What it is

The problem

Ticmint's numbers lived in a dozen tools that disagreed with each other: ads, CRM, outbound, product analytics and search. Reports took hours, and two people could quote different figures for the same week. The stack makes one source of truth that states where every figure came from.

How it works

1. Wake and resolve timeA schedule starts the run. It reads the date from the shell, never from model memory, and closes the previous month if a boundary passed.
2. Read stateIt fetches the current dashboard document and keeps its version number for the final write.
3. Pull everything in one passOne script fans out to GA4, Search Console, Meta Ads, Explee and Google Ads, then prints one JSON line. A failing source becomes null, never an estimate.
4. Call the tools only the agent can reachMixpanel funnels and Ubersuggest rank tracking come in through MCP.
5. Apply the rulesImplausible values are dropped and the last good value is kept. Fields that need a human are on a denylist and never touched.
6. Write once, then notifyThe whole months array is rebuilt in memory and written pinned to the version it read. The founder digest goes to Slack at most once a day.

Architecture

Legend

• HTTPS pull
• MCP tool call
• Versioned write
• Schedule trigger

Layer 3 · Schedulerdashboard refresh · ads report · weekly sheets · form test · SEO tracking · one-off audits

Sources · 10+

Meta AdsGraph API · long-lived token

Google AdsGAQL · OAuth2 + dev token

GA4Data API · OAuth2 refresh

Search ConsoleOAuth2 · paged rows

FreshsalesREST · API key · 1k req/h

ExpleeREST · lifetime totals only

Instantly · LemlistREST v2 · key / Basic

MetabaseQuestion API · token

MixpanelMCP · Insights query

HubSpotMCP · forms, CRM, AEO

UbersuggestMCP · SSE · rank tracking

WordPressREST + MCP · AIOSEO

Layer 2 · Adapters

connectors/ · BaseConnector

• authenticate()
• fetchAllPages() cursor / offset
• normalize() to standard schema
• getMetrics(start, end)
• withRetry() 429 to 2^n backoff
• safeDivide() never NaN
• 12 files · 6,000+ lines

MCP gateway

• tools/list discovery
• typed tool schemas
• OAuth granted by owner
• agent-invoked, not scripted

Processing

pull_api_data.js

• fetch() only · zero deps
• env: .env or cloud secrets
• stdout: one JSON line
• failed source → null + errors[]
• idempotent · date arg

Agent runtime

• joins API + MCP payloads
• applies rules.ajinkya/*
• manual-field denylist
• anomaly → skip, keep last good
• months[] rebuilt in memory
• write pinned to if_version
• lastRunNotes on every skip

Layer 1 · Context

• CLAUDE.md business brain
• skills report · ads · blog
• memory 30 saved rules
• read at the start of every run

Layer 4 · State and output

Dashboard · V1, V2

• doc dashboard/snapshot
• compare-and-set: if_version
• V2 = exact mirror of V1

Google Sheets × 5

• raw-cell writes only
• formula guard · no col insert

Slack digest

• founders · max 1 per day

ClickUp

• tasks · 100 calls/day cap

WordPress writes

• backup → changelog → write
• verify · 301s are permanent

Layer 5 · Guardrailsmanual-field denylist · optimistic concurrency · anomaly skip keeps last good value · backup → changelog → rollback · read-only audits · bot checks are never bypassed

Impact

North-star metrics it tracks: for enterprise, sales-qualified lead to won deal to activated account. For self-serve, signup to first event created to first event published.

Built with

Decisions that mattered

• Direct connectors for any figure that must be date-exact, because wrappers can silently ignore date ranges.
• A failed source returns null. The job keeps the last good value and logs one line. It never estimates.
• Dashboard writes use compare-and-set, so a manual edit and a scheduled run cannot overwrite each other.
• Thirty rules, each one fact with its reason, each from a number that was once wrong.

What is next

• Open-source the connector contract.
• Add more sources as the team asks new questions.