naulon

@naulon/wayfarer-mcp

A Model Context Protocol server that lets any LLM discover, quote, pay, and cite naulon-tolled sources. Bring your own wallet, local stdio, the wayfarer brain running in-process.

Point any MCP-capable client at this server and the model gains tools to find tolled articles, get a quote, pay the 402 toll from a wallet you control, and cite what it bought: the same budgeted buying loop the CLI agent runs, exposed as callable tools and slash commands.

Works with any MCP client: Claude Code, Claude Desktop, Cursor, Windsurf, Cline, VS Code, OpenCode, Codex, Gemini CLI, or your own host. Setup for each is below.

Listed in the MCP Registry as app.naulon/naulon.

Quick start

npx -y @naulon/wayfarer-mcp        # runs the stdio MCP server

The package is scoped (@naulon/wayfarer-mcp); the binary it installs is wayfarer-mcp. Zero-config, it starts and mocks settlement (no wallet, no spend), but mock settlement is not a mock catalog: discover, research and verify need a real place to look, so with nothing configured they refuse (they never fabricate sources). To actually find and read something, pick one:

• Buy from the fleet (hosted): point a client at the hosted endpoint with an agent token and use naulon_ask, which buys across the whole naulon fleet directory, real corpus, no local wallet. See Hosted endpoint.
• Single publisher (stdio): set RSS_URL, PUBLISHER_URL, or CATALOG_URL to that site's feed/catalog, plus TOLLGATE_URL to pay. A lone TOLLGATE_URL pins every payment to one gate, which is single-publisher by design.
• Fleet / many publishers (stdio): to buy across more than one publisher locally, set WAYFARER_ALLOW_DOMAINS to the publisher hosts you trust (and point discovery at a fleet directory via CATALOG_URL) instead of pinning a single TOLLGATE_URL, and each discovered publisher is then paid at its own gate, bounded by the allowlist. See Configuration.
• Full loop offline: run the repo's make demo (or make dev) to drive the whole discover → toll → pay → settle loop against a local stub publisher.

The canonical registration, which every client below is a variant of:

{

"mcpServers": { "naulon": { "command": "npx", "args": ["-y", "@naulon/wayfarer-mcp"] } } }

Slash commands (prompts)

Every prompts-capable client surfaces these as native, argument-taking slash commands, with no per-user config. In Claude Code and Desktop they appear as /mcp__naulon__ (the naulon segment is whatever you named the server):

\* ask is only present on the hosted endpoint (it drives the cloud naulon_ask tool). The stdio server exposes research / discover / verify.

Per-client setup

Claude Code

CLI (recommended, since --scope project writes a shared .mcp.json and user makes it global across your projects):

claude mcp add naulon --scope user -- npx -y @naulon/wayfarer-mcp

Then /mcp inside Claude Code to confirm it connected, and type / to see the research / discover / verify prompts. Or add it by hand to .mcp.json (project) using the canonical block above.

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{

"mcpServers": { "naulon": { "command": "npx", "args": ["-y", "@naulon/wayfarer-mcp"] } } }

Restart Claude Desktop. Prompts appear in the + / slash-command menu.

Cursor

~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project), key mcpServers, the same canonical block as above.

Windsurf

~/.codeium/windsurf/mcp_config.json, key mcpServers, the same canonical block.

VS Code (native MCP / Copilot agent)

.vscode/mcp.json (workspace). VS Code uses the servers key (not mcpServers); the type is inferred as stdio from command, so it's optional:

{

"servers": { "naulon": { "command": "npx", "args": ["-y", "@naulon/wayfarer-mcp"] } } }

Cline

Cline → MCP Servers → Configure (cline_mcp_settings.json), key mcpServers, same canonical block.

OpenCode

opencode.json in your project root. OpenCode uses the mcp key, a type of local, and the command as one array:

{

"$schema": "https://opencode.ai/config.json", "mcp": { "naulon": { "type": "local", "command": ["npx", "-y", "@naulon/wayfarer-mcp"] } } }

Environment variables go under environment, not env.

Codex

~/.codex/config.toml:

[mcp_servers.naulon]

command = "npx" args = ["-y", "@naulon/wayfarer-mcp"]

codex mcp list shows it as enabled.

Gemini CLI

~/.gemini/settings.json (global) or .gemini/settings.json (per-project), key mcpServers, the same canonical block. gemini mcp list shows it connected.

Any other MCP host

Spawn the stdio binary and speak MCP over its stdio transport:

command: npx   args: ["-y", "@naulon/wayfarer-mcp"]

Hosted endpoint (no local wallet)

The hosted naulon service exposes the same brain over Streamable HTTP at /_naulon/mcp, authenticated with an agent token. Tolls are signed by naulon's custody-free session key, so no private key ever touches your machine. This endpoint also adds the hosted-only naulon_ask tool + its ask prompt.

Clients that support remote/HTTP MCP with headers:

# Claude Code. `claude mcp add` refuses a name that exists, so clear it first;

the same two lines then work again after you rotate the token.

claude mcp remove naulon -s local 2>/dev/null; claude mcp remove naulon -s user 2>/dev/null claude mcp add --scope user --transport http naulon \ https:// /_naulon/mcp \ --header "Authorization: Bearer "

// Generic HTTP MCP config

{ "mcpServers": { "naulon": { "type": "http", "url": "https:// /_naulon/mcp", "headers": { "Authorization": "Bearer " } } } }

// OpenCode (opencode.json)

{ "$schema": "https://opencode.ai/config.json", "mcp": { "naulon": { "type": "remote", "url": "https:// /_naulon/mcp", "headers": { "Authorization": "Bearer " } } } }

Mint the agent token from your naulon buyer wallet / dashboard. Spend is bounded by the server budget and the token's sub-cap, so the model can lower a run's budget, never raise it past either.

Configuration

Env read by the stdio server (all optional, so omit for the offline mock):

Budget and wallet are server config, never tool arguments, so the model plans spend within the envelope but can't widen it.

Tools

\* hosted endpoint only. Every tool states readOnlyHint, destructiveHint and openWorldHint. The free ones are read-only; the ones that spend are destructive, because a payment cannot be taken back, so a client that confirms irreversible actions asks first.

MIT.