Larvyn MCP server

For agents that prefer MCP over a shell, larvyn mcp serves every larvyn command as an MCP tool, on stdin and stdout. The tools are the commands with an underscore: terminal read is terminal_read, agents wait is agents_wait, and so on.

Add it to your agent

Run these in a Larvyn terminal, where larvyn is on the PATH.

Claude Code

claude mcp add larvyn -- larvyn mcp

Codex

codex mcp add larvyn -- larvyn mcp

Cursor, Windsurf and other clients with an mcpServers file (.cursor/mcp.json in Cursor):

{
  "mcpServers": {
    "larvyn": { "command": "larvyn", "args": ["mcp"] }
  }
}

OpenCode (opencode.json):

{
  "mcp": {
    "larvyn": { "type": "local", "command": ["larvyn", "mcp"] }
  }
}

An agent you start from Larvyn finds larvyn on its PATH. A client started anywhere else needs the full path to the binary, in the bin folder of Larvyn’s state folder: ~/.config/larvyn/bin/larvyn on Linux, ~/Library/Application Support/larvyn/bin/larvyn on macOS and %APPDATA%\larvyn\bin\larvyn.exe on Windows. There it works from a project’s folder or one of its worktrees, while the app runs.

Tools

Tools that only read are marked with readOnlyHint, so clients can run them without asking. A long agents_wait or stack_wait doesn’t block the other calls.

Tool	Reads only	What it does
context	yes	Project, worktree, branch, local URL and dev stack.
version	yes	Larvyn’s version.
ports	yes	The ports the project’s terminals and services listen on, with their URL.
terminal_list	yes	The project’s terminals.
terminal_read	yes	A terminal’s last lines, or what its screen shows.
terminal_search	yes	Matching lines of a terminal’s whole output.
terminal_write		Types text into a terminal.
terminal_open		Opens a terminal running a command or a shell.
terminal_close		Closes a terminal the caller opened.
stack_list	yes	The dev stack and what runs.
stack_wait	yes	Waits until a stack item is ready.
stack_start, stack_stop, stack_restart		Starts, stops or restarts a stack item, optionally waiting until it is ready.
artisan		Runs an Artisan command that finishes.
test		Runs the tests in a terminal Larvyn shows and returns the failures.
logs_errors	yes	The latest errors in the Laravel log.
models	yes	Eloquent models with table, relationships, fillable, casts and row count.
db_tables, db_table	yes	Tables with row counts; one table’s columns, indexes and foreign keys.
db_query	yes	A query that reads, in a read-only transaction that is rolled back.
agents_kinds, agents_list, agents_status	yes	The agents Larvyn can start, the project’s agents, one agent’s status and screen.
agents_wait	yes	Waits for one or several agents to be done or need attention.
agents_start, agents_send		Starts an agent in a new worktree; sends it a follow-up.
lock_list	yes	The locks held in the project.
lock_acquire, lock_release		Takes turns on what agents share, such as the database.
specs_list, specs_show	yes	Specs and their reviews.
specs_draft, specs_review, specs_revise, specs_build, specs_drift, specs_fix		The Spec first workflow.
agents_setup, agents_remove		Installs or removes the larvyn skill and session hook.

Each tool’s parameters are the command’s options, with the same names and defaults. Your client shows them, or see the command reference.

Security

larvyn mcp talks to the app through a socket only your user can open, with a token that changes at every start. It never opens a port, so a web page in your browser can’t reach it. Agents get the permissions you set in Larvyn: an agent started by another agent uses your current permission setting, and agents can start agents at most two levels below the ones you start.