OpenRouter Fusion — MCP server

A tiny MCP server that exposes OpenRouter Fusion (multi-model deliberation) to any MCP client (MCPHub, Claude Desktop, Claude Code, …) when you need a very good answer.

What is Fusion? It turns your prompt into a small multi-model deliberation: a panel of models answers in parallel (with web search/fetch), then a judge ("Fuse with") model synthesizes consensus, contradictions, unique insights and blind spots into one final answer. It beats any single model on hard questions — at roughly 4–5× the cost of a single completion. Docs: https://openrouter.ai/docs/guides/features/plugins/fusion · model page: https://openrouter.ai/openrouter/fusion

Tools (v2 — async only)

Tool	Role
fusion_list	List the available configs (name, label, panel, judge, reasoning_effort, temperature).
fusion_start	Start a deliberation in the background and return a job_id immediately (never times out).
fusion_result	Long-poll a job_id (~45 s/call) → the synthesized answer, or {status:"running"} to retry.

The old synchronous wrappers (fusion_quality / fusion_ultra / fusion_perso) were removed — sync calls were cut by client/proxy timeouts. Everything goes through fusion_start + fusion_result.

fusion_start({ prompt, preset:"quality" }) → { job_id }
fusion_result({ job_id }) → answer   (re-call while {status:"running"})

fusion_start forces the deliberation pipeline (tool_choice:"required"). Per-call overrides: preset, reasoning_effort, temperature, analysis_models, judge_model, system.

Configs (Quality / Budget / Custom)

Mirrors OpenRouter's "Model Fusion" UI tabs. Two are built-in in server.mjs; the rest are named custom configs, one per environment variable.

Config	Source	Panel	Judge	Reasoning
quality	built-in	Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro	Opus 4.8	high
budget	built-in	chosen by OpenRouter (general-budget)	OpenRouter	medium
maths, medecine, code, code-eco, perso	env var	see fusion-presets.json	—	high

Each config carries analysis_models (panel) · judge (orchestrator) · reasoning_effort (xhigh|high|medium|low|minimal|none, default high) · temperature (null = model default, or 0–2) · optional system.

Configuring — one env var per config

The server scans every OPENROUTER_FUSION_<NAME> env var and registers it as a config named <name> (lowercased; override with a name field — e.g. code-eco). The value is the JSON object {...}; the parser also tolerates a "name":{...} fragment or a {"name":{...}} wrapper.

OPENROUTER_FUSION_MATHS    = {"name":"maths","analysis_models":[...],"judge":"...","reasoning_effort":"high","temperature":null,"system":"..."}
OPENROUTER_FUSION_CODE     = {...}
OPENROUTER_FUSION_PERSO    = {...}

• fusion-presets.json — canonical, documented copy of all configs.
• fusion-env-vars.json — the exact compact values to paste into the mcphub env (one per config).
• Extra vars: OPENROUTER_API_KEY (an inference key with credit), OPENROUTER_FUSION_DEFAULT_REASONING.
• Legacy OPENROUTER_FUSION_PRESETS (single map) and OPENROUTER_FUSION_PERSO_CONFIG are still read for backward-compat.

Diagnostic: FUSION_DUMP_PRESETS=1 node server.mjs prints the resolved configs and exits.

Install

npm install

No build step — plain ESM (node server.mjs). Requires Node ≥ 18 (global fetch).

Set your key via the client's env, or copy .env.example → .env:

OPENROUTER_API_KEY=sk-or-...

Use with MCPHub

Add a stdio server pointing at the absolute path of server.mjs:

{
  "name": "openrouter-fusion",
  "config": {
    "type": "stdio",
    "command": "node",
    "args": ["/absolute/path/to/openrouter-fusion/server.mjs"],
    "env": {
      "OPENROUTER_API_KEY": "sk-or-xxxxxxxxxxxxxxxx",
      "OPENROUTER_FUSION_MATHS": "{...}",
      "OPENROUTER_FUSION_CODE": "{...}"
    }
  }
}

If node isn't on the launching process's PATH, replace "node" with an absolute path.

Use with Claude Code / Claude Desktop

Two ways, depending on whether you also want the interactive selector UX.

A) MCP server only

claude mcp add openrouter-fusion -e OPENROUTER_API_KEY=sk-or-... -- npx -y github:tboome33/openrouter-fusion-mcp

You get the 3 tools. The interactive selector behavior travels with the server via the tool descriptions — when you ask to use Fusion without naming a preset, the model is told to call fusion_list and ask you to choose preset → reasoning effort → temperature (it should not auto-pick). This is the only layer that also works on claude.ai / Cursor / other MCP clients.

B) As a Claude Code plugin (MCP + skill + /fusion command)

This repo is also a Claude Code plugin (.claude-plugin/plugin.json). Installing it bundles the MCP server (.mcp.json, run via npx), a model-invoked skill (skills/fusion-selector) that auto-triggers the selector whenever you want Fusion, and an explicit /fusion slash command.

/plugin marketplace add tboome33/openrouter-fusion-mcp
/plugin install openrouter-fusion@tboome33

Set OPENROUTER_API_KEY in your environment first (the plugin's .mcp.json reads ${OPENROUTER_API_KEY}).

Why a plugin? Slash commands and skills do not travel with a plain MCP install (only tool descriptions do). A plugin is what bundles the MCP server together with its command/skill so a user gets the full interactive UX in one install.

Optional — package as a one-click .mcpb bundle

npx @anthropic-ai/mcpb pack

The bundle prompts the user for their OpenRouter API key on install (user_config).

Examples

// start a quality deliberation
{ "prompt": "Explain the trade-offs between ridge, lasso and elastic net.", "preset": "quality" }

// a custom config with a per-call reasoning override
{ "prompt": "Review this clinical case and list differential diagnoses.", "preset": "medecine", "reasoning_effort": "xhigh" }

Notes

• Model slugs use OpenRouter syntax — pick valid slugs from https://openrouter.ai/models.
• Fusion is billed per the panel + judge it runs — quality (~$0.09 on a trivial probe) costs more than budget (~$0.04). Keep the heavy configs for the cases that justify them.
• A tool-set change only shows client-side after the connector reconnects (cached manifest); calls still hit the live server.
• Jobs live in memory only. If the server process restarts, in-flight job_ids become Unknown — just call fusion_start again. Finished jobs are pruned after ~30 min (and on a background timer).