LangChain on November 19, 2026 released LangSmith Tool Gateway in public beta on its LangSmith Agent Platform. The announcement was written by Karthic Subramanian. Tool Gateway is a managed layer that sits between agents and the tools they call, handling credentials, identity and access so that teams do not have to build authentication separately for every agent. LangChain frames it as the tool-side companion to LLM Gateway, the existing LangSmith feature that governs model calls with spend limits, fallbacks, rate limits and redaction of personal data and secrets.
The company's argument is that agents are shifting from chat windows to scheduled and event-triggered work that runs without someone watching each step. That raises three requirements for tool use: access, meaning which tools an agent may call and with which credentials; control, meaning what each user may do and when approval is needed; and visibility, meaning a record of what was called. LangChain describes the common shortcut of hard-coding an OAuth flow per provider and storing tokens next to the agent, and says it breaks as the number of tools, users and agents grows.
In practice, a team registers an MCP server by URL or switches on a built-in server, then bundles selected tools into a kit that is published at a single MCP URL. An agent connected to a kit sees only the tools in it. For large libraries there is a search mode, in which the agent queries the kit and loads only the tools relevant to the task. LangChain says the gateway works with any MCP client and names LangChain, Deep Agents, Managed Deep Agents, Claude Code and Cursor. Built-in servers hosted by LangSmith cover Google, Microsoft, Salesforce, GitHub and Linear, with others such as Exa, Tavily and Apollo available through an API key.
Each call resolves to one of two identities. When acting as the user, the gateway applies that person's own credentials, identified through an identity provider such as Okta or Entra ID. If the person has not connected an account yet, the call returns a sign-in link, waits and then continues. When acting as the agent, it uses credentials tied to the agent, set up through agent auth or through a profile of shared credentials that a workspace admin connects once.
Controls come as deny policies, which can block tools by user, service key or combination at organisation or workspace level, and approval policies, which ask a person to confirm a sensitive call such as writing to a data warehouse before it runs. Every call is traced inside the agent's own LangSmith trace with tool, server, user, outcome and duration, and LangChain says the data sent to or returned by the tool is not stored. On Enterprise plans an audit log records which tool ran and as whom. LangChain says customers must contact it for access to deny policies and the audit log.
The post uses an unnamed large US retailer as its example, with more than 50 MCP servers and over a thousand tools, and offers no measurements of speed, adoption or security outcomes. Its claims about faster setup and simpler audits are the vendor's own, and the beta label means behaviour and pricing may still change.