FAQ

No. looot search and looot inspect read the public catalog without a token, and so does GET /v1/public-catalog. You need an account, a token and a funded balance to run anything. See Browse without an account.

No. A new workspace starts at $0. Add at least the minimum top-up (shown by looot balance) before your first run. See Before your first run.

Node.js 20 or newer. looot doctor checks it.

looot login saves it to ~/.config/looot/config.json, readable only by you. looot logout revokes it on the server and deletes the file. Every token is also listed in Settings, Agent tokens, on looot.ai, where you can revoke it. See Sign in.

Usually not. Claude Code, Claude.ai, Claude Desktop, ChatGPT, Codex, Cursor, VS Code, Gemini CLI, Windsurf and other MCP clients that support browser sign-in keep their own token, which renews on its own. You only put a token in a header for CI, scripts, or a client without browser sign-in. See Connect your agent.

No. There is no --token flag, so the secret never lands in shell history. Use looot login or set LOOOT_TOKEN.

looot inspect <endpoint-id> (or MCP inspect, or GET /v1/operations/{id}) shows the price formula and, when signed in, the estimated maximum cost of one run. Search results show the price per call or per result.

No. A run that fails at the provider settles at $0 (or the provider’s evidenced partial cost) and its hold is released. A run refused before it starts takes no hold at all. See Money: free calls and failures.

Not with the same idempotency key. The same key with the same input returns the original run (replayed: true) and charges once. The CLI prints the key it used on every run. See Idempotency.

A job (also called a capability, like people.email.find) is the task several providers can do, not one provider’s specific operation. Send endpointId: "job:<job id>" to run and looot picks the first provider of that job that can run now and accepts your input. You don’t pick an endpoint id yourself. See Jobs.

Add fallback to a run to let it move on to the next provider of the same job when the first one misses or errors. It’s one hold for the whole route, and only the attempts that actually ran are charged; an error, a 402, or a rejected call is never charged. See Fallback. --fallback on the CLI is coming in the next release; use the MCP run tool or the REST body today.

A completed run of some jobs (currently email verify/find, phone find, company enrich, scrape markdown and backlinks summary) can carry a normalized object next to result: the same answer mapped onto one shared field set, so you don’t have to read every provider’s own response shape. It’s absent, never null, when it doesn’t apply to that job or that gateway hasn’t turned it on. See Normalized output.

Some endpoints are switched off for a while, for example when the provider is down. They return provider_disabled, and search lists the endpoints that run now by default. Pick another endpoint for the same job, or send job:<id> and let looot pick one that can run.

runs_evidence / looot runs evidence <run-id> shows every attempt with the provider’s own response status, latency, receipt id and cost. Quote the run id and the requestId when you ask for help. Also check outcome: "miss" means the provider answered but found nothing, and "weak" means a flagged answer (a catch-all email, a guessed pattern) worth verifying before you use it.

Any MCP client that speaks Streamable HTTP works: add https://api.looot.ai/mcp and sign in with the browser. See Connect your agent for the setup page for Claude Code, Claude.ai, Claude Desktop, ChatGPT, Codex, Cursor, VS Code, Gemini CLI, Windsurf, Grok, Copilot Studio and more, and CLI vs MCP if you’d rather script against the CLI or REST directly.

The plugin adds ready-made skills for Claude Code and compatible clients on top of the MCP tools: setup, finding and running the right endpoint, money awareness, and troubleshooting. It is not required; the MCP tools alone are enough to search, inspect and run. See Plugin.

Its token lacks the runs.execute scope. If the agent signed in with the browser, revoke its “(connected)” token in Settings, Agent tokens, sign in again from the agent and keep runs.execute ticked on the consent page. If it uses a header token, create one that includes it, with looot login or in Settings, Agent tokens. See Access.

Ask for it with the capability_request MCP tool, or POST /v1/capability-requests.

Pipe the command, or add --format json (or jsonl). Errors come as one JSON object on stderr; see Errors.

Yes. https://docs.looot.ai/llms.txt is a plain-text index of this site, and every page also has a .md mirror (append .md to the page’s URL) for an agent that would rather read Markdown than rendered HTML. There’s also a docs MCP server at https://docs.looot.ai/mcp an agent can connect to and search directly. Separately, the gateway itself publishes https://api.looot.ai/llms.txt describing the API and tools, and each catalog endpoint has an agent-readable card at /catalog/{endpointId}.md.