LangSmith launches Tool Gateway: agents reach tools through a single URL
An agent that reconciles warehouse stock overnight needs to know which tools it may use and whose credentials to use, yet developers still hand-write login flows for every service. LangSmith has released Tool Gateway, a gateway through which all agent tool calls pass. You assemble the tools you need, get a single MCP address (URL), and the gateway injects credentials and logs every call. You no longer have to code OAuth login for each agent and each service.
AI-processed from LangChain Blog; edited by Hamidun News
Imagine an agent that reconciles warehouse stock by itself every night at three in the morning. It needs email, the warehouse system and a schedule, and you need to know which of these tools it is allowed to use, under whose account, and what exactly it did. Today developers handle this by hand: they write OAuth (the standard "sign in with an account" protocol) login for each service and each agent. LangSmith, the LangChain platform for building agents, has released Tool Gateway. It is a gateway: all tool calls pass through one address (URL) over the MCP protocol, and the gateway decides who may do what, injects the right credentials and logs every call.
What problem the gateway solves
While an agent is single and runs locally, things are simple: the developer hardcodes keys into the code, writes OAuth authorization for each service and adds a route for every new tool. That is enough for a prototype. In production the picture is different. The agent serves hundreds or thousands of users, each with their own permissions, tokens must be refreshed, and access must be recorded for audit. Doing this separately for each agent is slow, and every homemade integration becomes another security hole.
LangChain gives a real-world example. A large American retailer is building agents for its stores. The first agent is an employee assistant. It answers questions and works with the permissions of a specific person: warehouse, schedule, email. The second is a warehouse agent that reconciles stock at three in the morning under its own service account. The team has already written more than 50 MCP servers with over a thousand tools in total. Each agent needs only a small approved part. If you give an agent a whole server or hardcode keys into the code, nothing stops it from taking someone else's credentials or an unapproved tool. Meanwhile the security team cannot see what is happening.
How Tool Gateway works
All calls go through one URL, and the gateway handles three tasks.
Access. An administrator assembles a set of tools (a kit) and publishes it at one address. An agent connected to the kit sees only what is included in it. If there are hundreds of tools, search mode kicks in: the agent does not load the full list at once but requests only the tools that fit the task. This reduces latency and lowers the risk of picking the wrong tool.
Identity. The gateway knows on whose behalf to make a call. For the employee assistant, it uses the person's personal access. The user connects their account once, and their identity is confirmed by a corporate provider such as Okta or Entra ID. If the person has not yet signed in, the call returns a login link, waits and continues by itself. For background agents, profiles are used: shared credentials that an administrator connects once.
Policies. Deny rules can be set for a user, a service key or a combination of the two, at the organization or workspace level. There are also approval rules. For example, writes to the data store can be allowed only after a human confirms them in the MCP client.
Logs and ready-made integrations
Every call lands in a LangSmith trace inside the agent's conversation. You can see the tool, the server, the user, the result and the duration. The contents of the data sent to a tool or returned from it are not stored. A question like "which tools did the agent use at three in the morning and under whose credentials" becomes an ordinary search in one place, not a matter of assembling logs from several systems. Tool Gateway works alongside LLM Gateway, which already manages model calls: spending limits, fallback models, and masking personal data.
For common services there are built-in servers hosted by LangSmith, requiring no servers of your own and no OAuth app registration. Google, Microsoft, Salesforce, GitHub, Linear and LinkedIn connect via account sign-in. Exa, Tavily, Apollo, Ashby and Pylon, for example, connect via an API key.
Setup takes five steps: register an MCP server or enable a built-in one, connect an account, select the tools, get a single MCP URL, and point the client at it. Any MCP client works, including Claude Code and Cursor.
What this means in practice
For teams running several agents in production, this is a way to remove repetitive authorization code and get audit without separate development. There are caveats. The solution is tied to the LangSmith ecosystem. Access to deny policies and the audit log, according to the announcement, is offered separately, so the terms should be checked in advance. The announcement contains no independent figures on speed or security gains: it is the vendor's own description of capabilities.
The practical takeaway is simple. If you have more than one agent and more than one service, list the tools for each agent and decide on whose behalf it calls each of them. You will have to maintain this table anyway. The only question is whether it lives in code or in a separate layer with policies and logs.
- Nov 21, 2026Claude Opus 5.5 Picked a School in Iran as a Military Target in 22 of 32 Tries. OpenAI Models Never Did.
- Nov 21, 2026OpenAI Launches Wallet: Dot Pays for Purchases with a Token, Not a Card Number
- Nov 21, 2026ChatGPT cited Ahrefs pages, but in 43% of answers it recommended other companies' conferences
- Nov 21, 2026Gemini in Task Mode Calls Auto Shops, Pays for Purchases and Parses Bills
Key Questions Answered
What problem the gateway solves?
While an agent is single and runs locally, things are simple: the developer hardcodes keys into the code, writes OAuth authorization for each service and adds a route for every new tool. That is enough for a prototype. In production the picture is different. The agent serves hundreds or thousands of users, each with their own permissions, tokens must be refreshed, and access must be recorded for audit. Doing this separately for each agent is slow, and every homemade integration becomes another security hole.
How Tool Gateway works?
All calls go through one URL, and the gateway handles three tasks. Access. An administrator assembles a set of tools (a kit) and publishes it at one address. An agent connected to the kit sees only what is included in it. If there are hundreds of tools, search mode kicks in: the agent does not load the full list at once but requests only the tools that fit the task. This reduces latency and lowers the risk of picking the wrong tool. Identity. The gateway knows on whose behalf to make a call.
What are the key details on Logs and ready-made integrations?
Every call lands in a LangSmith trace inside the agent's conversation. You can see the tool, the server, the user, the result and the duration. The contents of the data sent to a tool or returned from it are not stored. A question like "which tools did the agent use at three in the morning and under whose credentials" becomes an ordinary search in one place, not a matter of assembling logs from several systems.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.