Virtual MCP Servers

Build a tool set for each agent. From any app, behind one endpoint.

Pick the tools one agent role needs, from any connector or your own API, and serve them as one endpoint.

One endpoint.
Assembled from anything in your catalog.

A deal desk agent needs an opportunity, an email thread, a Slack note and an invoice. Give it those five tools, not all 112.

Every run, the model chooses from five.
Not from twenty thousand.

Give an agent only the tools its job needs, and it picks the right one the same way, every run.
Clear context
More room for the task
Around 400 tokens of tool definitions instead of 8,000. The rest of the window goes to the work.
Just the right access
Access you set, not the prompt
A prompt asking the model not to delete can be ignored. A tool that isn't in the session can't be called.
Faster, more accurate
Fewer candidates, fewer wrong turns
Five candidates instead of forty means fewer wrong calls, fewer retries, and runs that finish sooner.

Set it up once.
Start a new session each run.

Define the tool set once per agent role. Each run then gets a session for one user, and your agent always calls the same URL.
Once per role
Define the config
Name the connections and the tools you want from each. Omit the tool list on a connection to expose all of them.
actions.mcp.create_config()
Per user, once
Check their accounts
See which of the required connections this user has authorized, and get an auth link for the ones they haven't.
list_mcp_connected_accounts()
Per run
Create a session token
Scoped to that user's connected accounts and to this config. Fresh token before every run — never reuse one.
create_session_token()
Per run
Point the agent at it
The static server URL, with the session token as a bearer credential. Your agent code doesn't change between users.
Authorization: Bearer …
Token lifetime
Defaults to about an hour, set it yourself with expiry. Give it more room than the run needs.
No refresh endpoint
Tokens can't be extended in place. Mint a new one rather than holding a long-lived credential.
Stable identifier
An email, a UUID, whatever you use for a user — just keep it consistent across every call for that person.

Where a smaller tool set
matters most.

Every agent works better with fewer tools. These four kinds of agents feel it most.
Agents that run on their own
No one reviews each step, so the agent can only use the tools you gave it. Anything else isn't there to call.
One agent, thousands of users
Set up the tool set once. Each user gets their own session with their own accounts, with no per-user setup.
Agents users are waiting on
Fewer tools means a smaller prompt and fewer wrong picks, so answers come back faster.
Apps serving many customers
Every customer runs the same agent code, and each one only reaches their own data. The gateway enforces it.

Build your first tool set.
It takes one call.

Free to start, no credit card.
Pick the tools, mint a session, and point your agent at the endpoint.