XTesterExternal Connections
Learn how MCP connects an AI client to XTester, what tasks it supports and where to check the project, connection and access.
External Connections
MCP connects your AI client to XTester Studio tools. Instead of manually moving code and reports into a chat, ask an agent to inspect a project, explain a strategy, make a change or examine an actual backtest result.
You keep working in Studio and can see the selected project. The AI client requests tools; XTester performs supported actions within the access you approve.
Start with Studio running, a project open and a client that supports local MCP, such as Claude Code, Codex, Claude Desktop or Cursor. The separate connection guide covers the setup and configuration formats.
Screenshots show a development build of Studio after the connection interface update. Your installed release may arrange controls differently; use the configuration generated by your own installation.
How the connection works
The flow is AI client → local MCP process → selected XTester Studio window. The client starts the process using the configuration supplied by the application. The transport is stdio, using the process input and output rather than a public website address.
MCP is not a language model. Your AI client supplies its model and account. Connecting an external client does not require configuring the provider of Studio’s built-in chat first; features that use internal AI can have their own setup.
Tasks an agent can help with
Understand a project: inspect its structure, source and parameters, find events and check available API signatures.
Develop a change: edit a strategy through project tools, compile it and explain returned diagnostics.
Prepare data: inspect source capabilities and available history, then perform a supported download with the required access.
Test a strategy: configure and run a backtest, then retrieve its run ID, metrics, trades and equity curve.
Investigate results: compare variants, trace orders through source and logs, and work with saved research and models using the available tools.
The catalog depends on the host version and mode. An agent should use search_tools and get_tool_schema to discover the current contract. An available tool still enforces its permissions and run conditions.
Where to inspect connections
AI Models sidebar → External Connections.
- External Connections shows the connected client and its access.
A row shows the connected client and its access. “Reported by client” means the client supplied the name; it is not independent identity verification. The educational example displays a generic client as Generic MCP client.
Read-only access is for inspecting the project and results. Project changes require suitable access. Writer status is separate: an observer and the current single writer have different roles.
Refresh the list, inspect activity or disconnect a connection here. Check the selected row and window before disconnecting. Writer transfer is offered only for an eligible connection; an old row or read-only connection cannot bypass approval.
Project, access and connection lifetime
A connection selects one exact live Studio window. With several projects or duplicate project windows, use the project name and exact window tag shown by the client.
- Read what the requested access permits before approving.
In Studio’s request, check the project, client name and requested access first. Reading is sufficient for an initial inspection. Access to project changes covers files and calculations but does not replace separate permission for exchange accounts or other protected actions.
Ordinary access belongs to the current connection and ends when it disconnects. A remembered grant, when offered, belongs to a specific agent process and project. Manage it in Settings → External Agents.
Project data and the client model
The local MCP channel and model processing are different parts of the flow. Your AI client may send retrieved context to its model provider. Follow its data settings and your own project policy.
The generated XTester configuration contains the MCP process command and arguments, not exchange secrets. Account and key access is controlled separately. A client may also have its own filesystem and terminal tools, whose permissions are configured in that client.
Next steps
Use configuration generated by your own Studio installation. Start by asking the agent to explain the project, then progress to changes and verification using actual results.