LangSmith Tool Gateway is entering public beta as a managed control point for tools used by AI agents. Instead of embedding separate authentication flows and credentials in every agent, teams can register Model Context Protocol servers once and publish a limited tool kit through a single URL.

Each call can run either with the user’s credentials or with an identity assigned to the agent. Administrators can deny a server or individual tool for selected users, require human approval before sensitive operations, and keep autonomous background agents on separate service credentials. Search mode lets an agent load only relevant tools when a kit contains hundreds or thousands of options.

The gateway records the tool, server, identity, outcome, latency and surrounding agent trace. LangChain says request and response payloads are not stored in that tool-call record, reducing the chance that sensitive content enters an audit log. Enterprise plans add an organization-wide audit view.

Built-in connections cover services including Google, Microsoft, Salesforce, GitHub and Linear. The gateway works with MCP clients such as Claude Code and Cursor as well as LangChain products. It reduces authentication plumbing, but administrators still need to design narrow permissions and decide which actions deserve explicit approval.