mabl test data becomes eligible to appear in that context, alongside the code the AI client already sees. Whatever terms your organization has with that model provider govern mabl data the same way they govern your code.
mabl's models
mabl's own capabilities, such as agentic test authoring, results analysis, and auto-heal, use commercial off-the-shelf models through Google Vertex AI. Neither mabl nor Google Cloud uses your data to train those models. The mabl agent runs in the mabl cloud and doesn't send data to your AI client's model. See How mabl enhances your testing with AI.