Connect AI assistants

Start here. This is the first step for using Configview with an LLM. Once you’ve connected an assistant, head to Saved Queries to learn what your assistant can run — and how to add your own.

What it is

Configview exposes an MCP (Model Context Protocol) server so AI assistants — Claude, ChatGPT, Cursor, or any MCP-aware client — can answer questions about your infrastructure by running your saved queries. Once connected, you can ask the assistant things like “Which AWS users haven’t logged in for 90 days?” or “Show me the Okta groups missing MFA” and it will pick the right saved query, run it against your live Configview data, and summarize the result. MCP is an open standard. The same endpoint works with any client that supports it — these instructions cover the most common ones.

Your MCP endpoint

Every Configview customer gets a dedicated MCP endpoint at:

https://<your-subdomain>.configview.com/mcp

Replace <your-subdomain> with whatever your team uses to log into the Configview dashboard. For example, if your dashboard is acme.configview.com, your MCP endpoint is https://acme.configview.com/mcp. You’ll paste this URL into your MCP client in the next section.

Available tools

The server exposes two read-only tools. The assistant decides which to call based on your question.

Tool	Purpose
`list_queries`	Lists every saved query on your Configview instance — title, description, parameter schema, category, and cache TTL. The assistant calls this first to see what’s available. Optional `category` filter (e.g. “Identity”, “Cloud”, “Access”).
`run_saved_query`	Executes a saved query by its `seed_key` and returns the result rows. Honors the result cache by default; pass `force_live=true` for a fresh run.

Only saved queries are exposed. The assistant cannot run arbitrary SQL, modify data, change settings, or access your integrations directly. See the Saved Queries guide for how queries get into the catalog — today the MCP connector exposes the curated catalog Configview ships; user-saved queries appear on the dashboard and Slack bot but are not yet MCP-callable.

Connect your MCP client

Every option below uses the same OAuth flow — your client redirects to Configview, you sign in with your normal credentials, and the client gets a per-user access token. You don’t need to copy API keys around.

Claude.ai (web)

Go to https://claude.ai and sign in
Click your profile avatar > Settings > Connectors
Click Add custom connector
Name: Configview
URL: https://<your-subdomain>.configview.com/mcp
Click Add — a Configview login window opens
Sign in with your Configview credentials and click Allow on the consent page

The connector now shows as connected. In any chat, toggle it on from the message composer.

Claude Desktop

Open claude_desktop_config.json (find it under Settings > Developer > Edit Config) and add the server:

{
  "mcpServers": {
    "configview": {
      "type": "http",
      "url": "https://<your-subdomain>.configview.com/mcp"
    }
  }
}

Restart Claude Desktop. The first time you use the connector it opens a browser window for OAuth sign-in.

Claude Code (CLI)

1. Register the server. Run once to add Configview at user scope (available in every Claude Code session):

claude mcp add --transport http --scope user configview https://<your-subdomain>.configview.com/mcp

--transport http is required — without it, Claude Code defaults to stdio and the connection will fail. If you’ve added the server before without that flag, remove and re-add it:

claude mcp remove --scope user configview
claude mcp add --transport http --scope user configview https://<your-subdomain>.configview.com/mcp

Alternatively, commit a project-scoped .mcp.json to share with your team:

{
  "mcpServers": {
    "configview": {
      "type": "http",
      "url": "https://<your-subdomain>.configview.com/mcp"
    }
  }
}

2. Restart Claude Code. Claude Code only loads MCP servers at startup — a freshly added server won’t appear in the current session. Exit and relaunch. 3. Authenticate. In the new session, run /mcp. The Configview entry shows as needs auth. Select it and choose Authenticate:

A browser tab opens to https://<your-subdomain>.configview.com
Sign in with your normal Configview credentials
Review the consent page (it asks for the mcp scope — read your saved queries and run them on your behalf) and click Allow
The tab closes and /mcp now shows the server as connected

Tokens are stored in Claude Code’s per-user credential store, not in .mcp.json — so committing the JSON to a repo never leaks credentials. Each teammate authenticates as themselves. 4. Use it. Ask Claude something Configview-related, e.g. “List the saved queries on Configview”. The list_queries tool fires and returns the catalog. If you ever need to re-authenticate (signed in as the wrong user, token revoked, etc.), run /mcp → pick configview → Authenticate again. Tokens auto-refresh in the background otherwise.

Other MCP clients (Cursor, Continue, custom integrations)

Any client that supports Streamable HTTP transport with OAuth 2.1 works. The standard config block is:

{
  "mcpServers": {
    "configview": {
      "type": "http",
      "url": "https://<your-subdomain>.configview.com/mcp"
    }
  }
}

The client will discover the OAuth authorization server automatically from the WWW-Authenticate header on the first request (RFC 9728 protected-resource metadata). No client ID, secret, or auth URL needs to be configured manually — dynamic client registration handles it. If your client only supports stdio MCP servers (not HTTP), you can bridge with mcp-remote:

{
  "mcpServers": {
    "configview": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://<your-subdomain>.configview.com/mcp"]
    }
  }
}

Running queries

Once connected, you don’t need to memorize query names. Ask in plain language:

“What queries are available on Configview?” — lists the catalog
“Show me queries in the Cloud category” — filtered list
“Run the Identity drift report” — picks and runs the matching query
“Which AWS IAM users were created in the last 30 days?” — picks a query and fills in the parameters from your question
“Re-run that query with fresh data” — bypasses the cache (force_live=true)

The assistant calls list_queries to discover what’s available, then run_saved_query to execute it. You’ll see the tool call in the chat, and the result rows are summarized inline. If a query expects parameters (e.g. an email address, a date range), the assistant will either pull them from your question or ask you for them.

Security model

The connector uses OAuth 2.1 with PKCE and Dynamic Client Registration — the same auth standard used by GitHub, Atlassian, and other modern SaaS connectors.

Property	How Configview handles it
Per-user tokens	Every user gets their own access token. There are no shared API keys to manage or rotate.
Scoped access	Tokens grant the `mcp` scope only — read your saved queries and execute them. They cannot change settings, manage users, or view secrets.
Time-limited	Access tokens expire after 1 hour. Refresh tokens rotate on every use and are revoked on suspicious activity.
Revocable	Disconnect the connector in your MCP client’s settings at any time. Tokens stop working immediately.
Auditable	Every MCP tool call is recorded in your Configview activity log, attributed to the authenticated user.
No data leaves your tenant unprompted	The assistant only sees query results when you explicitly ask a question. Configview does not push data anywhere.

Result rows returned through the MCP server pass through the LLM provider (Anthropic, OpenAI, etc.) for inference, the same as any other content you paste into a conversation. If your saved queries return sensitive data, treat the connector with the same care you would treat any other AI chat.

Troubleshooting

Claude Code: server shows as failed in /mcp — Almost always one of two things: (1) the server was added without --transport http, so Claude Code is trying stdio. Run claude mcp remove --scope user configview and re-add with the flag. (2) You haven’t restarted Claude Code since adding it — MCP servers only load on startup. Claude Code: “MCP server configview already exists in user config” — You already added it once. Either authenticate the existing entry (/mcp → select configview → Authenticate) or remove and re-add: claude mcp remove --scope user configview. Claude Code: the assistant says it doesn’t know what Configview is — The server is registered but not loaded in the current session. Restart Claude Code and try again. “Couldn’t reach the MCP server” — Your Configview instance must be reachable from the public internet on port 443. Check that your firewall allows inbound HTTPS from the LLM provider, not just from your office IP. If you’re self-hosting behind a VPN, the cloud-hosted LLM cannot reach it. OAuth window opens but immediately fails — Make sure your Configview account is verified. Unverified accounts cannot complete OAuth. Check with your Configview admin if you’re not sure. “Invalid Host header” / 421 response — The server’s allowed-hosts list is derived from the configured SITE_URL. If you’ve recently changed your subdomain, ask your Configview admin to confirm SITE_URL matches the URL you’re connecting to. Connector connects but tools return “no queries found” — Confirm at least one saved query exists (Dashboard > Saved Queries). The connector exposes saved queries only; an empty catalog means nothing for the assistant to call. “Token is unknown, expired, or revoked” — Your access token aged out or was revoked. Reconnect the connector in your client’s settings; the OAuth flow will issue a fresh token. Need to reset everything — Remove the connector from your client and re-add it. Any old tokens are invalidated automatically. Your saved queries and dashboard data are unaffected. Still stuck? — Every MCP tool call is logged in your Configview activity log with the user, tool name, and outcome. Check it for the specific error, or share the timestamp with Configview support.

Next step

Connection done? Head to Saved Queries to see what your assistant can call today and learn how to add your own.

Platform

Integrations

What it is

Your MCP endpoint

Available tools

Connect your MCP client

Claude.ai (web)

Claude Desktop

Claude Code (CLI)

Other MCP clients (Cursor, Continue, custom integrations)

Running queries

Security model

Troubleshooting

Next step

Platform

Integrations

Documentation Index

​What it is

​Your MCP endpoint

​Available tools

​Connect your MCP client

​Claude.ai (web)

​Claude Desktop

​Claude Code (CLI)

​Other MCP clients (Cursor, Continue, custom integrations)

​Running queries

​Security model

​Troubleshooting

​Next step

What it is

Your MCP endpoint

Available tools

Connect your MCP client

Claude.ai (web)

Claude Desktop

Claude Code (CLI)

Other MCP clients (Cursor, Continue, custom integrations)

Running queries

Security model

Troubleshooting

Next step