Connect your own agent

Ask your SourceMedium data questions from the AI you already use. Every agent uses one URL and your SourceMedium sign-in. It reads, never writes, and you never paste BigQuery, Shopify, or Meta credentials. Want SourceMedium to write the SQL for you? Use SourceMedium Agent in Slack.

Any workspace member can connect. Sign in with your membership email. The session is read-only and sees only what you already can.

The connection URL

https://mcp.sourcemedium.com/mcp

Keep the /mcp. The AI Agents page in your workspace has it ready to copy and confirms when your agent connects. Another MCP client may work with the same URL; these are the ones we keep steps for.

When you connect SourceMedium to an AI service, queries and the resulting data may be sent to the AI provider you choose. Access remains limited by your existing SourceMedium organization membership and connected-source permissions. This connection does not provide semantic PII detection or field-level masking. Do not request customer contact, address, or other direct personal data unless it is necessary and authorized for your use case.

Chat apps

Claude
ChatGPT

On Team or Enterprise, an Owner first adds SourceMedium under Organization settings → Connectors → Add → Custom → Web. Members then connect it under Customize → Connectors.

Open connectors

In claude.ai, go to Customize → Connectors.

Add the connector

On Free, Pro, or Max, click + and choose Add custom connector. On Team or Enterprise, click Connect on the connector your Owner added and skip the next step.

Paste the URL

Paste https://mcp.sourcemedium.com/mcp, click Add, then Connect.

Ask

In a chat, click +, open Connectors, switch SourceMedium on, and ask.

Claude’s Free plan allows one custom connector.

On paid plans the live pages Claude builds can read SourceMedium through this connection. Each viewer needs their own SourceMedium connection to refresh live charts. Check Claude’s sharing options for your plan and workspace. On Team or Enterprise, an Owner may first need to turn on Dashboards (Organization settings → Artifacts) and artifact connectors (Organization settings → Capabilities).

Vendor guide: Claude custom connectors

Developer mode depends on your plan and workspace policy. No toggle? Ask your workspace admin.

Turn on Developer mode

Settings → Security and login → Developer mode.

Add the connection

On the Plugins page, click +. Name it SourceMedium and paste https://mcp.sourcemedium.com/mcp under Connection.

Create it, review the tools, and complete the SourceMedium sign-in.

Ask

In a new chat, enable the SourceMedium plugin and ask.

Vendor guide: Connect to ChatGPT

Agents

Add it

claude mcp add --transport http sourcemedium https://mcp.sourcemedium.com/mcp

Start Claude Code, run /mcp, pick sourcemedium, and sign in.

Ask

Ask a question about your data.

claude mcp remove keeps your saved token. Run claude mcp logout sourcemedium to clear it.

Vendor guide: Claude Code MCP

Add it

codex mcp add sourcemedium --url https://mcp.sourcemedium.com/mcp

codex mcp login sourcemedium

Ask

Restart Codex and ask. Prefer config? Add this to ~/.codex/config.toml instead of the add command:

[mcp_servers.sourcemedium]
url = "https://mcp.sourcemedium.com/mcp"

Send one message

In a new chat, send:

Connect SourceMedium MCP at https://mcp.sourcemedium.com/mcp

Approve and sign in

Approve adding SourceMedium, then sign in with your SourceMedium email.

Ask

Ask a question about your data.

The AI Agents page with Grok Bot selected, showing the message to send and the account to sign in with

Send one message

In a new chat, send:

Connect SourceMedium MCP at https://mcp.sourcemedium.com/mcp

Approve and sign in

Approve adding SourceMedium, then sign in with your SourceMedium email.

Ask

Ask a question about your data.

Add it

hermes mcp add sourcemedium --url https://mcp.sourcemedium.com/mcp --auth oauth

hermes mcp test sourcemedium

Ask

Ask a question about your data.

Hermes Agent CLI only; the Hermes chat apps have no connector settings.

Vendor guide: Hermes Agent MCP

Add it

Add this to ~/.cursor/mcp.json (or Customize → MCPs → Add New MCP Server):

{
  "mcpServers": {
    "sourcemedium": {
      "url": "https://mcp.sourcemedium.com/mcp"
    }
  }
}

In Cursor’s MCP settings, click Login next to sourcemedium.

Ask

Ask Cursor’s agent a question about your data.

Vendor guide: Cursor MCP

Want your coding agent working in your warehouse directly? Add SourceMedium Agent Skills too: it keeps this connection, and builds dashboards and pipelines with your own gcloud login.

When your agent opens SourceMedium

Every agent above sends you to the same SourceMedium sign-in at its sign-in step:

Continue with Google, or enter the email you use for SourceMedium. Use the same account you use in your workspace.
Click Allow access. One connection reaches one organization. If you belong to more than one, pick it under Organization access. To use a different organization later, open a private window and connect again.

A sign-in step on the AI Agents page, showing the SourceMedium sign-in screen and the Organization access screen with Allow access

If connecting fails

Sign-in failed? Remove SourceMedium, add it again (or resend the message), then sign in.
Claude Code? Run claude mcp logout sourcemedium before adding it again.
401, ACCESS_DENIED, or the wrong organization? See Agent connection troubleshooting.

After you connect

One connection is one organization. In several workspaces? Pick one at sign-in. To switch, open a private window, connect again, and pick the other organization. A normal sign-out keeps the previous choice. In Claude Code, run claude mcp logout sourcemedium before adding the server again so the old token is gone. The AI Agents page lists every agent you’ve connected, and when it was last used:

Your own agent on the AI Agents page, listing Claude and Claude Code with when each connected and was last used

What it can reach

Surface	When
SourceMedium reporting data (`sm_transformed_v2`, `sm_metadata`, and any other SourceMedium dataset `get_data_context` lists)	Always
Other datasets in your BigQuery project	Dedicated project, and you are a direct member. The assistant only queries datasets `get_data_context` lists.
Google Sheets tables in those datasets	Once each sheet is shared with your SourceMedium address. See Google Sheets tables.
SourceMedium docs search	Always
Shopify	Once a store is connected
Meta Ads	Once an ad account is connected

Not connected? That tool answers SOURCE_UNAVAILABLE, and warehouse questions still work. Agency partners with a live grant can connect too; a brand- or store-scoped grant sees only those stores and can’t run raw SQL.

Google Sheets tables

A Google Sheet connected to your BigQuery project as a table works like any other table, once SourceMedium can open the sheet. A sheet starts out unshared. SourceMedium opens sheets as one address for your organization and checks each one for you.

Open AI Agents in SourceMedium. The Google Sheets section lists every Sheets table in your warehouse as Ready, Needs sharing, or Couldn’t check. Copy the address shown there.
For each table that needs sharing, click Open sheet, then Share.
Paste the address, choose Viewer, untick Notify people, and click Share.
Back on AI Agents, click Check again. The table shows Ready.

Just added a Sheets table? Click Check again to see it right away. SourceMedium also rechecks every sheet about every 15 minutes, and get_data_context tells your agent which tables it can’t read yet. Only someone who can share the sheet can do this. Not yours? Ask its owner. A sheet in a SourceMedium shared drive is shared by your SourceMedium team. If your Google Workspace blocks sharing outside your company, your admin has to allow the address first. If your agent queries a sheet before it is shared, it gets PROVIDER_PERMISSION_DENIED. The error links the sheet and gives the address, so the agent can tell you exactly what to share.

Your own tables

Your own tables are reachable in a dedicated project when you are a direct member of it, and the agent queries only the datasets get_data_context lists. A dedicated project comes with Pro; on Foundation your data lives in a shared BI project, and the agent reaches SourceMedium’s tables only (Foundation and Pro). The MCP tool reference explains how the catalog reports a table it can see but has no definition for. Your own tables have no SourceMedium definition, so the agent cannot find them by what they mean. Name the table in the question:

According to ops.store_sales, how many units did we sell in September 2026, by store?

Give the dataset and table, without the project: every query runs in the project get_data_context names for your organization, so ops.store_sales is enough. A question that only describes the table (“our stock table”) sends the agent to the catalog, and it may land on a SourceMedium table that covers similar ground. If you do not know the exact name, ask for it by part of the name first (“find the table with inventory in its name”). Example prompts. Each names the table, states the window, asks for the figure, and adds the one caveat its table needs:

According to ops.store_sales, how many units did we sell in September 2026, by store?
According to ops.daily_inventory, what is our latest inventory position? Use the most recent date in the table, and give the total available units and the number of SKUs on that date.
According to ops.purchase_orders, how many open purchase order lines do we have, and how many units are on them? This table keeps a copy of every sync.
According to ops.stock_snapshots, what were total units on hand and their total cost on the 2026-10-01 snapshot, and how does that compare with 2026-09-01?
A table of our own, ops.store_sales, also reports units sold. For September 2026, compare it with SourceMedium's order lines by store and explain any gap.
Which of our own tables hold inventory, and which of them is the freshest right now?

Expect the agent to call get_data_context, then search_data_catalog with the table name, which reports the table as live and uncatalogued, then read its columns (search_data_catalog with the table as dataset.table, or on the earlier tool set INFORMATION_SCHEMA.COLUMNS through run_bigquery_sql), and then the query runs. If the answer does not include a check (rows against distinct keys, nulls, a second method that should agree), ask for one. The numbers carry no SourceMedium validity filter. When you reconcile one of your tables with a SourceMedium table, ask for the match by store, day and SKU, and expect two differences: SourceMedium’s figures apply is_order_sm_valid, so draft, cancelled and refunded orders drop out, and your table may not; and SKU case can differ between the two. Four things can trip a question up, and all four come from how the data was loaded:

Fixed-width status values. Some system exports pad status values to a fixed width, so an exact match on Open can return nothing. If a status count comes back as zero, ask the agent to trim the values first.
Sync copies. A table replicated from a database can keep every sync, so a count of rows is a count of copies. Say “this table keeps a copy of every sync” in the question, and ask for the latest copy of each record with soft-deleted rows left out. If the answer does not mention duplicates, ask whether it checked.
Daily snapshots. A snapshot table holds one row per item per day, so a sum over it means nothing. Ask for “the most recent date in the table” and the figures on that date, as the inventory prompt above does.
UTC freshness. Load and rebuild times come from the warehouse clock, marked “UTC”. Ask for times in your store’s timezone if that is what you compare against.

Set these rules once in a Claude or ChatGPT project and every chat inherits them. See Set up a project for your data.

Your first question

Paste this in:

Compare net revenue for the last complete week against the week before, using SourceMedium’s own metric definition (query_metrics). Tell me which stores, dates, and timezone you used, and show the definition and any filters.

A good answer names your stores, the right dates, and the definition it used. Checking it against a dashboard? Match dates, timezone, stores, and filters first. The MCP tool reference explains every tool.

Limits

Limit	Value
Warehouse rows	500, then marked truncated
Shopify and Meta records per page	100
BigQuery billed bytes	10 GiB per query
Query time	60 s warehouse, 30 s Shopify and Meta
Response size	256 KiB

MCP tool reference

Every tool, and the order that gets correct answers.

Agent connection troubleshooting

Sign-in errors, source denials, and what to send support.

Set up a project for your data

Name your tables once in a Claude or ChatGPT project, with instructions to paste.

SourceMedium Agent

Ask in Slack and let SourceMedium write the SQL and charts.

Start here

SourceMedium Agent

SourceMedium Agent Skills

Your own agent

Reference

The connection URL

Chat apps

Agents

When your agent opens SourceMedium

If connecting fails

After you connect

What it can reach

Google Sheets tables

Your own tables

Your first question

Limits

MCP tool reference

Agent connection troubleshooting

Set up a project for your data

SourceMedium Agent

​The connection URL

​Chat apps

​Agents

​When your agent opens SourceMedium

​If connecting fails

​After you connect

​What it can reach

​Google Sheets tables

​Your own tables

​Your first question

​Limits

​Related

MCP tool reference

Agent connection troubleshooting

Set up a project for your data

SourceMedium Agent

The connection URL

Chat apps

Agents

When your agent opens SourceMedium

If connecting fails

After you connect

What it can reach

Google Sheets tables

Your own tables

Your first question

Limits

Related