Do I need to open ports or run a VPN to use this?

No. The Query Streams Network Agent runs inside your network and opens a single outbound, encrypted connection to mcp.querystreams.com. Your AI client connects to the cloud, never to your database. There is no inbound port, no static IP allowlist, and no VPN. The agent can run on a laptop, a VM, on-prem hardware, or inside any cloud that allows outbound HTTPS.

The MCP server adds a typical <500 ms of overhead on top of the underlying database query, dominated by the agent-to-cloud round-trip. Schema-tool calls (qs_get_connector_schema, qs_list_connectors) are usually under 50 ms because they hit cached metadata. Result streaming uses Server-Sent Events so the AI client sees rows as soon as the agent emits them. You don't wait for the whole result set before the LLM starts reasoning.

Which AI tools can I use with Query Streams MCP?

Any MCP-compatible client. The big ones tested end-to-end: Claude Desktop, Claude Code, Cursor, ChatGPT (Plus / Team / Enterprise via OpenAI Apps + Agents SDK), Grok (xAI Remote MCP Tools), Windsurf, Zed, Gemini CLI, Continue, Cline, VS Code Copilot, Codex, and Goose. Plus the long tail of 500+ MCP clients in the public directory. Query Streams ships a standards-conformant Streamable HTTP server with token-based auth, so any client that speaks MCP plugs in with one config block.

Can I revoke an MCP key?

Yes, instantly. Go to the MCP page in your portal, find the key, and click Revoke. Every in-flight call from that key starts returning UNAUTHORIZED within seconds. You can also disable MCP entirely for your user from Settings, or an org-admin can disable MCP for the whole org via the plan.

How is MCP usage billed against my data realm?

Query Streams bills MCP execute calls in uncompressed bytes, while every other access method (Excel add-in, Google Sheets add-on, Nova AI, Portal query builder) bills in compressed bytes. The reason is the wire protocol: the Model Context Protocol specification's standard transports (stdio and Streamable HTTP) do not reliably support payload compression to AI clients, so the bytes leaving the cloud over MCP are uncompressed. For queries that return repeating values, low-cardinality columns, or null-heavy rows, expect MCP to count roughly 5-10x more billable bytes than the same query run through Excel or Google Sheets. For high-cardinality numeric data the ratio is usually 2-3x. The per-client byte breakdown is on the Data Realms page in the Usage by client strip.

How does this compare to running postgres-mcp-server myself?

Open-source direct-DB MCPs (postgres-mcp-server, mysql-mcp, the long tail) are a 30-second npm install and they're great for poking at a dev DB from Cursor. For production, the trade-offs are: (1) they require the LLM or developer laptop to reach the database, which means exposing the DB to the public internet, running a VPN, or co-locating the LLM, and security teams typically reject that pattern; (2) they ship raw information_schema only, with no AI-curated metadata, no semantic types, no enum detection, and no implicit foreign keys; (3) they are scope-less and have no AI-loop guardrails, so a runaway LLM can pile thousands of accidental queries onto a DB before anyone notices; (4) 12 databases means 12 mcp.json entries the developer must manage. Query Streams gives you one outbound-only agent, one key, every connector, Schema Intelligence, per-key scopes, runaway-AI safety nets, and an audit trail in event_logs.

What happens if the AI tries to write or delete data?

The C# Network Agent runs every dispatched query through a hardcoded read-only validator before it ever reaches your database. SELECT, WITH, and EXPLAIN are accepted; everything else (INSERT, UPDATE, DELETE, DROP, TRUNCATE, ALTER, etc.) returns READONLY_VIOLATION and the LLM sees the rejection in the tool response. The validator runs in the agent, not in the cloud, so it can't be bypassed by a compromised cloud surface.

Do I have to run Schema Intelligence to use Query Streams MCP?

No. Schema Intelligence is opt-in per (connector, database) pair, and MCP works fine without it. Without SI, the AI gets bare schema (types, primary keys, formal foreign keys, indexes) and writes basic queries. With SI enabled, the AI gets six additional layers of curated metadata: AI-curated descriptions, table classifications and business domains, 18 semantic types per column, sample values from your real data, enum detection with frequency distributions, and AI-discovered foreign keys. Run time scales with table count: a small database under 100 tables completes in around 10 minutes; a typical mid-size database (a few hundred tables) finishes in 15-25 minutes; a large enterprise database with 1,500+ tables can take 45-60 minutes for a full scan. Subsequent refreshes after schema changes are incremental and much faster. SI never writes to your database, never changes your schema, and runs through your Network Agent, not in the cloud.

Can I try it without committing to a plan?

Yes. New accounts get a 30-day free trial of MCP on signup, no credit card required. Sign up at my.querystreams.com, install the Network Agent (Windows MSI / macOS .pkg / Linux deb), connect a database or a SaaS API, then generate a key on the MCP page. Total time from signup to first AI tool call is usually under 10 minutes. After the trial, ongoing MCP access is included on the paid Personal, Business, and Enterprise plans. The Personal plan is intentionally low-cost with generous included data-realm bytes for solo users and small teams.

What does Query Streams log when the AI uses MCP?

The cloud writes structural metadata only (tool name, scope used, key id, latency, error code, and the calling user/org) to event_logs. The cloud does not log the SQL the LLM generated or the rows the agent returned; those flow over the encrypted cloud connection and exist only in transit. Admins can answer 'who used MCP last week' with a single query against event_logs. Database-level audit, if your DBA has it enabled, sees the actual SQL the agent submitted, just like any other query.

Do I have to set up MCP just to chat with my data?

No. Nova AI is built into the Query Streams web portal and works against every connector in your account out of the box, without any MCP setup, JSON config files, or separate AI vendor subscription. Nova AI is included in every Query Streams plan with credits scaled by tier. Use Nova AI when you want to chat with your data inside Query Streams (no setup, no extra tools); use MCP when you want your existing AI client (Claude, Cursor, ChatGPT, Grok, Windsurf, Zed, etc.) to reach your data alongside whatever else it does for you. Same agent, same connectors, same Schema Intelligence underneath.

Model Context Protocol

Query your database from Claude, Cursor & ChatGPT.

Q: Which AI tools can I use with Query Streams MCP?

Any MCP-compatible client. The big ones tested end-to-end: Claude Desktop, Claude Code, Cursor, ChatGPT (Plus / Team / Enterprise via OpenAI Apps + Agents SDK), Grok (xAI Remote MCP Tools), Windsurf, Zed, Gemini CLI, Continue, Cline, VS Code Copilot, Codex, and Goose. Plus the long tail of 500+ MCP clients in the public directory. Query Streams ships a standards-conformant Streamable HTTP server with token-based auth, so any client that speaks MCP plugs in with one config block.

Q: Can I revoke an MCP key?

Yes, instantly. Go to the MCP page in your portal, find the key, and click Revoke. Every in-flight call from that key starts returning UNAUTHORIZED within seconds. You can also disable MCP entirely for your user from Settings, or an org-admin can disable MCP for the whole org via the plan.

Q: How is MCP usage billed against my data realm?

Query Streams bills MCP execute calls in uncompressed bytes, while every other access method (Excel add-in, Google Sheets add-on, Nova AI, Portal query builder) bills in compressed bytes. The reason is the wire protocol: the Model Context Protocol specification's standard transports (stdio and Streamable HTTP) do not reliably support payload compression to AI clients, so the bytes leaving the cloud over MCP are uncompressed. For queries that return repeating values, low-cardinality columns, or null-heavy rows, expect MCP to count roughly 5-10x more billable bytes than the same query run through Excel or Google Sheets. For high-cardinality numeric data the ratio is usually 2-3x. The per-client byte breakdown is on the Data Realms page in the Usage by client strip.

Q: How does this compare to running postgres-mcp-server myself?

Open-source direct-DB MCPs (postgres-mcp-server, mysql-mcp, the long tail) are a 30-second npm install and they're great for poking at a dev DB from Cursor. For production, the trade-offs are: (1) they require the LLM or developer laptop to reach the database, which means exposing the DB to the public internet, running a VPN, or co-locating the LLM, and security teams typically reject that pattern; (2) they ship raw information_schema only, with no AI-curated metadata, no semantic types, no enum detection, and no implicit foreign keys; (3) they are scope-less and have no AI-loop guardrails, so a runaway LLM can pile thousands of accidental queries onto a DB before anyone notices; (4) 12 databases means 12 mcp.json entries the developer must manage. Query Streams gives you one outbound-only agent, one key, every connector, Schema Intelligence, per-key scopes, runaway-AI safety nets, and an audit trail in event_logs.

Q: What happens if the AI tries to write or delete data?

The C# Network Agent runs every dispatched query through a hardcoded read-only validator before it ever reaches your database. SELECT, WITH, and EXPLAIN are accepted; everything else (INSERT, UPDATE, DELETE, DROP, TRUNCATE, ALTER, etc.) returns READONLY_VIOLATION and the LLM sees the rejection in the tool response. The validator runs in the agent, not in the cloud, so it can't be bypassed by a compromised cloud surface.

Q: Do I have to run Schema Intelligence to use Query Streams MCP?

No. Schema Intelligence is opt-in per (connector, database) pair, and MCP works fine without it. Without SI, the AI gets bare schema (types, primary keys, formal foreign keys, indexes) and writes basic queries. With SI enabled, the AI gets six additional layers of curated metadata: AI-curated descriptions, table classifications and business domains, 18 semantic types per column, sample values from your real data, enum detection with frequency distributions, and AI-discovered foreign keys. Run time scales with table count: a small database under 100 tables completes in around 10 minutes; a typical mid-size database (a few hundred tables) finishes in 15-25 minutes; a large enterprise database with 1,500+ tables can take 45-60 minutes for a full scan. Subsequent refreshes after schema changes are incremental and much faster. SI never writes to your database, never changes your schema, and runs through your Network Agent, not in the cloud.

Q: Can I try it without committing to a plan?

Yes. New accounts get a 30-day free trial of MCP on signup, no credit card required. Sign up at my.querystreams.com, install the Network Agent (Windows MSI / macOS .pkg / Linux deb), connect a database or a SaaS API, then generate a key on the MCP page. Total time from signup to first AI tool call is usually under 10 minutes. After the trial, ongoing MCP access is included on the paid Personal, Business, and Enterprise plans. The Personal plan is intentionally low-cost with generous included data-realm bytes for solo users and small teams.

Q: What does Query Streams log when the AI uses MCP?

The cloud writes structural metadata only (tool name, scope used, key id, latency, error code, and the calling user/org) to event_logs. The cloud does not log the SQL the LLM generated or the rows the agent returned; those flow over the encrypted cloud connection and exist only in transit. Admins can answer 'who used MCP last week' with a single query against event_logs. Database-level audit, if your DBA has it enabled, sees the actual SQL the agent submitted, just like any other query.

Point your AI assistant at every database and API in your account — even one behind a corporate firewall — over the Model Context Protocol. No VPN: the database can sit in a corporate office and your AI client at home. Only a revocable MCP key lives in your config — generate one and connect in about 90 seconds, read-only by default.

Try the Free Tier Team Pricing

No VPN — no inbound port Connect in ~90 seconds One key, all 59 sources Read-only at the agent

Your AI assistant

Claude · Cursor · ChatGPT · Grok

holds one MCP key

Query Streams cloud

mcp.querystreams.com

relays both ways

Network Agent

runs inside your network

read-only validator

— reaches your 59 data sources —

+ 51 more

The agent opens one outbound, encrypted connection — requests and results flow both ways over it. No inbound port, no VPN.

The big idea

No database credentials in your AI config. No VPN.

Open-source MCPs put a raw connection string in your client config — so the AI (and your laptop) must be able to reach the database directly. Query Streams puts only a revocable MCP key there. The real credentials live on the Network Agent inside your network, so a database in a corporate office is reachable from an AI client at home — securely, with no VPN and nothing exposed.

mcp.jsonyour AI client config

{
  "mcpServers": {
    "querystreams": {
      "url": "https://mcp.querystreams.com",
      "headers": {
        "Authorization": "Bearer YOUR_MCP_KEY"
      }
    }
  }
}

In your AI client config

One MCP key (a revocable token)
The cloud URL — mcp.querystreams.com
A per-key scope: read / analyze / execute

Never leaves your network

Database usernames & passwords
Connection strings & host names
The database itself — no inbound port, no VPN

What an `npm install` MCP can’t match

A developer can stand up a direct-database MCP in 30 seconds. Here’s what changes when you need it in production.

No VPN, across any network

The Network Agent opens one encrypted connection out to the cloud, so a database in a corporate office and an AI client at home connect with no inbound port, no IP allowlist and no VPN — the pattern security teams already approve.

Schema Intelligence, not `information_schema`

Every schema call returns AI-curated descriptions, a semantic type per column, sample values, enum distributions and discovered foreign keys — so the model writes correct SQL the first time instead of guessing.

Read-only & scoped per key

A hardcoded validator in the agent allows only SELECT / WITH / EXPLAIN — everything else is rejected. Each key carries its own scope, so a browse-only key never gets execute.

Query Streams MCP vs a direct-database MCP

Open-source direct-DB servers are great for poking at a dev database. For production data, the gaps add up.

Direct-DB MCP (`postgres-mcp-server`…)

The database must be reachable from the LLM or your laptop
A raw connection string sits in your client config
Bare information_schema — the model guesses columns
Scope-less — the string grants whatever it can do
One MCP entry per database; SaaS APIs not supported

Query Streams MCP

Connect across networks — no inbound port, no VPN
Only a revocable MCP key in the config; creds stay home
Schema Intelligence — descriptions, semantic types, FKs
Per-key read / analyze / execute scopes
One key, all 59 sources — databases and SaaS APIs as SQL

Direct-DB MCPs are a fine 30-second npm install for a dev database. Query Streams is what you reach for when the data is production, the database is behind a firewall, and a connection string in an AI config is a non-starter.

One key. Every source, every AI client.

38 database connectors and 8 API connectors, surfaced through one MCP endpoint and reachable from any standards-conformant MCP client. A federated query is a saved query like any other, so an assistant can run one and get the already-joined rows back — “how did each region do last week” answered across two databases, in chat.

38 database connectors

8 API connectors — cached as SQL

Claude

Cursor

ChatGPT

Grok

Claude Code Windsurf Zed Gemini CLI Continue VS Code Copilot Goose + many more MCP clients

Standards-conformant MCP server with token auth. Add a connector any time — the AI sees it on the next qs_list_connectors call.

No VPN. Connect in about 90 seconds.

Install the agent once; after that, generating a key and connecting an assistant takes about 90 seconds — even across networks.

Install the Agent

Run the outbound-only Network Agent inside your network — Windows, Linux or macOS. A one-time setup; no inbound port, no VPN.

Connect your sources

Add any of 11 databases and 8 API connectors. Schema Intelligence profiles them so the AI sees real columns.

Generate one MCP key

Create a key in the portal, pick its scope (read / analyze / execute), and revoke it any time.

Paste it & ask

Drop the key into Claude, Cursor, ChatGPT or Grok and ask — about 90 seconds, no VPN, no credentials in the config.

MCP Server FAQ

Do I need to open ports or run a VPN?

No. The Network Agent runs inside your network and opens a single outbound, encrypted connection to mcp.querystreams.com — requests and results travel both ways over that one agent-initiated channel. Your AI client connects to the cloud, never to your database, so there’s no inbound port, no static IP allowlist and no VPN. That’s what lets a database in a corporate office answer an AI client at home. The agent runs on a laptop, a VM, on-prem hardware, or any cloud that allows outbound HTTPS.

Are my database credentials in the AI config?

No. Only a revocable MCP key (a token) and the cloud URL go in the client config. The real database credentials stay on the Network Agent inside your network and are never sent to the AI client or the cloud. Revoke a key any time from the portal.

Which AI tools can I use?

Any MCP-compatible client. Tested with Claude Desktop, Claude Code, Cursor, ChatGPT, Grok, Windsurf, Zed, Gemini CLI, Continue, VS Code Copilot and Goose — and it works with any other MCP-compatible client.

What if the AI tries to write or delete data?

The Network Agent runs every query through a hardcoded read-only validator before it reaches your database. SELECT, WITH and EXPLAIN are accepted; everything else (INSERT, UPDATE, DELETE, DROP, ALTER…) is rejected. The validator runs in the agent, not the cloud, so it can’t be bypassed by a compromised cloud surface.

Which plans include the MCP Server?

MCP Server is included on the Business and Enterprise plans. New accounts get a 30-day free trial on signup with no credit card required, so you can generate a key and connect an assistant before committing.

Connect your AI assistant to your data.

Generate one MCP key, drop it into Claude, Cursor, ChatGPT or Grok, and ask — no VPN, one outbound-only agent, every connector you’ve added, read-only by default.