janas-mcp

Expose a Janas-LLM model as MCP tools over stdio

TLDR

Run as an MCP server with a chat model

$ janas-mcp [models/qwen3-4b.jns]

Add an embedding model (enables the `embed` tool)

$ janas-mcp [models/qwen3-4b.jns] --embedding-model [models/qwen3-embedding-0.6b.jns]

Cap the answer length and turn reasoning off

$ janas-mcp [models/qwen3-4b.jns] --max [2048] --think [off]

$ claude mcp add janas -- /path/to/janas-mcp /path/to/model.jns

janas-mcp offers a model running on this machine to Model Context Protocol clients (Claude Code, editors, other agents) as tools over stdio. The client starts the process, writes JSON-RPC on standard input, and reads answers on standard output; everything else goes to standard error.Tools:- generate — ask the chat model a question (prompt, optional system message, longest answer, temperature) and return the answer- embed — return the vector of a text, only when --embedding-model is givenEvery call starts a conversation of its own. The reasoning of a model that reasons is never returned, only the answer. A ping is answered while a generation runs; `notifications/cancelled` stops it.The server answers both eras of the protocol: a request that names revision 2026-07-28 in `_meta` is served statelessly, and a client that opens with `initialize` gets the handshake of the revision it asks for (2025-11-25, 2025-06-18, 2025-03-26, or 2024-11-05).Configure it like any other stdio MCP server:

$ {"mcpServers": {"janas": {"command": "/path/to/janas-mcp",
"args": ["/path/to/qwen3-4b.jns"]}}}

At least one of a chat model or --embedding-model is required.

PARAMETERS

model.jns

Chat model for the generate tool.

--embedding-model file

A model that gives embeddings (Qwen3-Embedding). Enables the embed tool.

--mtp file

The chat model's multi-token prediction block.

--ctx tokens

Context length (default 16384).

--cache GiB

RAM for streamed experts (default: automatic).

--reserve GiB

Memory left to other programs.

--temp t

Default temperature (default 0.7).

--max tokens

Default longest answer (default 2048).

--think on|off

Reasoning before answering, for models that do it (default off; the reasoning is never returned).

--timeout s

Longest a generation may run (default 600; 0: no limit).

-h, --help

Print usage.

CAVEATS

This is an MCP server, not an interactive chat. Do not run it in a terminal expecting a prompt; a client must spawn it. The chat model and the embedding model are optional individually, but omitting both is an error. Same hardware and format limits as the rest of Janas: `.jns` files, x86-64 Linux or Windows, experimental, hard on RAM and disk.

HISTORY

janas-mcp ships with Janas, written by Maurizio "camauri" Cammalleri (first public release September 2026, GNU GPL 3 or later). It links Mbed TLS so `https://` MCP servers can be used from other Janas programs; this binary itself speaks stdio to its client.

RESOURCES

Source code · Documentation