LangChain launches a gateway to control which tools AI agents can use
LangSmith Tool Gateway separates an agent’s available tools from whose credentials it uses. Organization-wide audit logs require an Enterprise plan.
Loading page…
LangSmith Tool Gateway separates an agent’s available tools from whose credentials it uses. Organization-wide audit logs require an Enterprise plan.
Listen to this story
LangSmith Tool Gateway organizes tool access into curated kits, so agents can be limited to approved tools while calls use either a person’s credentials or credentials configured for autonomous agents. It also supports blocking calls and requiring human confirmation for sensitive actions. LangChain says traces capture who called which tool, the outcome and duration, but not the tool’s inputs or outputs; deny policies and organization-wide audit logs are available on Enterprise plans.
LangChain opened the service’s public beta on November 19.
Teams can connect an MCP server by URL or use hosted integrations for Google, Microsoft, Salesforce, GitHub and Linear without operating a server or registering their own OAuth app.
Search mode lets agents retrieve relevant tools from large kits instead of loading hundreds or thousands at once.
LangChain launched LangSmith Tool Gateway in public beta on November 19, giving teams a managed checkpoint between AI agents and the tools they call. The service routes requests through a single URL and handles access, identity and logging—putting controls around both what an agent can do and whose credentials it uses.
The launch targets a problem LangChain illustrates with an unnamed large U.S. retailer. Its engineering team has built more than 50 MCP servers exposing over a thousand tools. Yet its store assistant needs individual associates’ access, while an overnight inventory agent needs service credentials. Each should reach only an approved subset of tools.
Tool Gateway organizes that access into tool kits: selected collections published at one URL. An agent connected to a kit sees only its bundled tools. Kits can be shared across a workspace or kept for personal use, rather than exposing every tool a server offers.
Teams can register a server by URL using MCP, the tool-connection interface supported by the gateway, or choose a LangSmith-hosted server. Built-in options include Google, Microsoft, Salesforce, GitHub and Linear. For those built-ins, teams do not have to operate a server or register their own OAuth application for account authorization.
For larger kits, search mode lets the agent query its collection and load tools relevant to the current task. LangChain says loading hundreds or thousands of tools at once increases delay and makes tool selection harder.
Tool availability and identity are separate settings. For each call, the gateway can use the individual user’s credentials or credentials assigned to the agent. The latter can come from agent authentication configured in LangSmith or a shared profile connected by a workspace administrator.
In the user-based path, an identity provider such as Okta or Entra ID identifies the caller. If that person has not connected the required account, the tool call returns a sign-in link, waits for authentication and then continues automatically. LangChain says this removes the need to write additional code for that authorization handoff.
Administrators can set deny policies at the organization or workspace level, targeting users, service keys or combinations of the two. They can also block access for everyone with named exceptions. Approval policies serve another purpose: asking a person to confirm a sensitive call in their MCP client before it runs.
LangChain’s example is a tool that writes to a data warehouse. An administrator can require human approval each time the agent attempts that write, rather than treating permission to access the tool as permission to execute every request.
For investigation afterward, gateway tracing records the tool, server, user, outcome and duration within the agent’s activity record. LangChain says data sent to or returned by the tool is not stored in that trace. The record therefore captures call details without retaining the tool’s inputs and outputs.
Organization-wide audit logs, which record which tool ran and whose identity it used, are available on Enterprise plans. The launch directs customers to contact LangChain for access to deny policies and the audit log.
The client side is not limited to LangChain’s own agents. Tool Gateway supports any MCP client, the company says, including Claude Code and Cursor: teams point their client at the kit URL to use the configured tools.
Loading discussion...
Join the conversation
Explain where you would draw the line between privacy and accountability.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.