Bring Your Own AI Agents to Shopify Data via MCP
Bring-your-own AI agents means your team already uses Claude, ChatGPT, Codex, Grok, or another MCP client, and you point those clients at company-owned, verified ecommerce metrics. Pasted CSVs and personal Shopify API keys drop out of the path. On Finsi, those clients connect over Finsi MCP (read-only) and follow the same approval rules as Finsi's Agent Library when money or customers would change.
This post covers the operating model. Connection steps and tool lists live on the MCP page; the full agent platform is on AI agents.
The laptop agent problem
A common pattern in 2026:
- Someone exports Shopify, Meta, and Klaviyo.
- They paste the CSV into a chat.
- The model answers with confident numbers that do not match finance's definitions.
- Or they create a personal Admin API token so a local agent can ship a change tonight.
That path is fast for one afternoon and expensive over a quarter. Data goes stale, metrics get invented, keys stick around after people leave, and prompts leave with them. Finsi's docs name those modes directly: invented numbers, orphan keys, know-how on one laptop, agents that never share context.
Company-owned agents on one metric layer
Finsi's model:
- Connect Shopify, Klaviyo, Meta Ads, Google Ads, Recharge and other sources once into a warehouse.
- Define metrics in a semantic layer and calculate them once.
- Attach agents — Finsi's or yours — with scoped access.
- Require human approval before execution; log what ran.
Bring-your-own still sits under governance. MCP clients on Finsi only receive read tools, so analysis can happen in Claude or ChatGPT without a silent write path. Writes still go through approval inside Finsi.
What Finsi MCP exposes (high level)
From the live MCP page:
| Area | Examples of read-only tools |
|---|---|
| Brand context | list_companies, get_company, list_brand_data_sources, get_brand_voice |
| Analytics | get_revenue_summary, get_ads_performance, get_email_performance, get_top_products, get_customer_cohorts |
| Metrics catalog | list_metrics_catalog, run_catalog_metric (no raw SQL; date ranges capped at 366 days) |
| Reports / recommendations | list_business_reports, list_business_recommendations, list_ads_recommendations, list_flow_recommendations |
Controls stated on the page: read-only; one brand per API key; every call logged (sensitive values stripped); cannot change ads, flows, listings, or prices via MCP.
Endpoint (as published): https://api.finsi.ai/mcp with Authorization: Bearer finsi_<your_key>. Claude Desktop, Claude Code, ChatGPT, and other streamable-HTTP MCP clients can attach the same URL and header.
How this coexists with the Agent Library
You do not have to choose only BYO or only library agents.
- Agent Library — ECHO, PULSE, MUSE, MAGNET, SCOUT, BEACON, SAGE, RIVAL, Lead Analyst — ready workflows for email, ads, creative, audiences, profit, search, research, competitors, reporting.
- BYO over MCP/API — your preferred model for analysis and drafting against the same metrics.
- Both — e.g. Lead Analyst for the weekly digest, Claude via MCP for ad-hoc questions, PULSE for budget proposals waiting in the approval queue.
Group chat across several agents on one question is part of the published platform: combined proposal, not five disconnected answers.
Setup sketch (details on /ai/mcp)
- Connect data sources in Finsi.
- Create an agent connection / API key for the brand.
- Add Finsi as an MCP server in your client (Claude Desktop JSON,
claude mcp add, or ChatGPT's MCP settings). - Ask questions that map to tools (revenue vs returning customers, ROAS/CAC, cohorts, open ads recommendations).
- When you want an action in Meta or Klaviyo, use the approval path in Finsi — MCP will not push the change itself.
FAQ
Can I bring Claude or ChatGPT as my own agents?
Yes. Finsi documents Claude, ChatGPT, Codex, Grok, and any MCP client over the same MCP endpoint and rules. They read verified metrics; they do not get a silent write path.
Is Finsi MCP read-only?
Yes. The MCP server only reads. Changes to money and customers go through approval in Finsi.
Shopify MCP vs Finsi MCP?
Shopify's own MCP or connectors, when you use them, center on Shopify Admin and store context. Finsi MCP sits on multi-source metrics already joined and defined in Finsi (Shopify plus ads, email, and subscriptions, with a semantic layer). Reach for Shopify-native tooling for store admin tasks. Reach for Finsi MCP when the question spans the stack and has to match finance's definitions. Shopify's product surface changes; verify current Shopify MCP docs separately.
How do I revoke an agent's access?
Revoke or rotate the brand API key and adjust per-agent access in Finsi. Company-owned keys beat leftover personal tokens.
Do BYO agents share context with Finsi agents?
On Finsi, agents can work from the same warehouse and semantic layer and can be grouped on one question. That is the published answer to "marketing's agent vs finance's inventory reality."