TL;DR

A fully operational solo trading stack in 2026 can run at zero recurring infrastructure cost — not zero total cost, because you still own a Mac Mini and pay for electricity, but zero monthly subscription line items. The components: built-in launchd (or systemd) for scheduling, Alpaca's free IEX tier for real-time data on covered names, local Llama/Qwen models via Ollama for cheap LLM research, BYO frontier API keys only for the paths where deep reasoning is actually needed, SQLite for state, and Cloudflare Pages' free tier for dashboards. The lines where the $0 model breaks cleanly: full-SIP consolidated quotes, Level-2 depth, 24/7 cloud compute, real-time options chains with Greeks, and historical tick data. Below: what each component does, what you give up, the vendor free-tier landscape as of April 2026, and the moment each paid tier becomes unavoidable.

Why this matters

A recurring-cost floor is a tax on experimentation. Every month you can't shut a subscription off is a month you can't pivot a strategy or kill a failing research loop without paying anyway. For a retail operator who is still funded by a day job, running at $0/month means the question "should I keep working on this?" has no financial coercion in it — only the actual signal.

This is a constraint that can be dropped once the stack is paying for itself. One-time purchases (a Mac Mini, a historical-tick dataset, a permanent license) are always on the table. The rule is about monthly line items, and it is time-bounded to the phase where the trading stack is not yet generating its own revenue.

The components

Compute: Mac Mini M4 (one-time)

A 16GB Mac Mini M4 idles at ~7W and costs under €700 in April 2026. It runs 24/7 for roughly €15-20/year of electricity in Germany at typical consumer rates. It is wildly more reliable than a cloud VPS because it's not competing with noisy neighbours and doesn't get migrated mid-trade.

What you give up vs cloud: redundancy. A single Mac Mini is a single point of failure. If the local ISP dies, the strategy dies. For long-horizon strategies where a one-hour outage is not existential, this is an acceptable trade. For anything latency-sensitive, it isn't.

Scheduling: launchd (macOS, built-in)

launchd ships with macOS. A launchd plist running StartCalendarInterval replaces every cron-as-a-service SaaS product, and replaces most $20/month cloud schedulers. A typical trading stack ends up with ~10-20 plists: one per scheduled job (pre-market scan, intraday heartbeat, end-of-day reconciliation, weekly re-calibration).

One foot-gun: launchd's StartCalendarInterval fires on Berlin local time, not UTC. If you want a job to run at 14:30 UTC daily, compute the local time for that UTC instant and keep it updated at DST boundaries. Or use StartInterval for relative cadences.

Market data: Alpaca free IEX tier

Alpaca's free tier provides real-time trades and quotes from IEX, which represents roughly 2% of consolidated US equity volume. For highly liquid names (SPY, AAPL, MSFT, NVDA, TSLA), IEX prints are generally within a few basis points of the consolidated NBBO most of the time. For thinly traded names, IEX can have minute-long gaps and is not a reliable price source.

Practical rule: for universes of 50-200 high-liquidity names and holding periods over 5 minutes, IEX data is sufficient for signal generation. For intraday strategies on illiquid names, or anything that needs the consolidated NBBO tick-by-tick, IEX is a trap.

Alpaca also provides free historical bar data (1-minute, 15-minute, daily) back several years. For backtest construction, this is enough for most retail strategies. Historical tick-by-tick is not free.

LLM research: Ollama + Llama/Qwen locally, BYO frontier keys

Ollama on the Mac Mini runs 7B-30B parameter open-weight models at usable speeds. Qwen2.5-32B and Llama 3.3-70B (quantized) both run on 16GB with swapping, though 32GB is smoother. These handle structured tasks reliably: JSON extraction from SEC filings, ticker disambiguation, classification, summarization.

For tasks that genuinely require frontier reasoning — multi-step causal analysis, complex prompt regression testing, anything where hallucination cost is high — fall back to Claude, GPT-5, or Gemini via BYO API key, and pay only for the tokens actually consumed. This is per-request billing, not a subscription. The Token-Cost Optimizer tool estimates the split between local and frontier paths for a given workload.

Rule of thumb from the production Claude agent scaffold (see Building a Production Claude Agent for Finance): 70-90% of agent steps can run on a local 7B-30B model, and the remaining 10-30% on Claude or GPT-5. That keeps monthly frontier spend in the $5-40 range for a single-operator research loop, which is BYO, not a subscription.

State: SQLite

SQLite with WAL mode handles every persistence requirement of a solo trading stack. Trade log, order book, strategy state, risk limits, model calibration tables — all SQLite. No separate database process to run, no cloud DB bill, no schema-migration service. It concurrently handles the trading loop and separate read-only monitoring queries without contention.

The limit: SQLite is a single-writer database. If the architecture requires multiple concurrent writers (multiple trading processes updating the same state), you need a different store. Most retail stacks can be architected around a single writer.

Dashboards: Cloudflare Pages free tier

A static HTML/JS dashboard generated nightly from the SQLite database, pushed to Cloudflare Pages. Free tier is 500 builds/month and unlimited requests. Pair with the free tier of Cloudflare Workers for a tiny proxy that reads from a read-replica (or a one-way-synced JSON snapshot) if you want semi-live updates.

For a private dashboard (just you), add a Cloudflare Access policy on the free tier and authenticate via email OTP. No auth code to write.

Broker execution: Alpaca paper, or IBKR/Tradier with BYO credentials

Alpaca's paper API is free and behaves like live. For real money on the $0 stack, Alpaca charges zero commissions (PFOF model — read the fine print on execution quality). Interactive Brokers charges low per-share commissions but no monthly data fees under typical retail activity, and Tradier is flat per-order on equities. None of these impose a monthly SaaS line item on the trader. See Broker APIs Compared (2026) for detail.

The free-tier landscape, April 2026

Vendor Free tier Hard limit that breaks $0
Alpaca Data Real-time IEX + free historical bars Full SIP consolidated quotes
Tiingo 1 year historical EOD + 500 calls/hr Tick-level, intraday > 1-min
Alpha Vantage EOD free, 25 calls/day Real-time, intraday at scale
Polygon Starter Historical bars, no real-time Real-time + Level 2
Databento Sample datasets only Any production workload
Binance / Bybit public Spot + futures REST + WS, generous limits OTC, archived tick data
Open-Meteo Weather, unlimited free N/A — stays free
FRED US macro, unlimited N/A — stays free
SEC EDGAR All filings, rate-limited free N/A — stays free
Yahoo Finance (unofficial) Everything, unreliable It breaks without notice

The Data-Vendor TCO Calculator quantifies per-symbol, per-request costs across the paid tiers when the free tier breaks.

The exact lines where $0 breaks

Line 1: Consolidated quotes (SIP)

Polygon's SIP tier starts at $199/month for historical + real-time consolidated quotes (April 2026 pricing). Databento is similar. If your strategy needs the true NBBO — not just the IEX print — this is where the free model ends.

When this breaks $0: intraday strategies on mid-cap and small-cap names, any market-making strategy, and anything latency-sensitive.

When $0 still works: large-cap-only strategies with 5+ minute holding periods, end-of-day strategies, rebalancing strategies, cross-sectional strategies where individual fill latency is dominated by portfolio construction.

Line 2: Level 2 depth

Full Level 2 (order book beyond top-of-book) starts around $500/month from Polygon, Databento, or CBOE direct. Alpaca does not offer L2 at any tier.

When this breaks $0: any order-book-imbalance strategy, any market-making attempt, any execution-optimization work.

When $0 still works: momentum, mean-reversion, and factor strategies that operate on bar data.

Line 3: 24/7 cloud compute

If the strategy must run when your local infrastructure (ISP, power, Mac Mini) is down, you need a VPS. Hetzner Cloud CX11 is around €4/month in 2026. DigitalOcean's smallest droplet is $6/month. AWS t4g.nano with a reserved instance is in the same range.

The $0 stack cannot reach this. If you need cloud compute, the floor is roughly €4-6/month.

When this breaks $0: strategies that trade through overnight or weekend market hours (crypto, forex), strategies where a one-hour outage is strategy-ending, strategies with regulatory uptime requirements.

When $0 still works: regular-market-hours equity strategies where a rare outage flattens positions at tomorrow's open and that's acceptable.

Line 4: Real-time options chains with Greeks

Tradier at $10/month provides real-time option chains including implied Greeks via their market-data bundle. Polygon, CBOE LiveVol, and ORATS are more expensive. Historical options data (needed for backtests) is a separate $300+/month spend from ORATS or equivalent.

When this breaks $0: any options strategy beyond theoretical exploration.

When $0 still works: equity-only strategies; options education using delayed chains.

Line 5: Historical tick data

A one-time purchase, not a subscription — so technically compatible with the "no monthly subscriptions" rule. Tick History from Databento, ICE, or Refinitiv ranges from $300 for a small symbol pack to $3,000+ for a broad universe.

When you need it: any strategy whose signal operates below the 1-minute bar — execution cost modeling, market-making backtests, HFT-adjacent work.

When $0 still works: everything at the 1-minute bar or above, which is where most retail strategies live.

A representative $0 stack

A concrete configuration that operates entirely within the free-tier envelope:

  • Mac Mini M4 16GB, macOS, on home fiber.
  • launchd plists scheduling: pre-market scan (08:30 local), intraday updater (every 15 min, 14:00-21:00 UTC during US hours), EOD reconciliation (21:30 UTC), weekly calibration (Sun 06:00 local).
  • Alpaca data + paper/live execution.
  • Python 3.12, SQLite for state, requests for HTTP, websocket-client for WS.
  • Ollama running Qwen2.5-32B for structured extraction; Claude BYO key for the 10-20% of steps that actually need frontier reasoning.
  • A nightly-generated static dashboard on Cloudflare Pages.
  • Telegram Bot API (free) for mobile alerts.

Recurring monthly cost: €0. Electricity: €1.50-2/month. Frontier LLM tokens: variable, typically $5-30/month, billed per-token as used. Domain + DNS: one annual fee if you want a vanity URL, otherwise free under .pages.dev.

When the $0 model ends

There are three patterns that graduate a stack to paid infrastructure:

  1. Signal requires data the free tier doesn't provide. You've validated the idea on Alpaca's IEX feed and the signal goes away when you add the rest of the market. Or the strategy is options-based. Pay for what the signal actually needs.
  2. Operational demands exceed single-machine reliability. You're running during hours your local infrastructure can't guarantee, or an outage is genuinely costly. Move the hot path to a VPS; keep research local.
  3. LLM workload shifts from research to inference-in-the-loop. If the strategy itself makes a Claude call on every bar, local models can't keep up on cost or latency. This is rare; most LLM-in-trading workloads are upstream of the execution loop.

None of these failures are reasons to have paid earlier. They're reasons to pay exactly when the constraint bites, and no sooner. The Trading System Blueprinter tool generates a stack diagram for a given strategy profile, including the expected free-tier/paid-tier split.

Connects to

References

  • Alpaca Data API documentation, IEX tier pricing (retrieved April 2026).
  • Polygon.io pricing, Starter through Advanced tiers (retrieved April 2026).
  • Databento pricing, tick and consolidated data (retrieved April 2026).
  • Tradier brokerage pricing, market data bundles (retrieved April 2026).
  • Cloudflare Pages and Workers free-tier limits (retrieved April 2026).
  • Apple launchd man page, launchd.plist(5) (macOS 15.x).