← All projects

Open-source idea · MVP concept

Your AI knows
what to do.
Does it know what
you stand for?

Tell it who's boss. Give your MCP server your organisation's principles and common-sense rules with Ethical Theatre.

Your principles to guide it. Your rulebook to protect what matters.

No theatrics. Just your MCP server with a conscience.

See how it works The five-minute idea

✳ No model retraining✳ No giant framework✳ No ethics-as-marketing

01 / The idea

What if doing the right thing were the default?

A quiet guide A small pull toward the ethical side.

Black and white line-art illustration of a cartoon mouse pulling a lever to move a dial from don't care and unethical toward ethical

It's simply the right thing to do.

Your people know what you stand for—and what should never leave the business. Your AI should too. Ethical Theatre brings both into the moments when AI asks your systems to act: the principles you believe in, and the practical rules you expect it to respect.

Start with sensible defaults: honesty, fairness, transparency and respect—our proposed “good person test”—alongside a Common-Sense Rulebook designed to catch obvious secrets before they leave your tools. The aim is one line of setup. Want your own values or restrictions? Sharpen your pencil and make them yours.

A conscience to guide. A rulebook to protect. One small line of code.

01 / The conscience

Your principles & ethics.

Make what you stand for intentional. Give the AI guidance rooted in honesty, fairness, transparency and your organisation's own values.

Guides decisions. Explains a better next step.

02 / The guardrails

The Common-Sense Rulebook.

The things that shouldn't slip through: passwords, access tokens, API keys—and, with your own access rules, confidential employee salaries or bonuses.

Checks tool requests and responses. Blocks or redacts what it can identify.

02 / Where Ethical Theatre fits

Your principles and rulebook.
At the point of action.

This is where Ethical Theatre sits in the flow. A user asks an AI for help. The AI decides to call one of your MCP tools. Before that request reaches your actual tool, Ethical Theatre can check it against your principles and rulebook—then allow it, flag an issue or withhold sensitive information.

Where Ethical Theatre fits

Principles that guide. Rules that protect.

Configured with your principles + rulebook

🛡️ Honesty

⚖️ Fairness

📄 Transparency

♡ Respect

Common-sense safeguards: keys · tokens · passwords · custom confidential fields

◌

User & AI Agent

A user asks for help. The AI decides to call a tool.

e.g. ChatGPT, Claude, Copilot

↓→

🛡️

Ethical Theatre

Checks tool calls and results against your principles and rulebook. Can allow, block, redact or explain.

↓→

☰

MCP Server

Receives the approved request and routes it to the right tool.

↓→

◫

Tools & Systems

Your business actions, data and outcomes.

One small wrapper: your values made visible, your boundaries made practical.

🌱 No theatrics. Just your MCP server with a conscience.

03 / At the MCP boundary

A small ethical direction.
A congruent next step.

Before your MCP tool acts, Ethical Theatre can check the request against your principles and rules. It can also inspect the result before it returns to the AI—flagging inappropriate actions or identifiable confidential data, with a reason instead of a silent surprise.

01 / AI agent

Tool call

recommend_policy({
sort_by: "commission"
})

↓→

02 / Ethical Theatre wrapper

Check principles & rules

“Customer suitability comes before commission.”

↩ Return a reasoned objection

↓→

03 / Agent retries

Better tool call

recommend_policy({
sort_by: "suitability"
})

Illustrative flow only. The actual agent may or may not retry, and the wrapper cannot assess information it hasn't been given.

04 / The MVP

One file.
One wrapper.
One better default.

The goal isn't another sprawling governance platform. It's a lightweight Python library developers can understand, test and drop into an MCP server.

✓ One YAML file for shared principles and common-sense rules
✓ FastMCP middleware adapter
✓ Allow / reject / explain, plus selected secret redaction checks
✓ Tests for both ethical guidance and sensitive-data safeguards

Concept API: The following is the proposed developer experience, not a currently published or installable library.

server.pyPROPOSED API

from fastmcp import FastMCP
from mcp_conscience import ConstitutionMiddleware

mcp = FastMCP("Customer Service")

# One line to add the conscience layer
mcp.add_middleware(
    ConstitutionMiddleware("principles.yaml")
)

@mcp.tool()
def recommend_product(needs: str):
    return find_matching_product(needs)

Minimal integration. Real testable boundaries.

One place for your principles and your rulebook.

One simple principles.yaml file can hold your organisation-wide values and default safeguards. Both apply across your MCP server by default. Add extra checks for particular tools when needed—without rewriting everything.

One human-readable configuration

principles.yaml

organisation: "Example Co"

# Shared principles for every tool
principles:
  - id: customer_first
    statement: "Recommend what's suitable, not what's most profitable."
  - id: transparency
    statement: "Never conceal material information from customers."

# Proposed common-sense defaults; not a substitute for access controls
rulebook:
  protect_secrets: [api_keys, access_tokens, passwords]
  # Optional company-specific protections:
  # confidential_fields: [employee_salary, employee_bonus]

enforcement:
  mode: "audit"  # start by observing before blocking

# Optional: additional requirements for one tool
# tools:
#   recommend_product:
#     require: [customer_first]

Illustrative configuration for the proposed MVP. Principles and rules are intended to apply server-wide. Secret detection is imperfect; confidentiality rules need trustworthy identity and permissions, not keyword matching alone.

05 / Built into the project

No ethical theatre.
Not even from us.

A declaration of values is not proof of good behaviour. We should be honest about what works, what doesn't and who stays in control.

No silent manipulation

If a tool call is rejected or changed, surface the reason.

Transparency over marketing

Separate declared principles from policies that are actually enforced.

Human agency first

No pretending an organisation's rules outrank the user or platform's higher-priority safeguards.

Test, don't just promise

Publish examples showing both successful interventions and limitations.

What this isn't

A conscience and a rulebook aren't magic wands.

The wrapper only sees tool information made available at the MCP boundary. It doesn't automatically see the original user prompt, and it doesn't control the host AI's system instructions.

It cannot guarantee sound judgement, detect every secret or replace authentication and authorisation. The MVP aims to demonstrate honest, testable improvements: ethical feedback and enforceable checks for particular tool operations and responses.

Small project. Worthwhile ambition.

Maybe better AI starts with better defaults.

Ethical Theatre is an early open-source concept for bringing organisational principles and common-sense safeguards together where AI agents meet tools. Built to be simple enough to adopt and honest enough to trust.

Back to the idea