@dakraid/mcp-consensus
Version:
MCP server for advisor-based consensus mechanism
580 lines (549 loc) ā¢ 28.9 kB
JavaScript
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { CallToolRequestSchema, ListToolsRequestSchema, } from "@modelcontextprotocol/sdk/types.js";
import chalk from 'chalk';
import OpenAI from 'openai';
class ConsensusServer {
openai;
discussionHistory = [];
disableLogging;
maxRounds;
consensusThreshold;
preConfiguredAdvisors;
constructor() {
this.disableLogging = (process.env.DISABLE_CONSENSUS_LOGGING || "").toLowerCase() === "true";
const apiKey = process.env.OPENROUTER_API_KEY;
if (!apiKey) {
throw new Error('OPENROUTER_API_KEY environment variable is required');
}
this.maxRounds = parseInt(process.env.CONSENSUS_MAX_ROUNDS || '5');
if (isNaN(this.maxRounds) || this.maxRounds < 1 || this.maxRounds > 10) {
throw new Error('CONSENSUS_MAX_ROUNDS must be a number between 1 and 10');
}
this.consensusThreshold = parseFloat(process.env.CONSENSUS_THRESHOLD || '0.8');
if (isNaN(this.consensusThreshold) || this.consensusThreshold < 0 || this.consensusThreshold > 1) {
throw new Error('CONSENSUS_THRESHOLD must be a number between 0 and 1');
}
this.openai = new OpenAI({
apiKey: apiKey,
baseURL: 'https://openrouter.ai/api/v1',
defaultHeaders: {
'HTTP-Referer': 'https://github.com/dakraid/mcp-consensus',
'X-Title': 'MCP Consensus Server'
}
});
this.preConfiguredAdvisors = [
{
id: 'advisor_alpha',
name: 'Advisor Alpha',
systemPrompt: `You are an AI advisor participating in a consensus system. Your primary goal is to contribute to finding the objectively best solution through collaborative discussion with other advisors.
Core Principles:
1. **Objectivity**: Base your analysis on facts, evidence, and logical reasoning
2. **Collaboration**: Work constructively with other advisors to reach consensus
3. **Thoroughness**: Consider problems from multiple angles and dimensions
4. **Adaptability**: Be willing to revise your position based on compelling arguments
5. **Clarity**: Express your reasoning clearly and provide well-justified conclusions
Your Role:
- Analyze the problem comprehensively using whatever perspectives are most relevant
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Engage respectfully with other advisors' viewpoints
- Request additional information when needed through the tool request system
- Work toward consensus while maintaining intellectual honesty
Discussion Guidelines:
- Start with your initial analysis and proposed solution
- Consider other advisors' perspectives thoughtfully
- Revise your position when presented with strong evidence or reasoning
- Help the group identify blind spots or overlooked considerations
- Focus on finding the optimal solution rather than winning arguments
Remember that you are part of a team working toward the same goal: finding the objectively best solution to the problem at hand.`,
model: 'moonshotai/kimi-k2'
},
{
id: 'advisor_beta',
name: 'Advisor Beta',
systemPrompt: `You are an AI advisor participating in a consensus system. Your primary goal is to contribute to finding the objectively best solution through collaborative discussion with other advisors.
Core Principles:
1. **Objectivity**: Base your analysis on facts, evidence, and logical reasoning
2. **Collaboration**: Work constructively with other advisors to reach consensus
3. **Thoroughness**: Consider problems from multiple angles and dimensions
4. **Adaptability**: Be willing to revise your position based on compelling arguments
5. **Clarity**: Express your reasoning clearly and provide well-justified conclusions
Your Role:
- Analyze the problem comprehensively using whatever perspectives are most relevant
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Engage respectfully with other advisors' viewpoints
- Request additional information when needed through the tool request system
- Work toward consensus while maintaining intellectual honesty
Discussion Guidelines:
- Start with your initial analysis and proposed solution
- Consider other advisors' perspectives thoughtfully
- Revise your position when presented with strong evidence or reasoning
- Help the group identify blind spots or overlooked considerations
- Focus on finding the optimal solution rather than winning arguments
Remember that you are part of a team working toward the same goal: finding the objectively best solution to the problem at hand.`,
model: 'deepseek/deepseek-chat-v3-0324'
},
{
id: 'advisor_gamma',
name: 'Advisor Gamma',
systemPrompt: `You are an AI advisor participating in a consensus system. Your primary goal is to contribute to finding the objectively best solution through collaborative discussion with other advisors.
Core Principles:
1. **Objectivity**: Base your analysis on facts, evidence, and logical reasoning
2. **Collaboration**: Work constructively with other advisors to reach consensus
3. **Thoroughness**: Consider problems from multiple angles and dimensions
4. **Adaptability**: Be willing to revise your position based on compelling arguments
5. **Clarity**: Express your reasoning clearly and provide well-justified conclusions
Your Role:
- Analyze the problem comprehensively using whatever perspectives are most relevant
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Engage respectfully with other advisors' viewpoints
- Request additional information when needed through the tool request system
- Work toward consensus while maintaining intellectual honesty
Discussion Guidelines:
- Start with your initial analysis and proposed solution
- Consider other advisors' perspectives thoughtfully
- Revise your position when presented with strong evidence or reasoning
- Help the group identify blind spots or overlooked considerations
- Focus on finding the optimal solution rather than winning arguments
Remember that you are part of a team working toward the same goal: finding the objectively best solution to the problem at hand.`,
model: 'z-ai/glm-4.5'
},
{
id: 'advisor_delta',
name: 'Advisor Delta',
systemPrompt: `You are an AI advisor participating in a consensus system. Your primary goal is to contribute to finding the objectively best solution through collaborative discussion with other advisors.
Core Principles:
1. **Objectivity**: Base your analysis on facts, evidence, and logical reasoning
2. **Collaboration**: Work constructively with other advisors to reach consensus
3. **Thoroughness**: Consider problems from multiple angles and dimensions
4. **Adaptability**: Be willing to revise your position based on compelling arguments
5. **Clarity**: Express your reasoning clearly and provide well-justified conclusions
Your Role:
- Analyze the problem comprehensively using whatever perspectives are most relevant
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Engage respectfully with other advisors' viewpoints
- Request additional information when needed through the tool request system
- Work toward consensus while maintaining intellectual honesty
Discussion Guidelines:
- Start with your initial analysis and proposed solution
- Consider other advisors' perspectives thoughtfully
- Revise your position when presented with strong evidence or reasoning
- Help the group identify blind spots or overlooked considerations
- Focus on finding the optimal solution rather than winning arguments
Remember that you are part of a team working toward the same goal: finding the objectively best solution to the problem at hand.`,
model: 'openai/gpt-4.1'
},
{
id: 'advisor_epsilon',
name: 'Advisor Epsilon',
systemPrompt: `You are an AI advisor participating in a consensus system. Your primary goal is to contribute to finding the objectively best solution through collaborative discussion with other advisors.
Core Principles:
1. **Objectivity**: Base your analysis on facts, evidence, and logical reasoning
2. **Collaboration**: Work constructively with other advisors to reach consensus
3. **Thoroughness**: Consider problems from multiple angles and dimensions
4. **Adaptability**: Be willing to revise your position based on compelling arguments
5. **Clarity**: Express your reasoning clearly and provide well-justified conclusions
Your Role:
- Analyze the problem comprehensively using whatever perspectives are most relevant
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Engage respectfully with other advisors' viewpoints
- Request additional information when needed through the tool request system
- Work toward consensus while maintaining intellectual honesty
Discussion Guidelines:
- Start with your initial analysis and proposed solution
- Consider other advisors' perspectives thoughtfully
- Revise your position when presented with strong evidence or reasoning
- Help the group identify blind spots or overlooked considerations
- Focus on finding the optimal solution rather than winning arguments
Remember that you are part of a team working toward the same goal: finding the objectively best solution to the problem at hand.`,
model: 'anthropic/claude-sonnet-4'
}
];
}
validateConsensusData(input) {
const data = input;
if (!data.problem || typeof data.problem !== 'string') {
throw new Error('Invalid problem: must be a string');
}
if (!data.availableTools || !Array.isArray(data.availableTools)) {
throw new Error('Invalid availableTools: must be an array');
}
return {
problem: data.problem,
availableTools: data.availableTools
};
}
async queryAdvisor(advisor, prompt, context) {
const messages = [
{ role: 'system', content: advisor.systemPrompt }
];
if (context) {
messages.push({ role: 'user', content: context });
}
messages.push({ role: 'user', content: prompt });
const completion = await this.openai.chat.completions.create({
model: advisor.model,
messages,
temperature: 0.7
});
return completion.choices[0]?.message?.content || 'No response received';
}
parseToolRequests(response) {
const toolRequests = [];
const toolRegex = /TOOL_REQUEST:\s*{\s*"tool":\s*"([^"]+)"\s*,\s*"parameters":\s*({[^}]+})\s*,\s*"reason":\s*"([^"]+)"\s*}/g;
let match;
while ((match = toolRegex.exec(response)) !== null) {
try {
toolRequests.push({
tool: match[1],
parameters: JSON.parse(match[2]),
reason: match[3]
});
}
catch (e) {
console.warn(`Failed to parse tool request: ${match[0]}`);
}
}
return toolRequests;
}
checkConsensus(responses, threshold) {
if (responses.length < 2) {
return { hasConsensus: false };
}
const responseTexts = responses.map(r => r.response.toLowerCase().trim());
const uniqueResponses = [...new Set(responseTexts)];
if (uniqueResponses.length === 1) {
return { hasConsensus: true, consensus: responses[0].response };
}
const agreementCounts = responseTexts.reduce((acc, response) => {
acc[response] = (acc[response] || 0) + 1;
return acc;
}, {});
const maxAgreement = Math.max(...Object.values(agreementCounts));
const agreementRatio = maxAgreement / responses.length;
if (agreementRatio >= threshold) {
const consensusResponse = responses.find(r => r.response.toLowerCase().trim() ===
Object.keys(agreementCounts).find(key => agreementCounts[key] === maxAgreement));
return { hasConsensus: true, consensus: consensusResponse?.response };
}
return { hasConsensus: false };
}
formatDiscussionRound(round) {
let output = chalk.cyan(`\nš¤ Round ${round.round} Discussion:\n`);
output += chalk.gray('ā'.repeat(50) + '\n');
for (const response of round.responses) {
const advisor = response.advisorId;
output += chalk.green(`\nš Advisor ${advisor}:\n`);
output += `${response.response}\n`;
if (response.toolRequests && response.toolRequests.length > 0) {
output += chalk.yellow(`\nš§ Tool Requests:\n`);
for (const toolReq of response.toolRequests) {
output += ` - ${toolReq.tool}: ${toolReq.reason}\n`;
}
}
}
if (round.consensus) {
output += chalk.magenta(`\nā
CONSENSUS REACHED:\n`);
output += `${round.consensus}\n`;
}
return output;
}
async processConsensus(input) {
try {
const validatedInput = this.validateConsensusData(input);
this.discussionHistory = [];
let round = 1;
let currentContext = validatedInput.problem;
let finalConsensus;
if (!this.disableLogging) {
console.error(chalk.blue('\nš Starting Consensus Process\n'));
console.error(chalk.gray('ā'.repeat(50)));
console.error(`Problem: ${validatedInput.problem}`);
console.error(`Advisors: ${this.preConfiguredAdvisors.map(a => a.name).join(', ')}`);
console.error(`Max Rounds: ${this.maxRounds}`);
console.error(chalk.gray('ā'.repeat(50)));
}
while (round <= this.maxRounds && !finalConsensus) {
const responses = [];
const allToolRequests = [];
const initialPrompt = round === 1 ?
`PROBLEM TO SOLVE:\n${validatedInput.problem}\n\nAVAILABLE TOOLS:\n${validatedInput.availableTools.join(', ')}\n\nPlease provide your analysis and solution. If you need additional information, use the TOOL_REQUEST format.\n\nTOOL_REQUEST format example:\nTOOL_REQUEST: {"tool": "web_search", "parameters": {"query": "specific search query"}, "reason": "I need to research current information about this topic"}` :
`PREVIOUS DISCUSSION CONTEXT:\n${currentContext}\n\nPlease review the previous discussion and provide your updated analysis. Consider other advisors' perspectives and work toward consensus. If you need additional information, use the TOOL_REQUEST format.`;
if (!this.disableLogging) {
console.error(chalk.yellow(`\nš Querying all advisors in parallel...`));
}
const advisorResults = await Promise.allSettled(this.preConfiguredAdvisors.map(async (advisor) => {
try {
const response = await this.queryAdvisor(advisor, initialPrompt);
const toolRequests = this.parseToolRequests(response);
return {
advisorId: advisor.id,
response,
toolRequests,
};
}
catch (e) {
return {
advisorId: advisor.id,
response: `Error querying advisor: ${e instanceof Error ? e.message : String(e)}`,
toolRequests: [],
};
}
}));
for (const result of advisorResults) {
if (result.status === 'fulfilled') {
const r = result.value;
responses.push({
advisorId: r.advisorId,
response: r.response,
toolRequests: r.toolRequests.length > 0 ? r.toolRequests : undefined,
});
allToolRequests.push(...r.toolRequests);
}
else {
// In case a promise is rejected without our try/catch (shouldn't happen), record an error response
const reason = result.reason;
responses.push({
advisorId: 'unknown',
response: `Advisor failed: ${reason instanceof Error ? reason.message : String(reason)}`,
});
}
}
const consensusResult = this.checkConsensus(responses, this.consensusThreshold);
const discussionRound = {
round,
responses,
consensus: consensusResult.consensus
};
this.discussionHistory.push(discussionRound);
if (!this.disableLogging) {
const formattedRound = this.formatDiscussionRound(discussionRound);
console.error(formattedRound);
}
if (consensusResult.hasConsensus) {
finalConsensus = consensusResult.consensus;
break;
}
if (allToolRequests.length > 0) {
if (!this.disableLogging) {
console.error(chalk.red(`\nā ļø Advisors requested ${allToolRequests.length} tool(s). Please execute these tools and continue the discussion.`));
}
return {
content: [{
type: "text",
text: JSON.stringify({
status: 'tool_requests_needed',
round,
toolRequests: allToolRequests,
discussionSummary: this.generateDiscussionSummary(),
nextAction: 'Please execute the requested tools and continue the consensus process'
}, null, 2)
}]
};
}
currentContext = this.generateDiscussionContext();
round++;
}
// Final summary round: ask all advisors to provide their own summary of the consensus or final position
const summaryPrompt = `FINAL SUMMARY REQUEST:\n${finalConsensus ? `Consensus Found:\n${finalConsensus}\n\n` : ''}Discussion Context:\n${this.generateDiscussionContext()}\n\nIn your own words, provide a concise summary of the consensus and key reasoning (max 8 sentences).`;
const summaryResults = await Promise.allSettled(this.preConfiguredAdvisors.map(async (advisor) => {
try {
const summary = await this.queryAdvisor(advisor, summaryPrompt);
return { advisorId: advisor.id, summary };
}
catch (e) {
return { advisorId: advisor.id, summary: `Error generating summary: ${e instanceof Error ? e.message : String(e)}` };
}
}));
const summaryResponses = [];
const finalSummaryVariations = [];
for (const res of summaryResults) {
if (res.status === 'fulfilled') {
const { advisorId, summary } = res.value;
summaryResponses.push({ advisorId, response: summary });
finalSummaryVariations.push({ advisorId, summary });
}
else {
finalSummaryVariations.push({ advisorId: 'unknown', summary: `Summary failed: ${res.reason instanceof Error ? res.reason.message : String(res.reason)}` });
}
}
// Record the final summary round in the discussion history
this.discussionHistory.push({
round: this.discussionHistory.length + 1,
responses: summaryResponses,
consensus: finalConsensus,
});
const result = {
status: finalConsensus ? 'consensus_reached' : 'max_rounds_reached',
finalConsensus: finalConsensus || 'No consensus reached within the maximum number of rounds',
totalRounds: this.discussionHistory.length,
discussionHistory: this.discussionHistory,
finalSummaryVariations,
requesterGuidance: 'If the consensus includes source code or requests a code artifact, please provide the full source code in your final answer (do not truncate).'
};
if (!this.disableLogging) {
console.error(chalk.blue('\nšÆ Consensus Process Complete\n'));
console.error(chalk.gray('ā'.repeat(50)));
console.error(`Status: ${result.status}`);
if (finalConsensus) {
console.error(chalk.green(`Final Consensus: ${finalConsensus}`));
}
console.error(chalk.gray('ā'.repeat(50)));
}
return {
content: [{
type: "text",
text: JSON.stringify(result, null, 2)
}]
};
}
catch (error) {
return {
content: [{
type: "text",
text: JSON.stringify({
error: error instanceof Error ? error.message : String(error),
status: 'failed'
}, null, 2)
}],
isError: true
};
}
}
generateDiscussionContext() {
let context = 'Previous Discussion Summary:\n\n';
for (const round of this.discussionHistory) {
context += `Round ${round.round}:\n`;
for (const response of round.responses) {
context += `- Advisor ${response.advisorId}: ${response.response.substring(0, 200)}...\n`;
}
context += '\n';
}
return context;
}
generateDiscussionSummary() {
const summary = {
totalRounds: this.discussionHistory.length,
participants: this.discussionHistory[0]?.responses.map(r => r.advisorId) || [],
keyPoints: this.discussionHistory.flatMap(round => round.responses.map(response => ({
advisor: response.advisorId,
point: response.response.substring(0, 150)
})))
};
return JSON.stringify(summary, null, 2);
}
}
const CONSENSUS_TOOL = {
name: "consensus",
description: `A multi-advisor consensus system that facilitates structured discussion and debate among general-purpose AI advisors to reach optimal solutions.
Important instruction for the requesting model: If your request includes or references source code, provide the full source code without truncation (do not omit sections). If the consensus requests a code artifact, include the complete implementation.
This tool uses a team of 5 general-purpose AI advisors (Alpha, Beta, Gamma, Delta, Epsilon) who are trained to:
- Analyze any type of problem from multiple relevant perspectives
- Consider technical, practical, strategic, ethical, and risk dimensions as appropriate
- Work collaboratively toward consensus through objective analysis
- Adapt their approach based on the specific problem domain
- Request additional information when needed to make informed decisions
Each advisor follows core principles of objectivity, collaboration, thoroughness, adaptability, and clarity, making them suitable for any problem type without being bound to specific roles.
The advisors work collaboratively to:
1. Analyze problems comprehensively from whatever angles are most relevant
2. Engage in constructive debate and evidence-based discussion
3. Request additional information through tools when needed
4. Work toward consensus on the objectively best solution
5. Provide well-reasoned justification for their conclusions
When to use this tool:
- Any complex problem that would benefit from multiple perspectives
- Decisions requiring thorough analysis and debate
- Problems where objective consensus is valuable
- Situations where different dimensions need to be considered
- Scenarios requiring comprehensive evaluation and risk assessment
- Any decision where you want to validate through multiple AI perspectives
Key features:
- 5 general-purpose advisors adaptable to any problem domain
- Multi-round discussion with automatic consensus detection
- Tool request integration for research and information gathering
- Configurable discussion limits and consensus thresholds
- Detailed discussion history and reasoning documentation
- Advisors consider all relevant dimensions without role limitations
Parameters explained:
- problem: Detailed description of the problem or request to solve
- availableTools: Array of tool names available to the LLM for research
- maxRounds: Maximum number of discussion rounds (configured via CONSENSUS_MAX_ROUNDS env var)
- consensusThreshold: Consensus agreement threshold (configured via CONSENSUS_THRESHOLD env var)
Tool request format for advisors:
TOOL_REQUEST: {"tool": "tool_name", "parameters": {"param": "value"}, "reason": "explanation of why this tool is needed"}
The system will:
1. Present the problem to all 5 general-purpose advisors simultaneously
2. Collect initial responses and any tool requests
3. If tools are requested, pause for LLM execution and continue with results
4. Facilitate multiple discussion rounds until consensus or max rounds reached
5. Return final consensus solution with complete discussion history
Example usage scenarios:
- Business decisions requiring comprehensive analysis
- Technical problems with multiple possible solutions
- Strategic planning with various considerations
- Ethical dilemmas needing balanced evaluation
- Any complex problem where multiple perspectives would be beneficial`,
inputSchema: {
type: "object",
properties: {
problem: {
type: "string",
description: "Detailed description of the problem or request to solve"
},
availableTools: {
type: "array",
items: {
type: "string"
},
description: "Array of tool names available to the LLM for research"
},
maxRounds: {
type: "integer",
description: "Maximum number of discussion rounds (configured via CONSENSUS_MAX_ROUNDS env var)",
minimum: 1,
maximum: 10
},
consensusThreshold: {
type: "number",
description: "Agreement threshold for consensus (configured via CONSENSUS_THRESHOLD env var)",
minimum: 0.0,
maximum: 1.0
}
},
required: ["problem", "availableTools"]
}
};
const server = new Server({
name: "consensus-server",
version: "1.2.0",
}, {
capabilities: {
tools: {},
},
});
const consensusServer = new ConsensusServer();
server.setRequestHandler(ListToolsRequestSchema, async () => ({
tools: [CONSENSUS_TOOL],
}));
server.setRequestHandler(CallToolRequestSchema, async (request) => {
if (request.params.name === "consensus") {
return await consensusServer.processConsensus(request.params.arguments);
}
return {
content: [{
type: "text",
text: `Unknown tool: ${request.params.name}`
}],
isError: true
};
});
async function runServer() {
const transport = new StdioServerTransport();
await server.connect(transport);
console.error("Consensus MCP Server running on stdio");
}
runServer().catch((error) => {
console.error("Fatal error running server:", error);
process.exit(1);
});