Vertice AI

The connector is anadditional access channelto Vertice. It exposes the same procurement data and workflows you can already use through the Vertice web application, governed by the same role-based access controls.

Connecting an AI assistant does not grant access to anything you could not see in Vertice yourself.

Throughout this document we use Claude as the worked example. The same connector works with any MCP-compatible AI assistant that supports the MCP Streamable HTTP transport and OAuth 2.1 with PKCE (S256).

The connector is intended for procurement-related questions and workflows. Below are illustrative questions you might ask. Each maps to one of the tools listed in section 4.

- "Which contracts in our portfolio expire in the next 90 days?"
- "What is our total annual spend with this vendor, and which other vendors are in the same category?"
- "Find the latest signed contract for this vendor and summarize the renewal terms."
- "Open a new purchase request for an additional 25 seats of this vendor's product."

These are examples. The exact set of tools, parameters, and behavior is determined by Vertice's orchestration layer at the time of the request.

The connector itself is a hosted service. No software is installed on your infrastructure. No personal API keys are issued.

The connector exposes tools across the following procurement-relevant categories. Each tool advertises explicit MCPreadOnlyHintanddestructiveHintannotations, so compatible AI assistants can warn the user before invoking write actions.

The authoritative list of tools is discoverable directly inside your AI assistant — the assistant calls the standard MCPtools/listmethod against the connector and displays the names, descriptions, and annotations to you.

The exact set of tools may evolve. The annotations on each tool will continue to indicate whether it reads or writes.

Each Vertice tenant has its own connector URL of the form:

https://ai.vertice.one/accounts/<your-account-id>/mcp

Replace<your-account-id>with your tenant identifier. If you do not know your tenant identifier, please contact Vertice Support.

You can now ask Claude procurement-related questions about your Vertice data in any conversation.

For convenience, the same flow can be initiated by following this link:

https://app.vertice.one/preferences/user/integrations

Authentication uses OAuth 2.1 with PKCE (S256). The flow is:
- Your AI assistant fetches the connector's RFC 9728 protected-resource metadata to discover the authorization server.
- The AI assistant fetches RFC 8414 authorization-server metadata to discover the OAuth endpoints.
- The AI assistant initiates an OAuth 2.1 authorization code flow with PKCE.
- Vertice's/authorizeforwarder routes you directly to your tenant's authentication provider. If your tenant uses SSO via SAML, Vertice routes you straight to your configured SAML identity provider with no Vertice picker step in between. If your tenant uses Vertice's native authentication, you see the Vertice login form.
- After successful sign-in, Vertice's identity layer issues a short-lived access token bound to your user, your tenant, and the connector's resource scope.
- The AI assistant presents that access token as a Bearer credential on every MCP request to Vertice.

Tokens expire automatically. You can disconnect the connector at any time from your AI assistant's settings; this does not affect your Vertice account.

The connector enforces your existing Vertice role-based access controls. The data you can reach through the connector is exactly the data you can reach through the Vertice web application, no more.

The same access-control model is shared across all Vertice surfaces — the web application, the Vertice integrations for Slack and Microsoft Teams, and the Vertice AI MCP connector. Every action that modifies data, regardless of which surface initiated it, is captured in Vertice's existing audit trail.

To preserve service quality, the connector enforces application-layer rate limits:

When a limit is reached, the connector returns a JSON-RPC error with code-32004, mapped toHTTP 429 Too Many Requestswith the following headers:

Retry-After: <seconds> X-RateLimit-Limit: <bucket-capacity> X-RateLimit-Remaining: 0 X-RateLimit-Reset: <unix-timestamp-when-window-resets>

Well-behaved AI assistants honorRetry-After. In ordinary interactive use you should not reach these limits. They exist primarily to protect against runaway automation. If your tenant's legitimate workload requires higher thresholds, please contact your Vertice Account Manager.

When a rate limit is reached during a Claude conversation, the Claude.ai client may display the response to the end-user as a generic tool-error message rather than as an explicit rate-limit message. The wire-level response is spec-compliant; the user-facing surface text is determined by the AI assistant, not by Vertice. If you encounter such an error, wait a few seconds before retrying.

The connector targets the 25,000-token response cap defined by the Anthropic Connectors Directory guidance. To keep results usable for very large queries, a safety-net truncation is applied at approximately 20,000 tokens; if it engages, the response includes an explicit footer indicating the truncation took place and how to narrow the query. In normal use, the underlying Vertice orchestration is designed to paginate large list operations and you should not see the truncation footer.

When you ask any external AI assistant a question that calls the connector, your prompt and the connector's response are transmitted to that AI assistant's provider. Their privacy policy and terms govern what they do with that data.

Vertice does not control or audit downstream processing by the AI assistant provider. Please review the AI assistant provider's policies before enabling the connector.

Does the connector send my Vertice data to my AI assistant provider?

When you ask an external AI assistant a question that calls the connector, your prompt and the connector's tool response are transmitted to the AI assistant provider's infrastructure (for example, to Anthropic for the Claude.ai connector). This is the unavoidable nature of an MCP connector. Vertice does not control or audit what the AI assistant provider does with that data — that is governed by your agreement with the AI assistant provider. We strongly encourage you to review the AI assistant provider's privacy policy before enabling the connector.

Does the connector see my Claude chat history or any other data inside Claude?

No. The connector receives only the explicit MCP request envelopes that Claude sends — the tool name and the tool parameters. It does not have access to Claude's memory, conversation summaries, prior conversations, or files uploaded into Claude.

Will the connector make changes to my Vertice data without my approval?

Tools that perform writes carry explicit MCPdestructiveHintandreadOnlyHint=falseannotations, which signal to compatible AI assistants to confirm with the user before invoking them. Claude follows this pattern. If you use an AI assistant that does not honor these annotations, please use the connector only with read tools or disable it.

Can two users on the same tenant share a connector configuration?

Each end-user authenticates individually with their own identity. The connector URL is per-tenant; the connection itself is per-user. Each user only sees what their own Vertice role permits.

Does the connector work offline or in air-gapped environments?

No. The connector is a hosted service. It requires inbound HTTPS connectivity from the AI assistant provider's infrastructure. There is no on-premises or self-hosted option.

Where can I report a bug or request a new tool?