Serious AI-assisted development has a habit of turning into a stack of subscriptions: a coding-assistant plan, a market-data API, a backtesting service, model credits. Each one is reasonable on its own, and together they quietly become a monthly bill that also locks you into someone else's rate limits.
I wanted to know how far you can get without any of that. The answer turned out to be: surprisingly far. Here's the stack I ended up with, running at $0 recurring cost and scripted so it rebuilds from scratch.
The goal
A capable, agentic AI workflow — coding and quantitative research — running entirely on free and local infrastructure, and reproducible so it isn't a one-off config living on a single machine.
What I wired together
Agentic coding runs through models by default and fall back to the free tiers on Groq and OpenRouter when a task wants more horsepower. Same agent, same workflow, whether it's fully offline or tapping a free hosted model.
Market data and research run on OpenBB, with MCP servers set up so the AI agent can drive market-data queries directly. This is the part that changed how it feels to work: the agent doesn't just suggest code that would fetch data — it calls a structured tool and pulls the actual data. VectorBT handles fast, vectorized backtesting inside the same environment.
The whole thing is documented and scripted, so it rebuilds rather than rots.
Why MCP is the unlock
The difference between "AI that writes code about your data" and "AI that operates on your data" is a tool interface. MCP (the Model Context Protocol) gives agents a structured way to call real tools — market data, internal APIs, a database — instead of hallucinating what the response might look like. Once the OpenBB layer was exposed over MCP, the agent could answer research questions by actually running the query.
The tradeoffs (honest version)
Local models aren't frontier models. For heavy reasoning I still route to a free hosted tier, and there are days the free tiers are rate-limited. The win isn't "local is as good as GPT-class" — it's that local-first with free-tier fallback covers the large majority of real work at zero marginal cost, and keeps your code and data on your machine.
Takeaway
If you want agentic tooling without committing to a stack of subscriptions — or you want agents that can operate tools rather than just generate text — local-first routing plus MCP is a genuinely practical pattern in 2026. You get modern capability with the cost and privacy under your control.
I'm writing up each piece of this stack in more detail. If there's a part you want to see first — the Aider/Ollama routing, the OpenBB MCP server, or the VectorBT backtests — say so in the comments.
SOCIAL SHARE CARD GENERATOR