What LangChain shipped
On 19 November 2026 LangChain released LangSmith Tool Gateway as a public beta. It's a managed layer that sits between an agent and the tools the agent calls. The gateway holds the credentials, works out who is making each call, applies access rules and writes a trace. Agents, coding assistants and backend services all reach their tools through one URL, so each agent no longer needs its own copy of that plumbing.
The gateway speaks MCP (Model Context Protocol), the open standard for connecting AI clients to external tools and data. Any MCP client can use it. LangChain names its own frameworks, Claude Code and Cursor as supported clients. The product docs say the beta runs on LangSmith Cloud in the GCP US, GCP EU, GCP APAC and AWS US regions, but not Azure US, and requires a Plus or Enterprise plan.
How a call moves through the gateway
The docs set out a fixed order for every tool call. The gateway identifies the caller as a user, an agent or a shared profile, then blocks the call if a deny policy matches. Next it attaches the caller's stored credential, or asks the caller to connect an account if no credential exists. After that it asks for approval if an approval policy matches. Only then does it forward the call to the MCP server and record a trace.
Tools are grouped into "kits". A kit is a chosen set of tools, possibly drawn from several MCP servers, published at a single URL. You can register any remote MCP server by URL or switch on one of LangSmith's built-in servers. The launch post lists OAuth-backed built-ins for Google, Microsoft, Salesforce, GitHub, Linear and LinkedIn, plus API-key or public servers including Exa, Tavily and Parallel search. LangSmith handles the OAuth app registration for the built-ins. If a server sits inside your own network, the docs say you'll need a tunnel to reach it.
Identity works in two modes. In "acting as user" mode, Okta or Entra ID identifies the person, and the person connects each account once. If they haven't connected yet, the call returns a sign-in link, waits for them to authenticate and then carries on. In "acting as agent" mode, a background job uses the agent's own credentials or a profile that a workspace admin set up. Unite.AI's coverage repeats LangChain's example of a large US retailer with more than 50 MCP servers and over 1,000 tools. In that example a store assistant runs under each staff member's login, and a 3 a.m. stock reconciliation runs under service credentials.
Search mode for large tool catalogues
Loading hundreds of tool definitions into a prompt wastes context and confuses models. The client docs describe a search mode that you turn on by adding ?mode=search to the kit URL. In search mode the kit exposes three meta-tools in place of the full catalogue. search_tools returns matches (up to 50, default 10), get_tool returns one tool's input schema, and call_tool runs the chosen tool. Policies still apply to whatever call_tool runs. Search mode works on saved kits and on the personal "me" kit, but not on a single server's endpoint.
Connecting a coding agent
A client needs Streamable HTTP transport and support for custom headers. Authentication uses an X-Api-Key header carrying a LangSmith key, and organisation-scoped keys also need X-Tenant-Id. For Claude Code, the docs give this command:
claude mcp add --transport http support-agent \
"$LANGSMITH_ENDPOINT/v1/managed-tools/kits/<kit-id>/mcp" \
--header "X-Api-Key: $LANGSMITH_API_KEY"Cursor takes the same URL and header in mcp.json. Managed Deep Agents needs version 0.10.5 or later of managed-deepagents, according to the same page.
Policies and approvals in detail
The policy docs define two effects. FORBID stops the call and returns a 403 error. REQUIRE_APPROVAL lets the call run only after the caller approves it in their MCP client. A block always wins over an approval. A policy can target users, groups, service keys or agents, and can cover one tool, every tool on a server, or *. Each policy holds up to 100 callers and 100 tools, and an organisation can hold 1,000 organisation-wide policies plus 100 per workspace.
Approvals depend on MCP elicitation, the part of the protocol that lets a server pause a request and ask the user for input. The MCP specification defines a form mode for structured input. Tool Gateway uses form mode to show the tool name, its server, the matching policies and the call's arguments. A client that can't display that form gets a 403. LangChain's docs don't list which clients support the form, so test your client before relying on approvals.
Limits worth knowing before you adopt it
Several constraints come straight from the docs. The approval goes to the person making the call, with no route to a separate reviewer or admin, so an approval policy checks intent and gives you no second pair of eyes. An unattended agent can't approve anything, which means a deny policy is the only control for background jobs. Kits decide which tools an agent can see, but the policy docs state plainly that kits don't restrict access and that only policies enforce it. A blocked tool still shows up in the kit's tool list, and Claude Code lists tools from servers you haven't connected yet.
Traces record the tool, server, user, outcome and duration. The launch post says the payloads themselves are left out: "Sensitive data stays out of the record, since data sent to the tool or by the tool is not stored." That's good for privacy, but you can't replay a call's inputs from the trace. The organisation-wide audit log is Enterprise-only, and the launch post asks customers to contact sales for deny policies and the audit log.
On cost, LangChain hasn't published a separate price for Tool Gateway. The LangSmith pricing page lists Plus at US$39 per seat per month with 10,000 base traces, then pay-as-you-go. Enterprise pricing is custom. The gateway is also a hosted dependency: every tool call now passes through LangSmith's cloud, and the docs describe no self-hosted option for the gateway.
What it means for builders in New Zealand
For NZ teams, the APAC region matters most. A GCP APAC deployment keeps tool-call traffic and traces closer to home than the US default, though the docs don't name the exact data centre. Teams that need to say where their data is processed should confirm the location with LangChain before putting customer data through the gateway.
The gateway fits best where a small team runs several agents against the same SaaS systems, such as Google Workspace, Microsoft 365, GitHub or Salesforce, and is tired of managing OAuth apps and refresh tokens for each agent. A sensible first trial is to put your coding agent's GitHub and Linear access behind a kit, add a REQUIRE_APPROVAL policy on write tools, and check how your client handles the approval form. Teams already using LangSmith for tracing get the most from it, because tool calls appear in the same traces as model calls. Teams that want to keep everything in their own infrastructure should note the cloud-only beta and the missing pricing, and wait for both to settle. If you build something with it, you can share it at asai.nz/projects.
Sources
- LangSmith Tool Gateway: Secure, governed tool use for agents (LangChain blog)
- Tool Gateway overview (LangChain docs)
- Tool Gateway clients and search mode (LangChain docs)
- Tool Gateway policies (LangChain docs)
- LangSmith pricing
- MCP specification: Elicitation
- LangChain Launches LangSmith Tool Gateway to Govern Agent Tool Calls (Unite.AI)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.