Platform · Tools
Reach your data from an agent. Provision infrastructure from your terminal.
Two different jobs: the MCP server connects an agent to your data, the CLI provisions your infrastructure. Neither one can reach further than the API itself allows.
The MCP server
Agent-native access, without hand-building the integration.
Hand an agent a bare API key with no tool boundary, and it can call anything that key can reach. Wiring up a scoped tool layer yourself is real work, and easy to skip under time pressure. This ships with that boundary already built in, covering the whole data plane.
The whole data plane, natively callable
Search, records, documents, folders, and RAG: twenty-three tools, each wrapping a real data-plane operation. No web-search or third-party tool is present to call, by design, not by configuration you could loosen.
The same enforcement as everywhere else
Tenant isolation and scope enforcement hold exactly as they do on the API. Running an agent through this server doesn't widen what its credential can reach.
{
"mcpServers": {
"vectros": {
"command": "npx",
"args": ["-y", "@vectros-ai/mcp-server"],
"env": { "VECTROS_API_KEY": "ssk_live_..." }
}
}
}Connects Claude Desktop, Cursor, Cline, Continue, VS Code, and hosted agent platforms. Tool results are sized for a context window by default, smaller limits than the HTTP API gets, to protect both the window and the cost.
The CLI
No root key in the loop. No surprises when you run it.
Hand-rolled provisioning usually means your root key ends up somewhere scriptable, and the first sign something's wrong is a failed run against real infrastructure.
Bootstrap without the root key
A short-lived bridge token from the developer portal is enough to mint a least-privilege scoped key. Your root key never touches the flow, whether you run it interactively or from a script.
Validate before you provision
The blueprint commands are credential-free and offline: structural checks, the scope gate, and a linter all run before anything gets created, and plan previews the result without minting a thing.
# Provision a least-privilege credential without ever touching your root key:
vectros bootstrap
# Interactively asks what to set up, mints a narrow scoped key (ssk_*), and merges an MCP
# config entry pointing at it. Non-interactive: --blueprint task-management --yes.
# Author and validate a blueprint fully offline, no credentials needed:
vectros blueprint init my-app --from clinical-intake
vectros blueprint validate ./my-app.blueprint.yaml
vectros blueprint plan ./my-app.blueprint.yamlSee it in a real app
One command sets up both.
Fork the RAVV reference app and point bootstrap at its own blueprint file directly, no separate scaffold step: the same command that mints a scoped credential also merges the MCP config for it, so provisioning and agent access come up together, not as two setup steps.
Read: the RAVV stack →Both surfaces are ways to reach the same governed system, not a separate one.
See how it all fits together on the platform page.