AI NewsModels & agentsAnnouncement

LangSmith Tool Gateway opens a credential layer for agents

LangChain opened LangSmith Tool Gateway in public beta. It handles sign-in, credentials, approvals and audit for every tool an agent calls.

AI News

Editorial2 min read

LinkedInX
Signal field artwork from the Reveneau library, used for an agents story

Why it mattersEach agent your team writes usually hard-codes its own tokens and permissions. A shared layer moves that work out of the agent and into one place you can audit.

Every agent that calls tools on a user's behalf has to answer three questions. Whose credentials does it use? Who stops it when a tool call is dangerous? And where does the record of what it did live? LangChain published a new product that answers all three in one place. LangSmith Tool Gateway, now in public beta, sits between an agent and the tools it calls.

LangChain calls it "a managed tool layer that handles credentials, identity, and access" for every tool an agent calls. An agent makes requests to one URL. LangSmith manages the tokens, the sign-in flow and the audit record behind it.

What it holds

The release includes built-in tools for Google, Microsoft, Salesforce, GitHub, Linear and LinkedIn. It also accepts any MCP server a team already runs, so it is not a closed catalogue. For each tool, it stores either the user's own credentials or a credential tied to the agent itself. When a user has not signed in yet, the gateway sends a sign-in link. It waits for the user to authenticate, LangChain writes, then continues automatically.

The team added two policy hooks. A deny policy stops a tool call outright. An approval policy "ask the user to confirm a sensitive tool call" before it runs. Both apply to any tool behind the gateway, without changing the agent code.

The trace

Every call is recorded. Each trace captures "tool, server, user, outcome, and duration", LangChain says. The record shows up in the LangSmith UI next to the agent that made the call. A failed call carries the error message. A refused call carries the policy that refused it.

That matches what a team in a reviewed industry already has to produce by hand. A SOC 2 auditor asks for the same record: who did what, through which agent, against which account. Tool Gateway writes it as a side effect of the tool call.

What the post leaves unanswered is cost. LangChain gives no pricing for the public beta and says nothing about per-call or per-token rates. For a team already paying for LangSmith, the gateway appears to be another surface inside that subscription. For a team not on LangSmith, the entry price is the platform price.

Running an agent that signs into somebody else's Google account is not a demo step. It is a production system with a secret-rotation problem the agent framework will not solve on its own. A gateway is one of the two honest answers to that problem. The other is to build the same thing yourself.

Source

LangChain, LangSmith Tool Gateway: Secure, governed tool use for agents, by Karthic Subramanian.

SourceLangChain

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project