The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Strategy Lab listing page.
kellyruns.xyz · how it works · @KellyRunsIt
Most people learn leverage, liquidation and impermanent loss by losing money to them. The lab is the cheaper way round.
You submit a lowcap, perp or LP strategy as parameters, not prose. The lab stress-tests it across thousands of simulated paths and publishes the whole distribution — not one backtest number that hides the tail where the damage lives.
| Certification gate | ≥ 3 runs and ≥ 1,500 paths |
| How to size | against the 5th percentile the run prints, never the median |
| What a pass means | a statement about a distribution — not a prediction |
| Who pays | trading fees cover the $0.85 of compute per run. Not you. |
Kelly speaks MCP, so another agent can read the record without scraping HTML — and the reproduction seed comes with it.
Stateless streamable HTTP, protocol 2025-06-18, no auth and no key. Published
to the official registry as
xyz.kellyruns/strategy-lab.
Read-only by design. No tool here submits a strategy, funds a run, or moves anything — writes belong behind a wallet session, and an unauthenticated endpoint is the wrong place to re-implement one.
| Tool | What it returns |
|---|---|
get_agent_state | Treasury, open positions, recent trades, and the reasoning behind the latest cycle. Paper mode — a published track record, not a portfolio. |
get_certification_gate | The thresholds a strategy must clear, and what a pass does and does not mean. Read this first. |
get_lab_summary | Counts by status, testing-pool balance, and how many runs it can still fund. |
list_strategies | Strategies filtered by kind and status, each with its certification rollup — runs, paths, median PnL, worst case, ruin rate. |
get_strategy | One strategy in full: parameters, verdict, every check the gate ran, and each run's PRNG seed. |
The seed is the point. Re-running with it reproduces the result exactly, which is what makes a verdict something you can disagree with by running it rather than by arguing about it.
The Kelly Criterion (John L. Kelly Jr., Bell Labs, 1956) answers the question retail traders skip: not is this a good bet but how much should I bet. That is the direct ancestor of the risk clamp that gets the last word on sizing here, over anything the model proposed.
Which is also why the agent is a character and not a dashboard. Kelly says no
is a sentence. A tool can't be the subject of one.
Up front, because a project about publishing losses cannot be coy about its own readiness.
| Chain | Robinhood Chain — 4663, Arbitrum stack, ETH gas |
| ERC-8004 Identity Registry | 0x8004A169…a432 |
| Agent registration | /.well-known/agent-registration.json |
TypeScript · Next.js · viem · Postgres · Anthropic SDK
The art above is generated by scripts/build-art.mjs —
two self-contained SVGs, no external fonts and no remote assets, because
GitHub's image proxy blocks every outbound fetch. No third-party widget
services either, so nothing here breaks when someone else's server goes down.
Built on Robinhood Chain. Not affiliated with, endorsed by, sponsored by, or partnered with Robinhood Markets, Inc. The chain is permissionless — anyone can deploy on it without approval, and that includes this project.