The weekly check
get_insights with last_days: 7, then get_overview for the headline figures.
Insights first is the right order: the panel is sorted worst-first, so the answer starts with what is wrong rather than with a wall of numbers.
Find the tool worth fixing
list_tools over 30 days, then get_tool on the worst offender.
The useful ranking is calls × failure rate, not failure rate alone — a tool at 20% over 30 calls matters less than one at 60% over 900.
Diagnose one tool
get_tool. The breakdown separates three different problems: high bad_args is a schema problem, high empty is a silent-empties problem, and mostly-ok-but-still-retried is a description problem.
Price the server
get_overview with metrics: ["cost", "sessions", "response_size"], then metrics: ["tools"] for the per-tool breakdown.
Find dead weight
get_insights with last_days: 90. The dead-tool rule prices schema bytes against sessions in range.
Ninety days matters — a tool used monthly is dead in a 7-day window and alive in a 90-day one.
Compare two periods
get_overview with metrics: ["first_call_success"] and last_days: 14. The deltas are already computed against the preceding window of equal length — the agent should read them rather than making a second call and subtracting.
Investigate one tool’s behaviour in a session
get_tool, reading follows. Nightly, so it lags a day. See Tool pairs.
Audit everything
list_mcps, then get_insights per MCP.
Setting a server up
create_mcp, then the agent should send you to the dashboard for the key — there is no create_key tool, deliberately. See Tools.
Housekeeping
rename_mcp, then revoke_key. A key is addressed by its prefix, which is what list_keys shows — an ambiguous prefix is refused rather than guessed, because revoking the wrong key is a silent outage.
invite_member. Give the name: without it the team list shows an address where a person should be.
Writes and confirmation
Three tools are marked destructive, and your client should confirm before calling them:delete_mcp, revoke_key and remove_member.
Two things worth knowing before you let an agent loose:
Your role still applies. A member’s agent can read every number and change nothing, refused with the same sentence the dashboard would give. See Roles and permissions.
update_notifications replaces rather than merges. Ask the agent to read the current settings first, or every event it leaves out silently falls back to its default:
Prompting notes
Name the server. The tools resolve an MCP by name, so “the orders server” saves alist_mcps round trip. Two MCPs with the same name are refused rather than guessed at.
Say the window. “This week”, “the last 90 days”, “March” all work. Without one you get the last 30 days.
Ask for one metric when you want one metric. metrics: ["calls"] is one small answer; the whole overview is many. A good agent narrows on its own, and saying so helps.
Say what you want changed, not how. The connector covers everything the dashboard does except ingest, minting a key and deleting the account. If an agent says it cannot do something, check your role before assuming the tool is missing.
Expect a day’s lag on three figures. First-call success, retries and tool pairs come from the nightly pass. An agent reading nightly_as_of will say so.