> ## Documentation Index > Fetch the complete documentation index at: https://docs.aixy-gateway.com/llms.txt > Use this file to discover all available pages before exploring further. # Agent tools and MCP > Give agents one controlled connection to your external tools. Aixy's MCP connection lets agents use external services through one endpoint. Administrators choose the servers, project access, and allowed actions; members connect their own accounts. The agent keeps its model connection and execution loop. Use **Agent tools** to connect accounts, inspect connection status, and copy the MCP endpoint. Use **Enable tools** to administer the tool catalog and **Connect an agent** to issue a personal MCP token for an external client. ## What you can configure * Enable GitHub, Slack, Atlassian Jira, Notion, or a compatible custom remote MCP server. * Limit a connection to a project or make it accessible across the organization. * Allow selected actions, or explicitly allow all current and future tools. * Connect personal accounts so each member uses their own provider permissions. * Connect an external agent or try approved actions directly in Playground. Model API keys and MCP tokens are separate. Your model key determines inference access and budgets. Your MCP identity determines which external tools and accounts the agent can use. Enabling a model does not grant tool access. ## How discovery works The agent receives Aixy's search, execution, and connection helpers. It searches for a capability when needed and receives a compact response containing each matching action's name and description. The search service returns compact JSON, preserving complete parameters when included. Aixy forwards the service's text unchanged. Names are semantic, such as `github.get_me`, so the model can identify the action directly. The agent uses the same search helper with a returned `name` to read that action's complete parameters. A single match includes its parameters immediately; a two-action result can include them when the complete response is small. Small supporting actions can include their complete parameters in the same response. Task searches focus on the requested capability instead of offering pages of incidental matches. This reduces context spent on the catalog and on unselected parameters. Server IDs, duplicate labels, revision fields, and normal status envelopes stay out of the model's discovery results. Execution checks the current account, project, tool permission, and connection revision again. The discovery service retains the selected definition internally for the conversation. A changed, expired, or unavailable selection requires a new search. Searches and parameter lookups still consume context and can add model rounds. Token savings depend on the catalog, task, tokenizer, and cache usage. Assess the complete task using actual results and [usage evidence](/observe/usage). ## Set up the workflow Configure servers, account ownership, project access, and action allowlists. Create a personal token and use the dashboard's client configuration. Search for capabilities and approve external actions in a conversation. Configure inference through Aixy for your editor or coding agent. ## Inspect execution Open **Activity**, or the **Activity** action in Agent tools. Model requests and MCP actions share the same timeline. Tool rows show server, action, status, time, request ID and duration. Activity project and user permissions apply, with `mcp:use` required for tool metadata. Accounts allowed to inspect all users see calls within their accessible projects; other accounts see their own. Arguments, results and credentials are excluded from this history. Use **Sessions** to open explicitly correlated model and tool steps; see [Activity](/observe/activity) for the correlation headers, retention and consumption coverage. An unknown outcome or a call left running can mean the external action completed but its response was lost. Check the external system before retrying. Aixy does not automatically retry external tool actions. External agents control their own approval policy; Playground asks for approval before each external action. Keep the `Mcp-Session-Id` returned by MCP initialization on subsequent tool requests. Discovery and execution require that context; opaque action names and requests without context are rejected. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.