About Loadout

Loadout is a competitive platform where engineers solve public coding challenges using AI tools, in the open. Anyone can pick up an open challenge, work it however they normally would — with whatever model, skills, MCP servers, and rules they already reach for — and submit the result.

A companion Claude Code plugin captures the full working session behind a submission: every prompt, every tool call, timings, cost, and the loadout that produced it. That captured session is called a trace, and it’s paired with a runnable artifact — the actual thing the session produced — to form a submission.

Once a challenge moves into voting, the community judges submissions blind and head-to-head: no names, no loadouts, no telemetry, until after a vote is cast. Those votes compound into public, per-category ratings, a ladder, and a profile for every competitor — a record of how you build, not just what you shipped.

Every submission on Loadout carries one of two labels. “Verified” means only this: the trace arrived from an authenticated capture-plugin session and was hashed on our server the moment it landed. That’s the entire claim — it isn’t an endorsement of the content, a claim about code quality, or a guarantee that the session is free of errors or omissions. It tells you where the upload came from and that we recorded a hash of it on receipt, nothing more.

“Self-reported” is the other label, and it’s just as honest: it means a submission was entered manually rather than captured by the plugin, so there’s no trace behind it and no hash to check it against. Neither label is a judgment on the work itself — they describe provenance, not quality.

Ratings come from blind, head-to-head votes in the Arena: two submissions, judged with identity and loadout hidden until after the vote is cast. Each vote updates both competitors’ ratings with a standard Elo model — everyone starts at 1500, and each result moves both ratings by a K-factor of 32 scaled to how surprising the outcome was, the same online update applied one vote at a time.

Ratings are tracked separately per challenge category, and a competitor’s tier (bronze through diamond) is just a threshold on that category’s rating. The stored ratings table is a cache, not a source of truth: it’s always defined as replaying every vote, in the order it was cast, from a clean 1500 start — so anyone can verify a rating is correct by recomputing it from the public vote log.

The capture plugin isn’t published to a marketplace yet — you install it from a clone of the Loadout repository. These are the exact steps, copied from packages/capture-plugin/README.md.

1. Build the plugin

pnpm install
pnpm --filter @loadout/capture-plugin build

2. In Claude Code, add the local marketplace

/plugin marketplace add /path/to/loadout

Replace /path/to/loadout with your clone path, e.g., ~/src/loadout.

3. Install the plugin

/plugin install loadout-capture@loadout-dev

4. Log in once per machine

/loadout-capture:login

This opens a device-flow sign-in against the Loadout server — approve it in the browser, and credentials are written to ~/.loadout/credentials.json (owner-only permissions, never logged, never put in a URL). You only need to do this once per machine.

5. Work normally, then submit

Every prompt you submit snapshots a checkpoint automatically — nothing about your normal workflow changes. When you’re ready to enter a challenge, run:

/loadout-capture:submit

This walks you through the redaction review below before anything leaves your machine.

Before a captured session ever reaches our servers, it goes through a local secrets and PII scanner, then a mandatory redaction review you have to complete — the plugin shows you a readable summary of the trace and every scanner hit (masked previews, never more of a secret than that), and lets you approve as-is or add manual redactions. Nothing is ever uploaded without that review step.

We scan again on our end when the trace arrives, independent of what the client reported, and reject anything that still looks like a secret — the server never accepts a trace with no redaction manifest attached. That said, the scanner catches patterns it recognizes; it isn’t a substitute for your own judgment about what you’re allowed to make public, and it can’t promise a session is free of anything sensitive.