by Synenergy
Consilium Router
A multi-model AI panel for the questions that matter. A plural answer, with evidence.
For open judgement calls — several models answering independently. For a stack of files — cited answers with the cost in view. One entry point — API or MCP — and a provider aggregator.
Available via catalogs:
OpenRouter
Hugging Face
- OpenAI
- Anthropic
- DeepSeek
- Qwen
- xAI
Blind spots — and context that does not fit.
One model means one perspective and one set of hallucinations. And when the real work is a stack of specs, logs, and notes, cramming it all into one frontier chat wastes money and still loses the thread. Pick the Consilium mode that fits your job.
What it’s used for
Your documents — big context, controlled cost.
You (or your assistant) collect the files — specs, notes, logs, scans, database exports, pages you already saved. Consilium works through that set instead of pasting everything into one expensive model: answers with quotes from those files, checks that the quotes really appear there, and shows what the run cost. Cheaper helper models can prep messy material first. It does not search the internet for you.
A second opinion — model pluralism.
Several AI models — different families, different blind spots — answer the same open question. That pluralism is the point: you see where they agree and where they clash. More models does not mean more truth — check important facts yourself.
Audits and reviews.
Architecture, security, contracts, strategy — put it in front of the panel and each model hunts for weak spots independently. What one misses, another finds.
One API and a provider aggregator.
Consilium routes to OpenRouter and Hugging Face for you: one base_url, one bill, models from both catalogs in solo roles and in panels. Your own provider key is optional.
Code verification.
Pre-merge code review, bug hunting, risk assessment of tricky changes. Independent verdicts with no groupthink: no model sees what the others said.
High-stakes decisions.
Business models, finances, investments, choosing a strategy or a stack — when mistakes are expensive, convene a council instead of trusting a single opinion.
Modes
Your documents (Gather & Judge).
When the answer must come from your files — and the set is too large for one prompt. Consilium works the context in steps: real quotes from your files, a check that those quotes are really there, and a clear cost. Use cheaper helper models for cleanup so the main assistant does not burn frontier tokens on busywork. Not the multi-model panel. A passed quote-check means the quote is in your files — not that the conclusion is true worldwide.
Best for: large file sets, contradictions, “what do these files actually say — with quotes?”
Compare.
Several models answer the same prompt independently, side by side.
Best for: a quick sanity check — do the models agree, or is one hallucinating?
Deliberate.
The panel answers, a judge distills the strongest parts into one final answer.
Best for: everyday work — one reliable answer instead of several raw ones.
Vote.
Submit your candidate — each model returns an independent verdict.
Best for: validating a finished decision, text, or code before release.
Brainstorm.
Models generate ideas blind; the strongest get recombined. Treat claims as unverified until you check them yourself.
Best for: finding new ideas when you need a fan of options, not one opinion.
Extended — West vs East.
Top models of the Western and Eastern schools on one panel, through two independent channels.
Best for: decisions where shared blind spots are dangerous.
Council.
A structured panel session for your hardest calls.
Best for: architecture, strategy, naming — decisions that are expensive to reverse.
Route Lab.
Try which models and routes fit a class of tasks before you hard-code a default into an agent. Free helpers discover and propose; paid experiments run real routes and rank them. Not everyday chat.
Best for: picking a default for agents or a panel recipe.
Why Consilium
Model pluralism.
Not one school of thought — a deliberately mixed panel so shared blind spots are harder to hide.
Context without the dump.
Large file sets are worked in steps — with quotes and a quote-check — instead of pasting everything into one frontier chat.
Cost you can see.
Document runs show the bill. Cheap helpers prep messy material; expensive models stay for judgment, not busywork.
Works with your tools.
Plugs in like a regular OpenAI model: two lines in your settings — and your chat client, IDE, or script consults the panel.
Built for agents.
Connect once: “Your documents” for big context with a visible cost, or the panel for a plural second opinion.
Flexible onboarding.
Demo right away, no keys needed. Then bring your own provider key — or run everything through us.
Confidentiality mode.
On commercial plans: no prompt storage, no training on your data.
Fails loudly, not silently.
A failing model is visible immediately — and the rest of the panel still delivers.
How it works
- Send one requestyour prompt and how wide a panel you want.
- The panel worksindependent answers, cross-examination, fusion.
- You get the verdictone answer, with disagreements surfaced, not hidden.
Consilium as a provider for agents
Point Cursor, Claude, Codex, Hermes, or OpenClaw at Consilium — “Your documents” when the pile will not fit one prompt (quotes + cost control), a panel for a plural second opinion, or one model for everyday work. Broader is not always more accurate.
Subagent roles on different LLMs.
Architect, skeptic, implementer — different models through one gateway, or a shared-prompt compare when they share the brief.
Solo for chores, panel for judgment.
Routine edits stay on one model. Escalate to a small panel or deliberate only for open design, naming, and trade-offs.
Vote before merge.
Finished candidate + rubric → independent accept / reject / abstain. A gate signal — not a fused essay, not a substitute for tests.
Disagreement → escalate.
When the panel contradicts itself on an expensive call, widen the panel, vote a concrete candidate, or ask a human — do not fake consensus.
Drop-in second opinion.
Set base_url to Consilium /v1 and/or wire MCP. Leave the rest of the agent stack as-is.
Your documents.
When the pile will not fit one prompt: Consilium answers with quotes from those files, shows the cost, and can use cheaper helpers for prep — so the main model is not a dump truck for context. Keep the panel for open judgment.
Route Lab before hard-coding.
Research which route or solo default fits a task class before you ship the agent config. Free helpers first; paid experiments only with explicit approve — not for every user chat.
Route Lab
Research which models and routes fit a class of tasks — before you hard-code a default. Free: discover, propose, stub-benchmark. Paid: route_lab_run_experiment (approve spend). Not a substitute for compare, deliberate, vote, or “Your documents.”
Agent model picks
When you set base_url to Consilium /v1, start from these solo roles — not a single “best model.” Vendor agent claims (including Meta on Llama tool use) are hypotheses until we publish our own harness ratings. Panel fuse is separate: panel=diversified* via MCP.
| Role | OpenRouter | Hugging Face |
|---|---|---|
| Controller | openai/gpt-6.1-sol$2 / $10 | deepseek-ai/DeepSeek-V4-Pro-0813$1.32 / $3.96 |
| Planner | anthropic/claude-opus-5.5$4 / $20 | Qwen/Qwen3.8-2.4T-A95B$2 / $6 |
| Implementer | anthropic/claude-sonnet-5.5$2 / $10 | deepseek-ai/DeepSeek-V4-Pro-0813$1.32 / $3.96 |
| Skeptic | x-ai/grok-4.7$2 / $6 | zai-org/GLM-5.3$1.4 / $4.4 |
| Cheap worker | openai/gpt-oss-120b$0.037 / $0.17 | zai-org/GLM-5.3-Flash$0.15 / $0.5 |
| Fast / light | openai/gpt-6-luna$0.1 / $0.5xiaomi/mimo-v2.6-flash$0.14 / $0.28 | XiaomiMiMo/MiMo-V2.6-Flash-RL$0.14 / $0.28 |
Top picks as of 5 Oct 2026 · prices from the providers’ public catalogs, per 1M tokens in/out · picks are advisory
Subagents: architect → Claude Opus 5.5 · implementer → Claude Sonnet 5.5 · skeptic → Grok 4.7 · orchestrator → GPT-6.1 Sol.
Connect / verify
Already have a Consilium service Bearer from an operator? Download the auth smoke script, then run it locally — it checks /health, /v1/models, and a tiny non-stream chat. It does not create tokens and never prints yours.
Prefer download-then-run. Optional pipe: curl -fsSL …/connect.sh | bash -s -- "$CONSILIUM_TOKEN".
Hermes / OpenAI drop-in: base_url=https://consilium.synenergy.ai/v1. Default agent solo: model=meta-llama/llama-4-maverick (see agent model picks).
- OpenRouter
- Hugging Face
Providers behind Consilium: requests go through the top catalogs.
For assistants. Panel = second opinion on a question. Your documents = answers with quotes from files you provide (same connection; after an update, reconnect once so new tools appear). Route Lab = research a default route for a class of tasks before you hard-code it — not for one-shot user questions. A normal single-model chat link alone does not run “Your documents” or Lab.
Stop betting on a single model.
Early access is limited — join the waitlist and we’ll send your API key as soon as a seat opens.