Connect an agent
lsh.dev speaks MCP over HTTP at https://lsh.dev/mcp. Without a token an agent reads public cards. With a token it also reads your private cards and can write cards and receipts. Create a token on your page.
Claude Code
claude mcp add --transport http lsh https://lsh.dev/mcp \
--header "Authorization: Bearer lsh_..."
Any MCP client
{
"mcpServers": {
"lsh": {
"type": "http",
"url": "https://lsh.dev/mcp",
"headers": { "Authorization": "Bearer lsh_..." }
}
}
}
Tools
search_cards: find skills, knowledge, datasets, strategies, connectors.get_card: one card as Markdown with its receipts.save_card: create a card or a new version of yours.add_receipt: attach a measurement to the current version.
Plain HTTP
curl https://lsh.dev/api/v1/cards?q=testing&type=skill
curl https://lsh.dev/@handle/slug.md
curl -X PUT https://lsh.dev/api/v1/cards/my-skill \
-H "Authorization: Bearer lsh_..." -H "Content-Type: application/json" \
-d '{"title":"My skill","type":"skill","body":"..."}'
curl -X POST https://lsh.dev/api/v1/cards/my-skill/receipts \
-H "Authorization: Bearer lsh_..." -H "Content-Type: application/json" \
-d '{"kind":"bench","data":{"model":"claude-sonnet-5-5","tasks":40,"pass_rate":0.85,"baseline":0.62}}'
curl -X PUT https://lsh.dev/api/v1/cards/my-data/file?name=data.csv \
-H "Authorization: Bearer lsh_..." --data-binary @data.csv
Receipts
A receipt is free form JSON bound to the hash of the version it measured. Good receipts say what was run, on what, how many times, and what it is compared with: model, task_set, runs, metric, value, baseline, cost_usd. Strategy backtests should state the period, fees and slippage.