Skip to content
watchingthat
AI & MCP

Every metric,
queryable from Claude.

The Watching That MCP Server ships today. Any MCP-compatible assistant — Claude, ChatGPT, Claude Code — can query your live ad-ops data with an API key.

We call this Agentic Analytics. It’s the web UI you already use, exposed to the assistant you already work in. No replacement. No re-platforming. One API key.

Install the MCP server Try it yourself ↓
Shipping today Read-only by design Claude · ChatGPT · Claude Code
Cached replay · one real question

Not a video.
The actual thing.

One operator-shaped question, one real answer. Captured from a real tenant, sanitised, cached as an asset. No live backend.

claude@operator ~ mcp://watchingthat
cached transcript
$ claude --mcp watchingthat
→ Connected to mcp.acme.watchingthat/mcp
→ Authenticated · 7 tools available
> Which publishers missed their fill-rate threshold today, and what’s the revenue impact?
Thinking...
[tool] list_available_metrics()
[tool] search(metric="fill_rate", dim="partner", range="today")
[tool] search(metric="revenue_at_risk", dim="partner", partner=["pub_14","pub_22","pub_31"])
Three publishers are below fill-rate threshold today.
PartnerFillThreshold24h ∆rev
pub_14 · Regional FAST61.2%78%−$14,300
pub_22 · CTV Network49.8%72%−$31,700
pub_31 · Local Broadcast67.0%75%−$6,850
Combined 24h revenue impact: ≈ $52,850. pub_22 is driving 60% of the shortfall — the SSAI template change on Tuesday looks upstream. Want me to open an Inspect session on pub_22?
Replay captured from a real tenant · partner IDs and dollar values sanitised · no live backend 13s transcript
What AI does here

What AI does here.

Read-only by design. Human signs off. Always.

What it does

Conversational access to your live ad-ops data.

Plain-English questions against every metric, dimension, and threshold your platform tracks. Same read-only surfaces as your web UI — a new way to reach the same data.

What MCP adds

Plain-language reasoning over your live data.

Summarise an open anomaly. Suggest the likely root from prior incidents. Walk a colleague through what’s happening, in plain English. Suggested, not asserted. Statistical, not generative.

What it won’t do

Read-only by design. Human signs off. Always.

Cannot write back. Cannot ship a decision autonomously. No exceptions. The decisions stay with your team — the assistant changes the speed of the conversation, not the accountability.

What MCP is

The protocol, plainly.

MCP — the Model Context Protocol — is an open standard for letting an AI assistant talk to a service on your behalf. The assistant doesn’t replace the web UI; it calls the same read-only surface, with the same permissions, and streams the answer back inline.

Compatible clients
Claude (Desktop & Code), ChatGPT, Cursor, and any other MCP-compatible assistant.
Connection string
mcp.<instance>.watchingthat/mcp?api_key=…
What the assistant can do
Query metrics and dimensions. Look up thresholds. Pull revenue-at-risk figures. Open Inspect sessions by reference.
What it can’t do
Write back. Change configs. Mutate partner feeds. Ship ad-ops decisions autonomously.
Seven tools

What the assistant can reach.

Named plainly. Each one is a read-only call against the Platform surface you already have.

search

Run an ad-hoc query against the live data layer — metric × dimension × time range. The workhorse tool.

list_available_metrics

Every metric queryable in the current scope — fill rate, revenue, requests, impressions, and the rest of the tenant’s library.

list_available_dimensions

Every dimension queryable in the current scope — partner, campaign, daypart, surface, and beyond.

allowed_long_term_dim_combinations

Which metric × dimension combinations the platform allows for long-term historical queries — guards against unbounded sweeps.

list_accessible_dss

Enumerate the data sources the authenticated user can query — by tenant, by surface, by access level.

current_user

Identity and permission context of the authenticated user, including accessible tenants and roles.

save_search

Persist a query (or query result) so it can be re-run, shared with a colleague, or referenced from a later session.

Try it yourself

Install the server. Point it at your data.

If you’re a technical evaluator, this is the shortest path to verification. Install the MCP server in your preferred assistant, authenticate against your tenant, and ask it a question a report wouldn’t answer.

Open the MCP setup docs or walk it →
~/.claude/config.json 
{
  "mcpServers": {
    "watchingthat": {
      "url": "mcp.acme.watchingthat/mcp",
      "headers": {
        "x-api-key": "$WT_API_KEY"
      }
    }
  }
}
Replace acme with your tenant. API keys issued on connection.

Ask your
ad-ops data a question.

Agentic Analytics, shipping today. One API key, one assistant, one real answer.

One API key · one assistant · one real answer Walk it with a strategist or try the MCP server yourself →