# Lantern MD

Find people and companies, research markets and products, draft and send email or LinkedIn outreach, track replies, and take notes at authorized meetings.

## User-facing conversation

Keep onboarding brief and in plain language. Do routine setup and connection checks quietly. Do not narrate reading this guide, inspecting the installer, downloading files, or running checks. Do not include MCP, CLI, OAuth, tool names, commands, flags, file paths, hashes, provider configuration, raw JSON, or a technical “Done so far” checklist in progress updates or the final reply. These implementation instructions are for the agent, not copy to show the user. Explain technical details only if the user explicitly asks for them.

- While setting up, at most say: “I'll set up Lantern MD and open sign-in for you.”
- If sign-in is pending, give one actionable prompt: “Finish [signing in](<login URL>), then return here. Confirm the connection code is <code>.” Use the actual validated login link and code; do not expose them anywhere else or invent them. Do not claim connection is complete while waiting.
- After login and verification succeed, say: “You're connected to Lantern MD. Try ‘Find five devtools founders’ or ‘Draft an introduction’. Drafts stay unsent until you ask to send.” Use at most three short sentences; do not paste command output or repeat the welcome.
- If setup fails, state the practical issue and the next action in one or two sentences. Keep credentials and internal diagnostics private. Disclose any incomplete step or unavailable feature that affects the user; do not claim success without verification.

## Agent setup instructions

Requires Node 22.12 or newer. No repository checkout, provider keys or database are downloaded.

1. Preserve an existing connection's origin. If already installed, inspect `lantern-md auth status` and log in only if needed; do not silently overwrite another connection. Otherwise download https://people.lantern.md/install.mjs to a temporary file and inspect it quietly. Run `node install.mjs --agent codex` (or `--agent claude` when that is the user's agent). Omit --agent for a standalone terminal. The installer downloads the client bundle, verifies its SHA-256 digest, and installs lantern-md and workflow skills. If its bin directory is not on PATH, invoke the installed client with Node at ~/.local/share/lantern-md/cli.mjs. Handle this internally; explain terminal configuration only if requested.
2. Complete browser login: verify email and continue directly into a personal workspace. There is no workspace chooser; existing team connections retain their access. No copied API key is required. Use --no-login to install first and log in later with `lantern-md login --api-url https://people.lantern.md`.
3. Verify with `lantern-md doctor`, which checks server health and stored capabilities only. Do not send, provision or run paid research as a connection check.
4. Follow the short connected reply above, with plain-language examples. The user should know what Lantern MD can do immediately.

## Agent reference: everyday use

`lantern-md discover "find people"` returns a few matching operations.
`lantern-md inspect find_people` returns its input schema.
`lantern-md run find_people --input '{"request":"founders","limit":5}'` executes it.
Use `lantern-md billing plans` and `lantern-md billing balance` for sandbox plans and credits. Pricing is pending; paid features require workspace configuration.
Use --input-file for JSON files, --output results.html or results.json for artifacts, and `lantern-md run runs_get --input '{"runId":"..."}'` to check asynchronous work.

After setup, users can simply ask their agent: “Find five devtools founders”, “Research this company”, “Draft an introduction”, or “Show new replies”. Drafts stay unsent until the user asks to send. Available providers depend on the workspace.

## Agent reference: optional MCP

MCP endpoint: https://people.lantern.md/mcp. Add --mcp with --agent during installation to register it, then complete the agent's native OAuth connection. Only discover, inspect and run are exposed. Every named operation in a skill or prompt is invoked through run({name,input}); inspect loads one schema on demand. CLI works without MCP.

On a fresh device, give the agent https://people.lantern.md/setup or say “Set up https://people.lantern.md/SKILL.md”. A bare product name cannot install an undiscovered tool.