Following along? Open the live page

Finmagine — free to explore • premium for full access • no app needed

Published: September 12, 2026 | 6 min read | Workflow Guide | Developer & Agent Builder

Everything demonstrated elsewhere in this series can also just be built. The Finmagine Developer Portal turns the platform into API and MCP building blocks — for a script, a Claude connector, or an autonomous trading agent. This workflow follows one request through all four layers: a single live call, a chained recipe, a real multi-agent research run, and the same start/poll/report API that powers Investment Committee — proof that a person clicking through the UI and an unattended agent are running on the exact same backend.

Not investment advice. This video demonstrates platform functionality only, using real live data and a real live API key at the time of recording. Nothing shown constitutes a recommendation to buy, sell, or hold any security — please do your own research or consult a SEBI-registered investment adviser.

Four Layers, One API: Level 1 Raw Data Retrieval, Level 2 Pipeline Filtering, Level 3 Agentic Reasoning, Level 4 Risk Allocation — From a single endpoint to multi-agent consensus — one API key, four layers of increasing abstraction.

The Four Layers

1. Explorer — building blocks

Under Explorer in the Developer Portal, paste an API key once and run any operation directly in the browser — no code required to see a real response. Across Indian and US markets combined, that's 174 operations (60 REST + 62 MCP for India, 26 REST + 26 MCP for the US), and every call returns live data, never a mock.

Layer 1, API Explorer: 60 REST plus 62 MCP operations for India, 26 REST plus 26 MCP for the US, 174 total operations, runnable in the portal with no code, live data never mock — 174 operations across both markets — runnable in-browser before writing a single line of code.

2. Workflows — chained calls

Real agent work is rarely one call — it's a chain. Under Workflows, a named recipe like India Momentum Setup runs a pre-built four-step sequence — market breadth, RS changes, market momentum, conviction ranking — in order, with each step a real independent read call. No manual chaining, no glue code to write.

Layer 2, Agentic Workflows: Market Breadth, RS Changes, Market Momentum, and Conviction Ranking chained in one execution, zero manual chaining, zero glue code — The "India Momentum Setup" recipe — four real read calls, chained automatically.

3. Strategy Lab — reasoning

For actual reasoning rather than data retrieval, Strategy Lab kicks off a real multi-agent chain using your own AI provider key — DeepSeek, Gemini, Groq, or OpenRouter, never a Finmagine-paid key. Each agent's output feeds the next, ending in an actual written report, not a canned answer.

Layer 3, Strategy Lab: a multi-agent chain producing an actual written report, using your own API key with DeepSeek, Gemini, Groq, or OpenRouter — Your key, never Finmagine's — the multi-agent chain runs against the AI provider you choose.

4. Investment Committee — consensus

Strategy Lab's second tab runs on the exact same infrastructure as Investment Committee: three specialist lenses arguing out a bull and bear case, synthesized into a risk-audited allocation. The workflow underneath is start a run, poll for status, pull the report — the same pattern whether it's a person watching a progress bar or an agent running it unattended, on a schedule, with no one clicking through a UI at all.

Layer 4, Investment Committee: Bull Agent and Bear Agent feeding an adversarial debate, checked by a Risk Auditor, producing a risk-audited allocation — Three specialist lenses argue bull and bear cases; a risk audit turns the debate into an allocation.

Why this closes the series' last founding workflow: Long-Term Investor, Swing Trading, and Earnings Season all show what the platform does for a person. This one shows the same platform is also just an API — the Investment Committee run demonstrated end-to-end in the Smart Screener workflow is available to script, schedule, and run unattended, using the identical start/poll/report cycle shown here.