OpenRouter Fusion MCP server

OpenRouter Fusion MCP server

An Apify Actor running a Model Context Protocol server (Streamable HTTP transport) that exposes the OpenRouter Fusion Router as a single tool.

ask_fusion sends your prompt to a panel of frontier models in parallel; each panelist can search the web and cite sources, then a judge model fuses them into one answer plus a structured analysis — consensus, contradictions (attributed per model), unique insights, and blind spots. You get the fused answer, the cited source URLs, and every panelist's full response.

It reaches OpenRouter through the apify/openrouter proxy, so no OpenRouter API key is required — requests are authenticated and billed with your Apify token.

Why use it

Use it for hard, high-stakes, or ambiguous questions where one model is not enough: a single model can be confidently wrong, omit a key angle, or miss recent facts. Fusion runs several models, grounds them in live web sources, and surfaces exactly where they agree, disagree, and fall short — so you can trust the consensus and scrutinise the contradictions.

Tool: ask_fusion

Input	Type	Default	Description
prompt	string	— (required)	The question or task to deliberate on.
system	string	none	System instruction to steer the panel.
preset	"high" | "budget"	"high"	Panel preset. high = frontier models (best, pricey). budget = cheaper/faster panel with a frontier judge.
analysisModels	string[]	preset panel	Override the panel with explicit OpenRouter model slugs (≥ 2; fewer is ignored). Beats preset.
maxToolCalls	number	8 (OpenRouter default)	Max web search/fetch iterations per panelist (clamped 1–16). Lower = faster/cheaper, fewer citations.
web	boolean	true	Require the panel to web-search and cite source URLs for factual/current claims. Set false for pure reasoning with no grounding.

Presets

preset selects the panel that deliberates (the judge is always the frontier model anthropic/claude-opus-4.8). Override either with analysisModels.

Preset	Panel models	Use when
high (default)	~anthropic/claude-opus-latest, ~openai/gpt-latest, ~google/gemini-pro-latest	Best quality and most reliable citations. Pricey, slow (~$0.4–1.1/call).
budget	google/gemini-3.5-flash, deepseek/deepseek-v4-flash, openai/gpt-5.4-mini	Faster/cheaper. Weaker panelists cite less reliably (the judge still vets sources).

The ~…-latest slugs are OpenRouter floating aliases that always resolve to the current frontier model.

Output

ask_fusion returns the fused answer as text (with a Sources: list appended), plus structured fields:

Field	Description
answer	The judge's fused answer.
sources	Distinct source URLs from the judge's vetted answer. The judge drops citations it deems fabricated/unverifiable, so these are not the raw union of panel URLs (weak panel models can hallucinate plausible-looking URLs; those are filtered out). The full per-panel citations remain in panel.
analysis.consensus	Points all/most panelists agreed on.
analysis.contradictions	Topics where panelists disagreed, with each model's stance — plus an "evidence" stance that is the judge's own adjudication (not a panel model).
analysis.uniqueInsights	Points only one model raised.
analysis.blindSpots	Topics the panel under-addressed.
analysis.partialCoverage	Points some (not all) panelists made.
panel	Each panel model and its full response (where its own citations live).
model	The judge/router model that produced the fused answer.
usage	Token usage reported by the proxy.

Grounding & citations. With web: true (default), the panel is instructed to verify current facts/prices/versions via web search, cite the real source URL, and say "no source found" rather than guess. The judge then adjudicates the panel's citations — fabricated or unverifiable URLs (which weaker panel models can invent) are dropped from sources. For the most reliable citations, use preset: "high" (frontier panelists search and cite more reliably than the budget panel).

Example output

Prompt: "For tuning the hyperparameters of an expensive-to-train ML model, compare Bayesian optimization vs reinforcement learning … cite sources." (abridged)

{
  "answer": "# Bayesian Optimization vs. Reinforcement Learning …\n\nFor tuning a single expensive-to-train model, **Bayesian optimization (BO) is the right default** on sample efficiency, practicality, and tooling maturity. **RL is a specialized choice** … BO builds a probabilistic surrogate (GP or TPE) and finds strong configs in a few dozen to a few hundred trials ([Frazier, BO tutorial](https://arxiv.org/pdf/1807.02811)). Standard model-free RL is sample-hungry — Hyp-RL needed ~10M frames / 24 GPU-hours of pretraining ([Hyp-RL](https://ar5iv.labs.arxiv.org/html/1906.11527)) …",
  "sources": [
    "https://arxiv.org/pdf/1807.02811",
    "https://ar5iv.labs.arxiv.org/html/1906.11527",
    "https://arxiv.org/abs/1611.01578",
    "https://research.google/pubs/google-vizier-a-service-for-black-box-optimization/",
    "https://arxiv.org/pdf/1807.01774"
    // … 18 more
  ],
  "analysis": {
    "consensus": [
      "For tuning a single expensive model on a limited budget, Bayesian optimization is the better default.",
      "Standard model-free RL is sample-inefficient for one-off HPO; its cost only pays off when amortized across many tasks."
    ],
    "contradictions": [
      {
        "topic": "Has RL been 'replaced' by BO in neural architecture search?",
        "stances": [
          { "model": "~google/gemini-pro-latest", "stance": "A 2023 NAS survey says RL was outperformed and replaced by BO and evolutionary strategies." },
          { "model": "~anthropic/claude-opus-latest", "stance": "The field moved to one-shot/differentiable methods (DARTS), not specifically BO." },
          // "evidence" is the judge's own adjudication of the disagreement, not a panel model.
          { "model": "evidence", "stance": "In NAS, RL controllers were displaced mainly by one-shot/weight-sharing methods (ENAS, DARTS), not BO." }
        ]
      }
    ],
    "blindSpots": [
      "Population Based Training (PBT) — a leading method for dynamic schedules — is omitted.",
      "Parallel/async evaluation often dominates wall-clock efficiency more than raw sample count."
    ]
  },
  "panel": [
    { "model": "~anthropic/claude-opus-latest", "content": "# Bayesian Optimization vs. Reinforcement Learning …" },
    { "model": "~openai/gpt-latest",            "content": "## Short answer\nFor a single expensive-to-train ML model, pick Bayesian optimization …" },
    { "model": "~google/gemini-pro-latest",     "content": "When tuning the hyperparameters of an expensive-to-train model …" }
  ],
  "model": "anthropic/claude-4.8-opus-20260528",
  "usage": { "promptTokens": 15831, "completionTokens": 3103, "totalTokens": 18934 }
}

This Actor is an MCP server — the result is returned inline in the tool response (the structuredContent above). It does not write to an Apify dataset or key-value store, so there is no run output to download separately; consume it through your MCP client.

Endpoint

Deployed as a standby Actor, the MCP server is reachable at:

https://straightforward-under--fusion-mcp.apify.actor/mcp

Authenticate with your Apify API token as a Bearer token:

Authorization: Bearer <YOUR_APIFY_API_TOKEN>

Note: a fusion deliberation is slow (frontier panel + web search routinely runs ~100s+) and the proxy can occasionally return an incomplete body on long runs. The server streams keepalive notifications so MCP clients don't time out, and retries incomplete responses automatically. Use preset: "budget" and/or a low maxToolCalls for faster, cheaper runs.

Use with Claude Code

Add the server with claude mcp add over the Streamable HTTP transport. Replace YOUR_APIFY_API_TOKEN with your Apify token.

Project scope

Scoped to the current project and written to a .mcp.json file in the repo (shared with anyone who checks it out):

claude mcp add --transport http --scope project fusion \
  https://straightforward-under--fusion-mcp.apify.actor/mcp \
  --header "Authorization: Bearer YOUR_APIFY_API_TOKEN"

Security: --scope project stores the header in .mcp.json, which is committed to git. Do not hard-code a real token there. Claude Code expands environment variables in .mcp.json, so prefer a placeholder and let each user supply their own — set the header to Authorization: Bearer ${APIFY_TOKEN} (single-quote it in the shell so it is stored literally) and export APIFY_TOKEN in your environment.

User scope

Available across all your projects on this machine; stored in your user config, never in the repo:

claude mcp add --transport http --scope user fusion \
  https://straightforward-under--fusion-mcp.apify.actor/mcp \
  --header "Authorization: Bearer YOUR_APIFY_API_TOKEN"

After adding, restart Claude Code and run /mcp to confirm fusion is connected. Claude can then call the ask_fusion tool. Remove it any time with claude mcp remove fusion.

Running locally

npm install
APIFY_TOKEN=<your-apify-token> npm run start:dev

The server listens for MCP requests at http://localhost:3000/mcp. The token is needed locally because ask_fusion calls the apify/openrouter proxy; on the Apify platform it is injected automatically.

Deploying your own

Push the Actor to Apify with standby mode enabled:

$apify push

Your standby endpoint will be https://<username>--fusion-mcp.apify.actor/mcp.

Cost

This Actor does not charge a per-call fee. You pay for:

• Apify compute while the standby Actor is running, and
• OpenRouter model usage, billed by the apify/openrouter proxy against your Apify token.

A fusion call is much pricier than a normal model call — preset: "high" runs three frontier models plus a judge, each able to web-search. Expect roughly $0.4–$1.1 per call depending on prompt length and web usage; preset: "budget" is several times cheaper. To control cost, prefer budget, lower maxToolCalls, or override analysisModels with cheaper slugs.

A Pay Per Event schema is included in .actor/pay_per_event.json for anyone who wants to monetize a fork. It is not enabled on this Actor, and Actor.charge is not called.

Resources

• OpenRouter Fusion server tool
• OpenRouter web search & citations
• What is Anthropic's Model Context Protocol?
• How to use MCP with Apify Actors
• TypeScript MCP SDK examples
• Apify SDK documentation
• Apify MCP server documentation