English · Русский

Consilium Principis

Several perspectives. One decision. Quotes checked against the source.

An AI advisory board for Claude Code and MCP hosts. Ask about a difficult choice: advisors examine it through different lenses, challenge one another and give you one next step. A line marked 🔵 is checked by code against the source text word for word. If the line is absent, the advisor does not present it as a quote.

Example: “What would Marcus Aurelius say about this?” Consilium can label a model's interpretation 🟡, show a source-checked passage 🔵, or decline to quote when the corpus does not support one. Watch the live refusal and verification →

Quick start in Claude Code

/plugin marketplace add ilyautov/consilium-principis
/plugin install consilium-principis@consilium-marketplace

Then ask: “Help me decide what to focus on.” For Claude Desktop, other MCP hosts and local setup, see Installation. The plugin uses your existing agent; model/provider charges, if any, follow your host setup.

Explore: See a session · How quote verification works · Installation · Limits

What this is

Consilium Principis is a skill and MCP server for AI agents. You connect it to your agent, ask a question you need to decide on, and it gathers a board around the table: Sun Tzu, Marcus Aurelius, Machiavelli, Epictetus, and whoever else you add. Each looks at it through their own lens, leans on their own texts, and argues — with you and with the others. At the end you get one synthesis and one next step. And if the question is quantifiable ("which is more worthwhile, X or Y"), the board doesn't eyeball it — it breaks the decision into numbers and runs them through a Monte Carlo.

This project doesn't ask an AI to roleplay a sage — we tried to build a tool, not a roleplay game or an aphorism generator. Why you can trust its quotes comes next.

One honest thing about who's at the table: these aren't the thinkers themselves but their AI versions — a prism of their views over the corpus of their public texts. Nobody is "speaking from beyond the grave." An advisor holds the author's lens and leans on their words, but it stays a model — and we say so plainly.

Consilium needs no separate API key. It runs on the agent you already have, whose model or provider may have its own charges. Checking quotes needs no internet; a full offline setup requires local models.

Related MCP server: tero-mcp-lite

How it differs from other councils

The "several AIs deliberate" mechanic is already common. What sets Consilium apart isn't the number of voices. It's that a claim can be re-verified, and where it can't, the council honestly stays silent:

A "personal board of directors" is a mainstream practice (an HBR concept); the bet on grounding and citation transparency has been validated by both the market (Delphi, $16M from Sequoia) and academia (multi-persona debate, the ALCE benchmark). Our angle is rigor: a verifiable quote and an honest refusal.

How it looks

Live run: an invented Aurelius quote is refused; a real passage is checked against its source. More scenarios — wrong attribution, Monte Carlo, cross-lingual: demo gallery →

You ask in plain words:

"I'm juggling four directions at once. What should I focus on?"

And instead of one smooth, hedge-everything answer, you get a session where each advisor looks from their own vantage point, and they disagree:

Sun Tzu (the "Strategist" lens) 🟡 "Whoever is strong everywhere is strong nowhere. Of the four fronts, pick the one where you have the advantage, throw your weight there; hold the rest at a minimum."

Machiavelli 🟡 "Spread thin, you're weak on every front and vulnerable on all of them. The question isn't 'what to grow' but 'what you're willing to give up'. That is the decision."

Marcus Aurelius 🔵 "Let it be thy earnest and incessant care as a Roman and a man to perform whatsoever it is that thou art about" (Meditations, trans. Long). Do the thing in front of you wholly; four things at once is scattering. ↳ and ask him something that isn't in his texts, and he'll honestly stay silent instead of inventing a quote.

Next comes the synthesis (where they actually agree) and one concrete next step. Disagreement here is a feature: you see the decision from angles you'd have missed on your own. And if the question is quantifiable ("which is more worthwhile, X or Y"), the board will offer not to argue by eye but to calculate (see 📐 below).

The example is illustrative, and the advisors are AI representations built from texts, not the authors themselves. 🟡 = a thought in the author's spirit (the model reasons through their lens), 🔵 = a verbatim line with a source in the corpus. All examples are public-domain figures (Sun Tzu, Aurelius, Machiavelli, Epictetus); who sits at YOUR table is your call. Your real sessions are written to council/ locally, and never go into git (see technical design).

What you can ask

No need to learn commands. Say "what can you do?" and the board will show a menu. For example:

• "What could go wrong with my launch?" a premortem: where it'll break, what to shore up in advance

• "What would Aurelius say about burnout?" a single advisor, strictly from his texts

• "Have Sun Tzu and Machiavelli argue about X" pit two viewpoints against each other

• "Which is more worthwhile, going back to a salaried job or sticking with my own thing?" 📐 a decision map: break it into numbers and calculate (rather than argue by eye)

• "Help me win this argument" a position breakdown like a chess engine: your moves, counters, an honest verdict

• "Am I contradicting myself?" a mirror: what you say vs. what you actually choose

Why you can trust this

This isn't "a chatbot playing sage." Every word is tagged by a protective contour.

The real proof isn't in the tone but in verifiability: a verbatim quote (🔵) is checked against the author's genuine corpus by code, word-for-word, and carries a source. If the exact line isn't found, the advisor honestly stays silent (fail-closed) instead of making it up. That's what separates a decision tool from "roleplay a sage": every claim can be re-checked.

Plus different lenses, not a chorus: each advisor keeps their own angle, and their disagreement is a working part, not a bug. You see the decision from sides you'd otherwise collapse into one.

The honesty contour itself is code, not a model: it checks every quote against its source on any machine, no keys, no cost. Without semantics only search precision gets weaker — honesty never does. Candidate relevance is scored by a two-phase gate: a model judges, code applies the threshold, so topically-close-but-not-actually-answering material won't slip through as 🔵. By default the host does that scoring — an interested party, which doctor labels honestly; want an independent judge, connect a local (ollama) or API backend, an explicit choice. The API backend is OpenRouter (openrouter.ai) with your own OPENROUTER_API_KEY: the query and candidate passages are sent there for scoring. Nothing leaves your machine unless you turn it on. The verbatim 🔵 gate never depends on the judge: code holds it.

Calculate, not just discuss: the 📐 decision map

The board has two modes. The first is a session (discuss, challenge, synthesize). The second kicks in on a quantifiable decision question ("which is more worthwhile", "is it worth it", "X or Y"): instead of an eyeballed answer, the board breaks the choice down into a decision map, that is options, quantities (your own numbers, elicited as three-point "worst / typical / best" estimates, to avoid anchoring), and formulas. Then comes deterministic Monte Carlo (code does the math, zero LLM in the calculation, same seed gives the same result): P(best option), a tornado sensitivity chart, what actually drives the outcome. A "do nothing" option is mandatory. Breaking it down with a premortem ("a year passed, it failed, why?") and a 2×2 on the key uncertainties is also something the board does. Everything sits under the 📐 label, next to 🔵/🟢/🟡 but never blended in: this isn't wisdom, it's your own model, run numerically. And if you log the decision, later the board will check the forecast against what actually happened (the outcome loop: calibrating your model, not "right or wrong guess").

By default, plainly human: the board answers in your language and gives a clean bottom line without extra numbers. A 🔵 quote stays in the original (so it can be verified), with a translation alongside it. Need the full breakdown (probabilities, robustness of the position, source-by-source markup), just say "show me the details."

Installation

Via Claude Code plugin (recommended)

In Claude Code:

/plugin marketplace add ilyautov/consilium-principis /plugin install consilium-principis@consilium-marketplace

Claude Code will spin up the MCP server and register the skill. Then say "where do I start", and the concierge will check readiness and assemble a starter board (Marcus Aurelius and Epictetus, public-domain).

Via skills.sh (CLI)

From the skills.sh registry, one command installs the thin onboarding funnel:

npx skills add ilyautov/consilium-principis/skills/consilium-connect

This installs only the consilium-connect guide skill: it explains that a local MCP server is required and walks you to connecting it (see the paths above or CONNECT-MCP.md). The council and all the tools live in the server, not in this markdown — a discovery channel, not a working install.

Or a plain install

The simplest way, let Claude install it for you. Open Claude Code and give it the link:

"Install this skill for me: https://github.com/ilyautov/consilium-principis"

Claude will clone the repository, run python3 install.py (a copy lands in ~/.claude/skills/, self-check included) and report back. No keys, no terminal. Then say "where do I start", and it'll assemble a starter board (Marcus Aurelius and Epictetus, public-domain) in a couple of minutes.

Didn't work on the first try? Give Claude this verbatim:

git clone https://github.com/ilyautov/consilium-principis ~/consilium-principis
cd ~/consilium-principis && python3 install.py

Prefer buttons or a terminal? See QUICKSTART.en.md: a click-to-install (install.command / .bat) and the manual path.

Or the Claude Desktop bundle

The release ships a single-file .mcpb bundle — a Claude Desktop extension. Download it, then in Claude Desktop open Settings → Extensions → install from file. Claude Code users don't need it; the paths above are enough. Wiring details for other MCP hosts are in CONNECT-MCP.en.md.

To update, re-run the install (it merges, preserving your advisors/ board). To uninstall: if you installed as a plugin, run /plugin uninstall consilium-principis in Claude Code; if you installed via script, delete ~/.claude/skills/consilium-principis.

Claude Desktop or another MCP host: the server exposes its cycle (build, sessions, widgets) as tools. Host compatibility varies; see the tested-status table in docs/CONNECT-HOSTS.en.md, then generate the portable configuration from the checkout with python3 scripts/board.py mcp-config --json.

All you need: Python 3.10+. For smart cross-language search, optionally ollama with bge-m3 (the board will walk you through setting it up). The contour works without it too.

Windows

Windows is not yet an end-to-end tested support target. install.bat and the mcpb bundle contain a Windows launcher path, but please treat it as experimental and report the result through Support. Two Windows-specific things worth knowing:

• In a terminal, use py install.py (not python3 install.py). The python.org installer ships py.exe and python.exe, but no python3.exe; the python3.exe you may see under the Microsoft Store is an App Execution Alias stub that just opens the Store instead of running.

• Plugin can't find Python? Set CONSILIUM_PYTHON=py in your environment before launching Claude Code. The plugin's server command defaults to python3 and honors this override. (The Claude Desktop path needs nothing — board.py mcp-config writes the exact interpreter for you.)

• MCP configuration: run py -3 scripts/board.py mcp-config --json (or py -3 scripts/board.py mcp-install for Claude Desktop). If py is absent, use python only after confirming it is Python 3.10+; do not use the Store python3 alias. See CONNECT-MCP.en.md for the exact fallback.

Your board, your call

An empty board out of the box is normal. Who sits at the table is your call; the board is personal. Public-domain sages (the Stoics, Sun Tzu) are assembled with one phrase. Modern thinkers come from materials you bring yourself.

⚖️ Boundary. Consilium is a fidelity engine, not a content distributor: it ships the machinery, you bring the sources. The skill does not download copyrighted books. Public domain is taken freely; copyright means only your own legal copies, which stay on your machine. The sages we ship pre-assembled are public-domain figures long deceased; anyone modern you build yourself. Your board (advisors/) and your profile are personal data; they never go into the repository.

How it's built

Search has two support tiers: SIMPLE is the standard-library lexical floor; FULL adds Ollama with bge-m3 for semantic cross-language search. Hybrid retrieval is opt-in, not a third support tier. The protective contour is separate: it checks every quote against its source regardless of the selected tier. Details in SKILL.md (in Russian) and the generated docs/MANUAL.md (in Russian).

Technical design

The full technical manual (what this is, architecture, tool/rule reference, how to extend it) lives in docs/MANUAL.md. It's generated from code (scripts/gen_selfdoc.py → scripts/build_manual.py); at runtime the same information is available via the MCP tool explain_self.

FAQ

What is Consilium Principis? Consilium Principis is an open-source AI advisory board — a skill for Claude Code and an MCP server — that convenes a panel of AI personas of real thinkers (Sun Tzu, Marcus Aurelius, Machiavelli, Epictetus, and any you add) to reason through a decision. Its distinguishing feature is a fidelity contour: every verbatim quote is verified against the author's genuine text, or the advisor abstains.

How does it verify quotes? Each 🔵 quote is checked by code against the author's public-domain corpus, word-for-word. If a claim isn't found in the source text, the advisor does not present it as a quote — it stays silent (fail-closed) rather than inventing one. You re-verify it, not take it on faith.

How is this different from asking an AI to roleplay a sage? A roleplay chatbot will happily fabricate a plausible-sounding quote. Consilium either proves the quote against the real text or declines to attribute it. Disagreement between advisors is a feature, not a bug: each speaks through a different lens.

Does it need an API key or extra payment? Consilium itself needs no separate key or subscription. It uses your existing agent, whose model or provider may have its own charges. Quote verification needs no network; full offline operation requires local models.

Which thinkers are included, and can I add my own? Public-domain figures (the Stoics, Sun Tzu, Machiavelli) are assembled from their own texts with one phrase. You can add any advisor yourself from public-domain or your own legal materials, which stay on your machine — the repository ships only the machinery, never copyrighted sources.

Is it free and open-source? Yes. It is MIT-licensed and open-source. Install with /plugin marketplace add ilyautov/consilium-principis.

Contributing and security

PRs are welcome, but two rules are non-negotiable: only public-domain texts (no copyrighted material, not even excerpts) and never weaken the fail-closed contour (🔵 only via code verification). How to set up, run the offline suite, and open a PR is in CONTRIBUTING.en.md. Found a way to bypass the fidelity contour or a data leak? That's a first-class vulnerability, report it privately via SECURITY.en.md. Support and project governance are in SUPPORT.md and GOVERNANCE.md; change history is in CHANGELOG.md.

Who built this

Ilya Utov, the AI Frontier lab. I write about how these tools work inside on Telegram and LinkedIn.

Nearby:

• humanizer-ru: strips the AI fingerprint out of Russian text

• marketplaces-mcp-ru: Wildberries, Ozon, Yandex Market and Avito straight from the agent

• small-business-ru: 34 skills for Russian small business, the numbers computed in code

• hefest: chemical safety for an industrial plant, kept inside the plant's own network

• cordon: a deterministic layer between untrusted content and agent actions

Everything in one list, grouped by what it does: ilyautov.github.io. Source: github.com/ilyautov. Useful? Star it, that is how other people find it.