:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.
Python ab-testingai-analyticsanalyticscdpdata-warehouseexperimentsfeature-flagsjavascriptproduct-analyticspythonreactsession-replaysurveystypescriptweb-analytics
import { formatMs, formatNumber } from '../dashboard/formatters'
import type { KPIData, ToolRow } from '../mcpDashboardOverviewLogic'
// At/above this error rate we call the tool out as flaky in the headline.
const FLAKY_ERROR_RATE_PCT = 5
export interface FirstLookChip {
key: string
label: string
value: string
tone?: 'danger'
/** When set, render this harness's logo in place of an icon. */
harness?: string
}
const possessive = (company?: string | null, project?: string | null): string =>
company ? `${company}'s` : project ? `${project}'s` : 'Your'
/** One-line, personalized "your server is live" headline derived from real metrics. */
export function buildHeadline({
company,
project,
topTool,
client,
kpis,
}: {
company?: string | null
project?: string | null
topTool: ToolRow | null
client: string | null
kpis: KPIData
}): string {
const who = possessive(company, project)
if (topTool) {
const agents = client ? `${client} agents` : 'Agents'
const flaky =
topTool.error_rate_pct >= FLAKY_ERROR_RATE_PCT
? `, though it errors ${topTool.error_rate_pct}% of the time`
: ''
return `${who} MCP server is live — ${agents} leaned on ${topTool.tool} ${formatNumber(topTool.total_calls)} times, your busiest tool${flaky}.`
}
// No per-tool rows resolved yet — fall back to top-line counts.
return `${who} MCP server is live — ${formatNumber(kpis.toolCalls.value)} tool calls across ${formatNumber(kpis.sessions.value)} sessions so far.`
}
/** A concrete question for PostHog AI, tied to their actual worst tool so the answer is useful. */
export function buildMaxPrompt({
worstErrorTool,
topTool,
}: {
worstErrorTool: ToolRow | null
topTool: ToolRow | null
}): string {
if (worstErrorTool && worstErrorTool.error_rate_pct > 0) {
return `My MCP server's ${worstErrorTool.tool} tool errors ${worstErrorTool.error_rate_pct}% of the time over ${formatNumber(worstErrorTool.total_calls)} $mcp_tool_call events. Show me the most common errors and what's causing them.`
}
if (topTool) {
return `My busiest MCP tool is ${topTool.tool}. Show me how its usage and latency have trended over the last 7 days and whether anything looks anomalous.`
}
return `Summarize how agents are using my MCP server from $mcp_tool_call events — top tools, error rates, and what they're trying to do.`
}
// Friendly harness label → where to paste the prompt. Falls back to "your agent".
const HARNESS_PASTE_TARGET: Record<string, string> = {
'Claude Code': 'Claude Code',
'Claude Desktop': 'Claude Desktop',
'Claude Code (VS Code)': 'Claude Code',
Cursor: 'Cursor',
'OpenAI Codex': 'Codex',
'VS Code': 'VS Code',
Windsurf: 'Windsurf',
}
/** A copy-paste prompt for the customer's own coding agent, framed for their dominant client. */
export function buildEditorPrompt({ client }: { client: string | null }): { label: string; prompt: string } {
const where = (client && HARNESS_PASTE_TARGET[client]) || 'your agent'
return {
label: `Paste this into ${where}`,
prompt: 'Using the PostHog MCP, pull my $mcp_tool_call events and tell me which tools error most and why.',
}
}
/** Metric highlight chips; any chip whose data is missing is dropped. */
export function buildChips({
topTool,
worstErrorTool,
kpis,
client,
}: {
topTool: ToolRow | null
worstErrorTool: ToolRow | null
kpis: KPIData
client: string | null
}): FirstLookChip[] {
const chips: FirstLookChip[] = []
if (topTool) {
chips.push({ key: 'top-tool', label: 'Busiest tool', value: topTool.tool })
}
// Skip the flakiest chip when it's the same tool as the busiest one — the busiest chip
// and the headline's flaky clause already name it; a third mention reads as noise.
if (worstErrorTool && worstErrorTool.error_rate_pct > 0 && worstErrorTool.tool !== topTool?.tool) {
chips.push({
key: 'worst-error',
label: 'Flakiest tool',
value: `${worstErrorTool.tool} · ${worstErrorTool.error_rate_pct}%`,
tone: 'danger',
})
}
if (kpis.p95LatencyMs.value > 0) {
chips.push({ key: 'p95', label: 'p95 latency', value: formatMs(kpis.p95LatencyMs.value) })
}
if (client) {
chips.push({ key: 'client', label: 'Top client', value: client, harness: client })
}
return chips
}