Welding every vertical task into the core costs effort A ˣ for return X — complexity compounds, payoff crawls. A harness flips it: each task becomes an isolated workspace a launcher spawns and schedules. Effort X, return A ˣ.
The old path turned every job — monitor a symbol, characterize a strategy, write a report — into another workflow welded into the software. The deeper the stack got, the more a trading agent spent its days fixing a brittle chat shim that had nothing to do with trading. So we flipped the architecture: each job is now a harness, the core stays still, and the gains pile onto one rack instead of one monolith.
effort
Xreturn
effort
A ˣreturn
A harness is a directory + git repo + native CLI session. It can't entangle the launcher's core, so adding the hundredth harness is as cheap as the first. Difficulty grows linearly.
Harnesses don't know about each other. The launcher composes them — run twenty at once, chain their outputs, react to events. Capability compounds where the difficulty doesn't. Return grows exponentially.
Heavy, cross-cutting context is the rack's job, not the harness's. K-lines for every asset class shouldn't be a skill you torture an AI with — standardize it once, hand it to every workspace.
It does two things: manage and schedule harnesses, and inject the shared context and tools they need. Specialize the rack by domain — the same registry feeds every one of them.
Your one-person Wall Street.
A launcher specialized for trading harnesses. It maintains the heavy, cross-cutting substrate itself — cross-asset market data, charting, a unified trading account over many brokers — so every harness on top stays light. The AI never re-derives K-lines or re-implements a broker; it just trades.
The launcher, minus the trading desk.
The same launch-rack, with no trading substrate baked in. For harnesses that have nothing to do with markets — games, decks, docs, research loops. Bring your own context; NanoAlice schedules the workspaces and gets out of the way.
Most harnesses are self-contained — they run in any coding-agent sandbox: a launcher, or Claude Code / Codex (web included). The trading harnesses are the exception — they bind to OpenAlice for its home market-data supply.
A workspace wired to the full trading tool surface.
A chat workspace plumbed into OpenAlice's MCP server — the entire trading and market-data tool surface is available to the agent. The harness that used to be a brittle SDK shim is now just… another harness.
An autoresearch coworker for trading strategies.
Spawns as a local Auto-Quant clone on its own autoresearch/<tag> branch. The agent runs the prepare / backtest / optimize pipeline, edits strategies, logs results to results.tsv, and commits as it goes — every iteration leaves an auditable git trail. Each workspace gets isolated data so runs never mix. This was the first experiment that proved a vertical scenario could be swallowed whole by a harness.
Market data, valuation, earnings & sentiment skills, ready to go.
Bundles finance-skills (yfinance market data, valuation, earnings, social readers, sentiment) into a fresh workspace. SKILL.md files land in project-local .claude/skills/ and .agents/skills/, discovered natively by Claude Code and Codex. Cloned fresh per workspace; your global state stays untouched.
A headless RPG Maker for GalGame-shaped games.
An AI-first engine where a game is a folder of markdown scripts and YAML. Pure step(Game, State, Input) → (State, Output) — no graphics framework. Play it in a terminal, over a JSON API, or via built-in AI personas. Live-authoring reloads as you edit. A harness for a domain that has nothing to do with trading.
A slide deck is a folder of React components.
Make presentations by editing .tsx files and let your coding agent do the writing. Fixed 1920×1080 canvas scaled to the viewport, single-slide preview for focused editing, a headless CLI for content management. Minimal layout primitives, on purpose — every design choice is deliberate.
The registry is a Next.js route handler — the same data this page renders, served as JSON. A launcher fetches it to discover every harness, which rack each one targets, and where to pull it from. CORS-open, uncached, always live.
$ curl https://harness.openalice.ai/api/registry { "schema": "harness-registry/v1", "registry": "harness.openalice.ai", "counts": { "launchers": 2, "harnesses": 5 }, "harnesses": [ { "id": "auto-quant", "agents": ["claude", "codex"], "runtime": { "portable": false, "requires": ["OpenAlice market-data supply"] }, "source": { "kind": "standalone" } }, … ] }
