LangSmith Tool Gateway Controls What AI Agents Can Do
LangChain launched LangSmith Tool Gateway in public beta on 19 November 2026, a managed access layer that centralises credentials, enforces identity checks, and creates an audit trail for every tool an AI agent calls. The gateway supports MCP-compatible clients and includes built-in integrations for Google, Microsoft, Salesforce, GitHub, and Linear. It is available on LangSmith Plus and Enterprise plans.
Operator Insight
Every AI agent you deploy makes decisions in your name. Until this week, the only way to control what tools those agents could reach was to manage credentials manually, audit logs inconsistently, and hope no agent exceeded its intended scope. LangSmith Tool Gateway changes that by making governance a first-class feature, not a configuration afterthought. For a 10 to 200 person organisation deploying its first real agents, this is the missing infrastructure layer that makes it safe to give agents access to Salesforce, GitHub, and Google Workspace without also giving them free rein over your entire stack.
30-Second Summary
LangChain launched LangSmith Tool Gateway in public beta on 19 November 2026. It is a managed access layer that sits between AI agents and the tools they call. Every tool call is routed through the gateway, where LangChain enforces credentials, identity, access policies, and audit logging centrally rather than leaving each of those to individual developers to implement.
At a Glance
- Topic: Agent Systems
- Company: LangChain
- Date: 19 November 2026
- Announcement: LangSmith Tool Gateway enters public beta, providing centralised governance for AI agent tool calls
- What Changed: Agent tools can now be governed by a managed access layer with identity verification, policy enforcement, and a full audit trail
- Why It Matters: Enterprises deploying agents across business systems have had no standard way to control what those agents can access. This creates one.
- Who Should Care: Any organisation deploying AI agents that call external tools, APIs, or business applications
Key Facts
- Launch date: 19 November 2026 (public beta)
- Developer: LangChain
- Availability: LangSmith Plus and Enterprise plans
- Built-in integrations: Google Workspace, Microsoft 365, Salesforce, GitHub, Linear
- Protocol support: MCP-compatible clients
- Request limits: 1 MB request cap, 8 MB response cap, 100 requests per second rate limit
- Audit logging: Records tool name, caller identity, and duration. Tool inputs and outputs are not retained by LangChain.
- Policy support: Access controls, deny policies, and approval flows for tool calls
What Happened
LangChain released LangSmith Tool Gateway in public beta on 19 November 2026. The product sits between an AI agent and every tool it calls, handling credentials, identity, access policies, and audit logs from a central point rather than leaving each of those concerns to individual developers.
The gateway works with MCP-compatible clients, the protocol that has emerged as the standard interface for agent tool calls. It presents a single managed URL that serves all configured tool servers, so an agent calls one endpoint and Tool Gateway routes the request to the appropriate tool with the correct credentials.
LangChain manages credentials inside LangSmith, meaning agents never receive the raw API keys for the tools they call. When an agent calls a Salesforce tool, Tool Gateway authenticates the request against the caller's identity, checks whether that call is permitted by the access policy, and logs the outcome before the tool ever sees the request.
Built-in servers for Google Workspace, Microsoft 365, Salesforce, GitHub, and Linear are available from day one. For other tools, developers can add MCP servers that are publicly accessible or provide their own API keys. Third-party integrations such as Exa and Tavily require the developer to supply an API key or use a publicly accessible endpoint.
Why It Matters
Agent governance has been the missing enterprise blocker. Most organisations that have deployed AI agents have done so by embedding credentials directly into prompts or environment variables. This works in development but fails at scale: credentials are scattered, audit trails are nonexistent, and there is no way to revoke access to one tool without touching every agent configuration.
The MCP protocol created a problem Tool Gateway solves. As Model Context Protocol became the standard interface for agent tools, the number of tools any single agent could theoretically reach grew rapidly. Without a governance layer, every new MCP server a developer connects is another surface where an agent can act with unchecked scope.
Identity at the tool call level changes the accountability model. Tool Gateway routes calls through a named identity, not an anonymous API key. That means an audit log that shows which agent called which tool, on behalf of which user, in which workflow. For regulated industries, this is a minimum requirement for deploying agents in production.
Deny policies shift AI risk management from intent to enforcement. Previously, an operator who wanted to stop an agent from sending emails or deleting records had to rely on prompt instructions. Tool Gateway lets you express that intent as a machine-enforced policy that the agent cannot override through any form of jailbreak or prompt injection.
The LangSmith Plus pricing tier puts this in reach for small teams. Enterprise governance tooling in AI has historically required a sales conversation and an enterprise contract. The public beta on a Plus plan means a 15-person company can wire up governed agent tool calls without a procurement process.
This normalises a standard that will compound. Tool Gateway works with any MCP-compatible client. If the pattern succeeds, the expectation that agent tool access requires a governance layer, not just an API key, will become the floor for responsible AI deployment across the industry.
The David and Goliath View
The story of enterprise AI adoption in 2026 has been a story of tools arriving before the infrastructure to govern them. Agents can now write to your CRM, send emails in your name, query your databases, and create records in your project management system. What most organisations have not had is a way to say, with confidence, exactly what any agent can and cannot do, and to prove it after the fact if something goes wrong.
LangSmith Tool Gateway is the first product that treats this as a solvable infrastructure problem rather than a prompt engineering challenge. That matters. Telling an agent not to delete records is not governance. A policy that the system enforces, with an audit log that shows the policy held, is governance. For a business operator considering their first serious agent deployment, this is the difference between a pilot and a production system.
David and Goliath has been building Employee Amplification Systems for clients who want agents doing real work inside real business systems. The question every client asks is: how do I know the agent stays in its lane? Tool Gateway is the clearest answer yet. Build your agents, wire up their tools through a governed gateway, and have an audit trail that tells you, every day, whether the system is doing what you intended.
Where This Fits in the AI Stack
Tool Gateway occupies a new layer between the agent orchestration layer and the tool execution layer. Before this product existed, that gap was filled by documentation, convention, and hope. Now it is a managed service with machine-enforced policies.
It connects upward to any agent framework that speaks MCP, including LangChain, OpenAI Agents SDK, and Anthropic Claude with MCP support. It connects downward to any tool server that is reachable over HTTPS, with built-in managed connections to five common enterprise systems.
The natural comparison is to identity and access management platforms such as Okta or Microsoft Entra, applied to AI agents rather than human users. The key difference is that agent identities need to be tied to workflows and task contexts, not just user accounts, and Tool Gateway's model reflects that.
Questions Operators Are Asking
Is this only useful if I am already using LangChain? No. Tool Gateway works with any MCP-compatible client, including agents built with OpenAI's Agents SDK, Anthropic Claude, or other frameworks. You do not need to use LangChain's orchestration layer to use Tool Gateway for access control.
What happens if LangSmith is unavailable? The docs label the feature beta, and LangChain has not published SLA guarantees for the gateway in the beta period. This is a risk to assess before making Tool Gateway the access layer for a production agent. Test it in a staging environment before connecting it to live business systems.
Does Tool Gateway log what agents actually read or write? No. Per LangChain's published documentation, the audit trail records which tool was called, by whom, and how long it took. Inputs and outputs are not retained. This is a deliberate privacy design choice, but it means the log tells you a call happened, not what the agent actually did with the response.
Can I use the five built-in integrations without a separate contract? Yes. The built-in servers for Google, Microsoft, Salesforce, GitHub, and Linear are managed by LangSmith, so you connect them through the gateway without negotiating a separate integration contract. The standard OAuth authorisation flow for each provider applies.
When will deny policies and audit log self-service be fully available? As of the public beta on 19 November 2026, deny policies and advanced audit log access require engaging LangChain's team directly. LangChain has not published a roadmap date for full self-service access.
Citable Summary
LangChain launched LangSmith Tool Gateway in public beta on 19 November 2026. It is a managed access layer for AI agent tool calls that enforces credentials, identity, access policies, and audit logging centrally, rather than leaving each to individual developers. The gateway supports MCP-compatible clients and includes built-in managed integrations for Google Workspace, Microsoft 365, Salesforce, GitHub, and Linear. Access policies including deny rules, identity checks, and audit trails can be configured without embedding credentials in agent prompts or environment variables. The beta is available on LangSmith Plus and Enterprise plans.
Why This Matters for Operators
- ✓
Register your business-critical tools in Tool Gateway before connecting them to any agent. This creates the access layer you can audit and revoke, rather than credentials embedded in prompts or environment variables.
- ✓
Use deny policies for destructive or irreversible actions such as delete, send, and payment operations from day one. The cost of an agent mistake in production is higher than the time it takes to set a policy.
- ✓
Take advantage of the built-in MCP servers for Google, Microsoft, Salesforce, GitHub, and Linear. These are tested integrations with known permission scopes, faster to deploy safely than rolling your own.
- ✓
Review the audit trail weekly during your first 60 days of agent deployment. Tool Gateway logs which tool was called, by whom, and how long it took. Anomalies in that log are your earliest warning that an agent is behaving outside its intended scope.
Related Intelligence
Related Briefings
- Cloudflare Proposes Identity Standard for AI AgentsCloudflare | Agent Systems
- Google Releases Gemini 4 Argon With Background Agent Tasks and 40+ App IntegrationsGoogle | Agent Systems
- DigitalOcean Launches Managed Agents, Removing the Infrastructure Barrier to Running AI at WorkDigitalOcean | Agent Systems
- AWS Closes Bedrock Agents to New Customers. AgentCore Is the Replacement.Amazon Web Services | Agent Systems
Related Signals
- [High] Anthropic launches Claude Agent SDK
Standardised framework for deploying production AI agents with built-in tool orchestration and safety guardrails.
Related Comparisons
- AI Growth Agency vs In-House Team for Cybersecurity Vendors
How hiring an AI growth agency compares to building an in-house growth team for a cybersecurity vendor, across speed to pipeline, cost, security buyer fluency, and key person risk.
- David & Goliath vs Deloitte AI
How a boutique AI systems firm compares to a global consulting practice for AI implementation, speed to deployment, and ongoing support.
- David & Goliath vs PwC AI
How David & Goliath compares to PwC for AI strategy, implementation speed, and cost structure for mid market organisations.
Explore Related Intelligence
How This Maps to David & Goliath
Apply This to Your Business
Want to see what this means for your team?
Tell us a little about your business and we will map the specific opportunity for your sector and team size.