Use with AI

Publish an MCP Server

Give an MCP server on your machine a public URL with one command. The link is the key, and a share page lists the tools with setup steps for five clients.

Docs for LLMs

otterkit mcp publish gives an MCP server on your machine a public URL. The server keeps running locally and OtterKit never hosts your code. Use it to test a server you are building against Claude on the web, to reach your own tools from another device, or to let a teammate try a server before you deploy it.

Terminal
# A script: the runner is picked from the extension (.js, .ts, .py)
npx otterkit mcp publish ./server.js

# Any command that starts a stdio MCP server
npx otterkit mcp publish npx -y @modelcontextprotocol/server-everything

# A server that already speaks HTTP locally
npx otterkit mcp publish http://localhost:3000/mcp

The command starts your server, lists its tools, and prints one link such as https://tunnel-a1b2c3d4.otterkit.app/mcp/k7Qf3xZ9…. Paste that link into any MCP client that accepts a remote server. Opened in a browser, the same link shows a share page with the tool list and setup steps for Claude, ChatGPT, Cursor, VS Code and Claude Code.

The link is the key#

There is no sign-in. The link ends in 22 random characters, and anyone who has the full link can call the published tools, so treat it like a password and share it only with people you trust. Requests without the right key get a 404 and never reach your server.

If a link leaks, press r in the terminal to replace it. The old link stops working immediately. Press q or Ctrl+C to stop, which takes the server offline.

Choosing which tools are public#

A tool that your server marks as destructive (destructiveHint: true) is held back by default, and the command tells you which ones. Callers cannot see or call a tool that is held back.

FlagMeaning
--only <tools>Publish only these tools (comma-separated).
--exclude <tools>Publish everything except these tools.
--all-toolsAlso publish tools marked as destructive.
--name <name>Name on the share page. Default: the server's own name.
--description <text>One line shown under the name on the share page.
--subdomain <name>Use a stable subdomain. The link then stays the same across restarts.
--rotateWith --subdomain: replace the saved link with a new one.
--ttl <duration>Auto-stop after e.g. 4h or 3d, or "never". Default: 24h.
--no-recordDo not keep a local log of tool calls.
--daemonRun in the background and return once the link is live.
--jsonWith --daemon: print the link, tools and PID as one JSON line.

Flags for publish go before the target. Everything after the target is passed to your server unchanged.

Run it in the background#

With --daemon the command starts the server as a background process, waits until the link is live, prints it and returns. The server keeps running after you close the terminal, until its --ttl passes or you stop it. Its output, including the request feed, goes to a log file whose path the command prints.

Terminal
# Start in the background on a fixed address
npx otterkit mcp publish --daemon --subdomain my-tools ./server.js

# See what is running, with each server's link
npx otterkit status

# Stop it
npx otterkit stop my-tools

Add --json for scripts and agents: the command prints one JSON line with url, tools, held, pid and logPath. To replace the link of a background server, stop it and start it again with --rotate.

Record and replay tool calls#

Every tool call that reaches a published server is recorded with its arguments, its result, the client and how long it took. Use the recording to see exactly what an agent sent, and to run a call again after you change the tool, without going back to the client and prompting until it calls the tool.

Terminal
# List recorded calls for the most recent server (or name one)
npx otterkit mcp calls
npx otterkit mcp calls my-tools --tool run_query

# Arguments and result of one call
npx otterkit mcp calls my-tools --show 12

# Run call 12 again and compare the result with the recording
npx otterkit mcp replay my-tools 12

# Same tool, different arguments
npx otterkit mcp replay my-tools 12 --args '{"table":"orders"}'

A replay starts the server again from the command it was published with, so it runs your code as it is now, then shows whether the result is the same as the recording or what changed. It opens no link and costs no credits, and it works after the published server has stopped. Add --json to either command for scripts, and --url to replay against a server that is already running.

Limits#

LimitDetail
Tool call timeA call must finish within 28 seconds.
Server-initiated messagesProgress updates, sampling and list-changed notifications are not relayed. Tools, prompts and resources work.
UptimeThe server is reachable only while the command runs and the machine is awake.
PriceThe same as a tunnel: 1 credit per connected hour.