For agents
Check before you submit. Every time.
An agent that hunts bugs should not file a finding its own evidence cannot carry. Give it one tool, basanos_precheck, and it learns what proof is missing before a judge has to tell it.
git clone https://github.com/Cassxbt/basanos && cd basanos && npm ci
claude mcp add basanos \
--env SERV_API_KEY=… --env DOLPHIN_API_KEY=… \
-- npx tsx "$PWD/src/mcp.ts"
{
"mcpServers": {
"basanos": {
"command": "npx",
"args": ["tsx", "/path/to/basanos/src/mcp.ts"],
"env": { "SERV_API_KEY": "…", "DOLPHIN_API_KEY": "…" }
}
}
}
npx tsx src/cli.ts finding.md --contest karak
# reads keys from .env · exit 0 only for READY that routes automatically · 1 anything else · 2 usage · 3 failed
npx tsx src/cli.ts finding.md --json | jq .route
curl -s https://basanos-nu.vercel.app/api/precheck \
-H 'content-type: application/json' \
-d '{"contest":"karak","title":"…","report":"…"}'
What the agent gets back
verdictREADY · NEEDS_PROOF · BLOCKED_BY_CODE · KNOWN_OR_OUT_OF_SCOPE · CITATION_BROKEN · GUARD_HELD
routeautomatic only when three passes agree and no instruction-like text was found
missingsuggested checks, each with what would resolve it
receiptchecksummed JSON from every interface: finding digest, audited commit, votes, SERV request ids, cost
A real CLI run on Karak #93, recorded from the shipped pipeline.