LangSmith Tool Gateway: one gate for every AI agent tool call
LangSmith Tool Gateway is in public beta: agents call MCP tools through one URL, never hold API keys, and hit deny and approval policies. What to set up first.
Your AI agent should not hold your Salesforce token. On November 19, LangChain put LangSmith Tool Gateway into public beta. It is a managed layer between agents and their MCP tools. Every tool call goes through one URL. The gateway checks who is calling, picks the right credential, applies your policies, and logs the call. The agent never sees the key.
What actually happened
Per LangChain's launch post and docs:
- Two identities. A call acts as the user (their own sign-in through Okta or Entra ID) or as the agent (shared credentials an admin connects once). One identity per call.
- Tool kits. A curated set of tools at one URL. An agent connected to a kit sees only those tools. The docs are clear that kits are for curation, not access control. Policies do the blocking.
- Deny policies block tools per user, per service key, or for everyone with exceptions. Deny wins when policies conflict. Tools are allowed by default.
- Approval policies make the person in the MCP client confirm a sensitive call first. A separate reviewer is not supported yet.
- Visibility. Traces log tool, server, identity, outcome, and duration, but not the data sent or returned. An org-wide audit log is on Enterprise plans.
- Built-in servers include Gmail, Google Drive, Outlook, SharePoint, Slack, Salesforce, GitHub, and Linear (full list). Your own servers must be remote, Streamable HTTP over HTTPS.
- Limits: 1 MB requests, 8 MB responses, 100 requests a second.
- Clients: any MCP client, including Claude Code and Cursor.
LangChain published no price for the gateway.
Why it matters for your business
Most small teams wire agents to tools the fast way: paste an API key into a config file and hope. That key usually has full access, lives on a laptop, and leaves no record of what the agent did with it. A gateway fixes all three at once.
One line in the docs matters more than the feature list: the gateway passes each server's response through unchanged, so "treat tool output as untrusted input." The gate controls what your agent can do. It does not clean what your agent reads. A poisoned email or web page can still steer it.
If you test it:
- Start with deny. Allow is the default, so block write and delete tools first, then open what you need.
- Put approvals on anything that sends, pays, or deletes.
- Move keys out of config files and into the gateway, then rotate the old ones.
Key takeaways
- LangSmith Tool Gateway entered public beta on November 19, 2026
- Agents call MCP tools through one URL and never hold the credentials
- Deny policies win over allows; tools are allowed by default
- Approval policies ask the user in the MCP client to confirm first
- Tool output passes through unchanged, so still treat it as untrusted
Agents holding keys they shouldn't? We audit your agent-to-tool wiring, move credentials behind a gateway, and set deny and approval rules that match how your business actually works. See what we build or book a review.
Sources: LangChain blog, LangSmith docs: how Tool Gateway works, LangSmith docs: servers.
- #langsmith
- #mcp
- #ai-agents
- #agent-security
- #langchain
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenRouter drops 3 providers, pulls ZDR from 125 endpoints
OpenRouter off-boarded three inference providers and removed 125 endpoints from zero data retention after Dragos canary research. What to check in your stack.
Read itOpenAI Performance Factory: Codex agents debug, engineers merge
OpenAI's Performance Factory ran 1,700+ Codex investigations into ChatGPT slowdowns. 70% of audited root causes held up; about 15 agent PRs shipped as written.
Read it