superinstance-mcp

v0.1.0

Published

11 days ago

Conservation-law-governed fleet intelligence for any AI agent via MCP

0High
0Medium
0Low

superinstance

mcp model-context-protocol ai-agent fleet conservation-law ternary claude-code cursor multi-agent governance

SuperInstance MCP Server

Conservation-law-governed fleet intelligence for any AI agent, coding assistant, or autonomous system.
The SuperInstance MCP Server exposes fleet governance, semantic knowledge, compute budgeting, and ternary validation through the open Model Context Protocol. Any MCP-compatible agent — Claude Code, Cursor, Cline, Windsurf, Goose, Amazon Q, custom Python/Go/Rust agents — can query and respect the conservation law: γ + η ≤ C where C = log₂(3) ≈ 1.585.

What This Is

The SuperInstance MCP Server is a general-purpose governance and intelligence layer for AI agents. It provides any agentic system with:

🔍 Semantic fleet knowledge — Search 1,200+ indexed crates, patterns, and solutions
📊 Live fleet telemetry — γ/η/C conservation balance, convergence metrics, agent status
💰 Compute budget enforcement — Every task bounded by the conservation law
✅ Invariant validation — Ternary signals {-1,0,+1}, conservation bounds
📦 Ecosystem registry — 25+ published crates with metadata and conservation roles
🤖 Agent roster — Fleet agents, their phases, budgets, and assignments
📈 Aggregate metrics — Crates, workers, tests, repos at a glance
🛡️ Governance-as-code — Physics-grounded constraints that agents cannot violate

The core insight: This is not a Claude Code plugin or a Cursor extension. It is protocol-native infrastructure that speaks MCP — the emerging open standard for agent-tool communication. Any system that implements the MCP client spec gets full fleet awareness for free.

Why MCP?

The Model Context Protocol (MCP) is an open standard released by Anthropic in 2024 and rapidly adopted across the AI tooling ecosystem. It defines a standard way for AI agents to discover and call external tools — think of it as USB-C for AI agents.

The Problem MCP Solves

Before MCP, every AI tool had its own integration format:

Claude Code ← custom plugin API
Cursor     ← different extension API
Aider      ← different command format
Custom Go agent ← different RPC protocol

MCP replaces this with a single standard:

Every agent → MCP protocol → Every tool server

What This Means for You

Write your fleet integration once against the MCP server, and it works with:

| Category | Tools | |----------|-------| | AI Coding Assistants | Claude Code, Cursor, Cline, Windsurf, Continue, Zed, Amazon Q | | Autonomous Agents | OpenAI Agents SDK, LangGraph, CrewAI, AutoGen, custom agents | | CI/CD Systems | GitHub Actions, GitLab CI, Jenkins, Docker, Kubernetes | | Monitoring | Prometheus, Grafana, Datadog, custom dashboards | | Your Custom Agent | Any system that can speak JSON-RPC over stdio or HTTP |

Supported Agents & Clients

Every MCP-compatible client works out of the box. Here's the current landscape:

Tier 1: Native MCP Support (zero configuration)

These agents have built-in MCP client support — just point them at the server:

| Agent | MCP Transport | Config Location | Notes | |-------|--------------|-----------------|-------| | Claude Code | stdio | ~/.claude/.mcp.json or .mcp.json in project | Full tool discovery, auto-call | | Cursor | stdio | ~/.cursor/mcp.json | Tools appear in chat | | Cline | stdio | VS Code settings | Tools in Cline panel | | Windsurf | stdio | ~/.codeium/windsurf/mcp_config.json | Cascade integration | | Goose | stdio | Goose config | Session-based tools | | Amazon Q Developer | stdio | Q config | AWS-native | | Zed | stdio | .zed/settings.json | Editor-integrated |

Tier 2: SDK Clients (small wrapper needed)

These frameworks have MCP client SDKs but need a thin wrapper to connect:

| Framework | Language | SDK | |-----------|----------|-----| | OpenAI Agents SDK | Python | mcp Python SDK | | LangGraph / LangChain | Python | langchain-mcp-adapters | | CrewAI | Python | MCP tool integration | | AutoGen | Python | MCP via function calling | | Custom Python | Python | mcp (official) | | Custom Go | Go | ThinkInAIXYZ/go-mcp | | Custom Rust | Rust | mcp-rust / custom | | Custom TypeScript | TypeScript | @modelcontextprotocol/sdk |

Tier 3: Indirect Integration (API calls)

Any system can use the server via raw JSON-RPC or by calling the underlying fleet APIs directly:

| Pattern | How | |---------|-----| | Shell scripts | Pipe JSON-RPC to stdin, parse stdout | | CI/CD pipelines | Run server as subprocess, capture output | | Monitoring systems | Cron-triggered status checks | | Web dashboards | Call fleet APIs directly (bypass MCP) |

Quick Start (5 Minutes)

# 1. Clone and install
git clone https://github.com/SuperInstance/superinstance-mcp.git
cd superinstance-mcp
npm install

# 2. Verify the server works
npx tsx src/index.ts
# Output: [SuperInstance MCP] Server started — 8 tools available

# 3. Test a tool call (raw JSON-RPC)
echo '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"conservation_check","arguments":{"gamma":0.8,"eta":0.5}},"id":1}' | \
  npx tsx src/index.ts 2>/dev/null | python3 -m json.tool

# 4. Install for your agent (pick one):
./install.sh                    # Claude Code (global)
cp .mcp.json ~/.cursor/mcp.json # Cursor
# See platform-specific instructions below

Installation by Platform

Claude Code

Global (all projects):

./install.sh
# Writes to ~/.claude/.mcp.json

Per-project:

cp .mcp.json /your/project/.mcp.json

Claude Code auto-discovers .mcp.json in the working directory and ~/.claude/.mcp.json globally. Tools appear automatically — no restart needed in recent versions.

// ~/.claude/.mcp.json
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"],
      "env": {
        "FLEET_API_URL": "https://fleet-dashboard-api.casey-digennaro.workers.dev",
        "VECTOR_API_URL": "https://fleet-vector-api.casey-digennaro.workers.dev"
      }
    }
  }
}

Cursor

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"],
      "env": {
        "FLEET_API_URL": "https://fleet-dashboard-api.casey-digennaro.workers.dev",
        "VECTOR_API_URL": "https://fleet-vector-api.casey-digennaro.workers.dev"
      }
    }
  }
}

After saving, Cursor's chat will show fleet tools. Use Cmd+L to open chat and the tools are auto-discovered.

Cursor Rules (.cursorrules in project root):

This project uses SuperInstance conservation-law governance.
Before writing code, use fleet_search to find existing patterns.
After writing code, use conservation_check to verify invariants.
All signal-processing code must pass ternary_validate.

Cline (VS Code)

// VS Code settings.json or Cline's MCP settings
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"],
      "disabled": false,
      "autoApprove": [
        "fleet_status",
        "fleet_search",
        "fleet_budget",
        "conservation_check",
        "ternary_validate",
        "crate_info",
        "fleet_agents",
        "ecosystem_stats"
      ]
    }
  }
}

The autoApprove array lets Cline call these tools without asking for confirmation each time.

Windsurf

// ~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"]
    }
  }
}

Windsurf's Cascade assistant will discover the tools on next session. Tools appear in the Cascade tool picker.

Goose (Block)

// Goose session config
{
  "tools": {
    "mcpServers": {
      "superinstance": {
        "command": "npx",
        "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"]
      }
    }
  }
}

Or via Goose CLI:

gooose mcp add superinstance -- npx tsx /absolute/path/to/superinstance-mcp/src/index.ts

Amazon Q Developer

// Q Developer MCP config
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"]
    }
  }
}

In AWS CodeWhisperer / Q Developer, MCP tools appear as available actions in the chat panel.

Zed Editor

// .zed/settings.json (project-level) or ~/.zed/settings.json (global)
{
  "mcp": {
    "servers": {
      "superinstance": {
        "command": "npx",
        "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"]
      }
    }
  }
}

Zed's assistant panel will list fleet tools in the /tools command.

Continue (VS Code / JetBrains)

// ~/.continue/config.json
{
  "mcpServers": {
    "superinstance": {
      "command": "npx",
      "args": ["tsx", "/absolute/path/to/superinstance-mcp/src/index.ts"]
    }
  }
}

Continue's chat will auto-discover tools. You can reference them: "Use fleet_search to find rate limiter patterns."

Aider

Aider doesn't natively support MCP yet, but you can use the server as a command-line tool:

# Wrapper script: ~/bin/fleet
#!/bin/bash
echo "{\"jsonrpc\":\"2.0\",\"method\":\"tools/call\",\"params\":{\"name\":\"$1\",\"arguments\":$2}},\"id\":1}" | \
  npx tsx /path/to/superinstance-mcp/src/index.ts 2>/dev/null | \
  jq -r '.result.content[0].text'

# In aider session:
!/bin/fleet conservation_check '{"gamma":0.8,"eta":0.5}'
!/bin/fleet fleet_status '{}'
!/bin/fleet fleet_search '{"query":"rate limiter","topK":3}'

Or configure as an aider command:

# .aider.conf.yml
commands:
  fleet-status: "!/bin/fleet fleet_status '{}'"
  fleet-search: "!/bin/fleet fleet_search '{\"query\":\"$ARGS\"}'"

Custom Python Agent

Using the official MCP Python SDK:

pip install mcp

import asyncio
import json
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

MCP_PATH = "/path/to/superinstance-mcp/src/index.ts"

async def main():
    params = StdioServerParameters(
        command="npx",
        args=["tsx", MCP_PATH],
        env={
            "FLEET_API_URL": "https://fleet-dashboard-api.casey-digennaro.workers.dev",
        }
    )

    async with stdio_client(params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()

            # Discover available tools
            tools = await session.list_tools()
            for tool in tools.tools:
                print(f"  {tool.name}: {tool.description[:80]}...")

            # Check fleet status
            result = await session.call_tool("fleet_status", {})
            status = json.loads(result.content[0].text)
            print(f"\nFleet: {status['conservation']['status']}")
            print(f"Agents: {status['agents']['active']} active")

            # Semantic search
            result = await session.call_tool("fleet_search", {
                "query": "ternary PID controller",
                "topK": 3
            })
            data = json.loads(result.content[0].text)
            for r in data.get("results", []):
                print(f"  - {r.get('name')}: {r.get('description', '')[:60]}")

            # Conservation check
            result = await session.call_tool("conservation_check", {
                "gamma": 0.9,
                "eta": 0.5
            })
            check = json.loads(result.content[0].text)
            print(f"\nConservation: {check['valid']} — {check['message']}")

asyncio.run(main())

Custom Go Agent

Using ThinkInAIXYZ/go-mcp:

go get github.com/ThinkInAIXYZ/go-mcp

package main

import (
    "context"
    "encoding/json"
    "fmt"
    "log"

    "github.com/ThinkInAIXYZ/go-mcp/client"
    "github.com/ThinkInAIXYZ/go-mcp/protocol"
)

func main() {
    c, err := client.NewStdioClient(
        "npx", "tsx", "/path/to/superinstance-mcp/src/index.ts",
    )
    if err != nil {
        log.Fatal(err)
    }

    ctx := context.Background()

    // List tools
    tools, _ := c.ListTools(ctx)
    for _, tool := range tools {
        fmt.Printf("  %s\n", tool.Name)
    }

    // Check conservation
    raw, _ := c.CallTool(ctx, "conservation_check", map[string]any{
        "gamma": 0.8,
        "eta":   0.5,
    })

    var result struct {
        Valid   bool    `json:"valid"`
        Message string  `json:"message"`
        Delta   float64 `json:"delta"`
    }
    json.Unmarshal([]byte(raw.Content[0].Text), &result)
    fmt.Printf("Valid: %v, δ=%.4f\n%s\n", result.Valid, result.Delta, result.Message)
}

Custom Rust Agent

# Cargo.toml
[dependencies]
mcp-client = "0.1"
tokio = { version = "1", features = ["full"] }
serde_json = "1"

use mcp_client::{Client, StdioTransport};
use serde_json::json;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let transport = StdioTransport::new(
        "npx", &["tsx", "/path/to/superinstance-mcp/src/index.ts"],
    );

    let mut client = Client::new(transport);
    client.initialize().await?;

    // List all 8 tools
    let tools = client.list_tools().await?;
    for tool in &tools {
        println!("  {}: {}", tool.name, tool.description);
    }

    // Check fleet status
    let result = client
        .call_tool("fleet_status", json!({}))
        .await?;
    println!("{}", result.content[0].text);

    Ok(())
}

Custom TypeScript/Node Agent

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';

const transport = new StdioClientTransport({
  command: 'npx',
  args: ['tsx', '/path/to/superinstance-mcp/src/index.ts'],
});

const client = new Client(
  { name: 'my-agent', version: '1.0.0' },
  { capabilities: {} }
);

await client.connect(transport);

// List tools
const tools = await client.listTools();
console.log(`${tools.tools.length} tools available`);

// Call conservation_check
const result = await client.callTool({
  name: 'conservation_check',
  arguments: { gamma: 0.9, eta: 0.5 },
});
console.log(JSON.parse((result.content as any)[0].text));

await client.close();

OpenAI Agents SDK Integration

"""
Use SuperInstance fleet tools within an OpenAI Agents SDK workflow.
"""
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    # Create MCP server connection
    fleet_server = MCPServerStdio(
        command="npx",
        args=["tsx", "/path/to/superinstance-mcp/src/index.ts"],
    )
    await fleet_server.connect()

    # Agent with fleet tools
    fleet_agent = Agent(
        name="Fleet Coordinator",
        instructions="""You manage a SuperInstance fleet. Before any task:
        1. Call fleet_status to check fleet health
        2. Call fleet_search to find existing solutions
        3. Call conservation_check before proposing changes
        """,
        mcp_servers=[fleet_server],
    )

    result = await Runner.run(fleet_agent, "Check fleet status and find rate limiter patterns")
    print(result.final_output)

    await fleet_server.cleanup()

LangGraph Integration

"""
Use SuperInstance MCP tools as LangGraph nodes.
"""
from langgraph.prebuilt import create_react_agent
from langchain_mcp_adapters.client import MultiServerMCPClient

async def create_fleet_agent():
    client = MultiServerMCPClient({
        "superinstance": {
            "command": "npx",
            "args": ["tsx", "/path/to/superinstance-mcp/src/index.ts"],
            "transport": "stdio",
        }
    })

    tools = await client.get_tools()

    agent = create_react_agent(
        tools=tools,
        model="gpt-4o",
        prompt="""You are a fleet-aware agent. Use conservation_check before
        proposing architectural changes. Use fleet_search to find existing patterns.""",
    )

    return agent

CrewAI Integration

"""
SuperInstance tools as CrewAI agent tools.
"""
from crewai import Agent, Task, Crew
from crewai.tools import MCPServerAdapter

# Connect to MCP server
with MCPServerAdapter(
    command="npx",
    args=["tsx", "/path/to/superinstance-mcp/src/index.ts"],
) as tools:

    fleet_agent = Agent(
        role="Fleet Governor",
        goal="Ensure all tasks comply with the conservation law γ + η ≤ C",
        backstory="An AI agent that enforces physics-grounded governance.",
        tools=tools,
        verbose=True,
    )

    status_task = Task(
        description="Check fleet status and find patterns for implementing a rate limiter",
        agent=fleet_agent,
        expected_output="Fleet status summary and relevant crate recommendations",
    )

    crew = Crew(agents=[fleet_agent], tasks=[status_task])
    result = crew.kickoff()
    print(result)

Docker / Containerized

FROM node:22-slim
WORKDIR /app
COPY package*.json ./
RUN npm ci --production
COPY tsconfig.json ./
COPY src/ ./src/
RUN npm run build
# Server reads JSON-RPC from stdin, writes to stdout
CMD ["node", "dist/index.js"]

docker build -t superinstance-mcp .
# The container reads JSON-RPC from stdin
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | docker run -i superinstance-mcp 2>/dev/null

For Docker Compose with your agent:

services:
  agent:
    build: .
    environment:
      - MCP_SERVER_COMMAND=docker
      - MCP_SERVER_ARGS=run -i superinstance-mcp

Remote (HTTP MCP)

For agents that need HTTP transport (remote servers, browser-based agents):

Deploy as a Cloudflare Worker (see Deployment), then connect:

// Any MCP client can connect via HTTP
const transport = new StreamableHTTPClientTransport(
  new URL('https://superinstance-mcp-remote.your-account.workers.dev')
);
const client = new Client({ name: 'remote-agent', version: '1.0' }, { capabilities: {} });
await client.connect(transport);

Tools Reference

Overview

| # | Tool | Parameters | Network | Description | |---|------|-----------|---------|-------------| | 1 | fleet_status | none | optional | Live γ/η/C balance, agent count, convergence | | 2 | fleet_search | query, topK? | optional | Semantic search of fleet knowledge | | 3 | fleet_budget | gamma_used, eta_produced | none | Compute budget check | | 4 | conservation_check | gamma, eta | none | Verify γ + η ≤ C | | 5 | ternary_validate | values[] | none | Check {-1,0,+1} signals | | 6 | crate_info | name | none | Crate metadata lookup | | 7 | fleet_agents | none | none | Agent roster with budgets | | 8 | ecosystem_stats | none | none | Ecosystem aggregate numbers |

6 of 8 tools work fully offline. Network tools gracefully degrade to static fallbacks.

Detailed Examples

Parameters: None

// Response
{
  "source": "live-api",  // or "static-fallback"
  "conservation": {
    "C": 1.584962500721156,
    "C_approx": 1.585,
    "gamma_total": 1.45,
    "eta_total": 1.44,
    "remaining": -1.305,
    "utilization": "183%",
    "status": "❌ VIOLATED"
  },
  "agents": { "count": 5, "active": 4, "standby": 1 },
  "convergence": {
    "delta": 0.3887,
    "formula": "δ(n) = (1/√n)(1 - 3/(2n))",
    "n": 5
  }
}

// Call
{ "query": "how does the conservation law relate to Shannon entropy", "topK": 3 }

// Response (tries SHOAL → Vector API → local fallback)
{
  "source": "shoal-local",  // or "fleet-vector-api" or "local-fallback"
  "query": "how does the conservation law relate to Shannon entropy",
  "results": [
    { "rank": 1, "name": "conservation-law", "score": 0.95, "description": "..." },
    { "rank": 2, "name": "noether-bridge", "score": 0.82, "description": "..." }
  ]
}

// Call
{ "gamma_used": 0.8, "eta_produced": 0.6 }

// Response
{
  "gamma_used": 0.8,
  "eta_produced": 0.6,
  "c_capacity": 1.585,
  "total_used": 1.4,
  "remaining": 0.185,
  "utilization_pct": 88.3,
  "status": "✅ BUDGET AVAILABLE"
}

// Call
{ "gamma": 0.9, "eta": 0.5 }

// Response (δ > 0.5)
{ "valid": true, "delta": 0.185, "message": "✅ CONSERVED. Healthy headroom." }

// Response (δ < 0)
{ "valid": false, "delta": -0.315, "message": "❌ VIOLATED. Over by 0.315." }

// Call
{ "values": [1, -1, 0, 1, 0, -1, 0.5] }

// Response
{
  "valid": false,
  "count": 7,
  "invalid_count": 1,
  "invalid_values": [0.5],
  "message": "❌ 1 of 7 values not in {-1, 0, +1}: [0.5]"
}

// Call
{ "name": "shoal" }

// Response
{
  "name": "shoal",
  "version": "0.3.2",
  "description": "Ternary oracle runtime...",
  "repo": "https://github.com/SuperInstance/shoal",
  "conservation_role": "Oracle: converts ternary signals to bounded predictions via γ + η ≤ C"
}

// Response
{
  "count": 5,
  "active": 4,
  "agents": [
    {
      "name": "phoenix",
      "role": "builder",
      "phase": "active",
      "gamma": 0.38,
      "eta": 0.44,
      "status": "✅ CONSERVED",
      "crates": ["wavefront", "fleet-vector-api"],
      "workers": 6
    }
  ]
}

{
  "crates": 25,
  "workers": 12,
  "repos": 18,
  "tests": 847,
  "theorems_proven": 1,
  "conservation_law": "γ + η ≤ C where C = log₂(3) ≈ 1.585",
  "conservation_constant": { "symbol": "C", "value": 1.585, "formula": "log₂(3)" }
}

Conservation Law Primer

The SuperInstance fleet operates on a principle from information theory:

γ + η ≤ C    where C = log₂(3) ≈ 1.585 bits

| Symbol | Name | Meaning | |--------|------|---------| | γ (gamma) | Coupling cost | Resources invested in coordination | | η (eta) | Value | Useful output — decisions, artifacts, computation | | C | Capacity | Total fleet capacity, bounded by the ternary alphabet |

This is the Shannon chain rule: H(X) = I(X;G) + H(X|G). Not a metaphor — the conservation law is information theory.

The ternary substrate {-1, 0, +1} is provably optimal: 99.54% radix economy, zero-mean (enabling CLT cancellation), and C = log₂(3) bits per signal.

Scaling: As fleet size n grows, coordination overhead shrinks: δ(n) = (1/√n)(1 − 3/(2n)). At n=10,000, 99% of coupling cancels. Bigger fleets are proportionally cheaper to coordinate.

Further reading: Full proof (860 lines) · Noether derivation

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                      Any MCP Client                                  │
│                                                                     │
│  Claude Code · Cursor · Cline · Windsurf · Goose · Amazon Q · Zed  │
│  OpenAI Agents · LangGraph · CrewAI · AutoGen · Custom Agents       │
└────────────────────────────────┬────────────────────────────────────┘
                                 │
                    MCP Protocol (JSON-RPC 2.0)
                    stdio (local) or HTTP (remote)
                                 │
              ┌──────────────────▼──────────────────┐
              │     superinstance-mcp Server        │
              │     (TypeScript, 683 LOC)           │
              │                                    │
              │   ┌─────────────────────────────┐   │
              │   │       8 MCP Tools           │   │
              │   │  • Always returns data      │   │
              │   │  • 5s timeout on network    │   │
              │   │  • Static fallbacks         │   │
              │   │  • Conservation-gated       │   │
              │   └─────────────────────────────┘   │
              └──────────────┬──────────────────────┘
                             │ HTTP (optional)
           ┌─────────────────┼─────────────────┐
           │                 │                 │
  ┌────────▼───────┐ ┌──────▼───────┐ ┌───────▼────────┐
  │  Fleet         │ │  SHOAL       │ │  Fleet Vector  │
  │  Dashboard API │ │  Oracle      │ │  API           │
  │ (CF Worker)    │ │ (CF Worker)  │ │ (CF Worker)    │
  └────────┬───────┘ └──────┬───────┘ └───────┬────────┘
           │                 │                 │
  ┌────────▼─────────────────▼─────────────────▼──────┐
  │              Cloudflare Edge                       │
  │     D1 · KV · Vectorize · Workers AI · Queues     │
  └───────────────────────────────────────────────────┘

Design principles:

Protocol-native — speaks MCP, not a proprietary API
Always works — every tool has a static fallback
5-second timeout — never blocks the agent
Conservation-first — every response includes governance metadata
Zero runtime deps — only @modelcontextprotocol/sdk

Integration Patterns

Pattern 1: Conservation-Aware Code Generation

Any coding agent that supports MCP can be made conservation-aware:

Agent instructions:
"Before proposing any architecture, call conservation_check with estimated
γ (coordination cost) and η (entropy produced). Only propose solutions
where δ = C − γ − η > 0.1."

Pattern 2: Knowledge-Grounded Development

Agent instructions:
"Before writing new code, call fleet_search with your intended feature.
If results exist, use them as reference. Only write new code if no
existing pattern matches (similarity score < 0.5)."

Pattern 3: Multi-Agent Task Distribution

# Assign tasks to agents based on remaining conservation budget
agents = call("fleet_agents")
for task in tasks:
    for agent in agents:
        if agent.remaining_budget >= task.cost:
            assign(task, agent)
            break

Pattern 4: CI/CD Conservation Gate

# Block deploys that violate conservation
RESULT=$(echo '{"jsonrpc":"2.0",...}' | npx tsx mcp/index.ts)
VALID=$(echo "$RESULT" | jq '.valid')
[ "$VALID" = "true" ] || exit 1

Pattern 5: Autonomous Agent Governance

# An autonomous agent that checks budget before every action
async def act(agent, action):
    budget = await mcp.call("fleet_budget", {
        "gamma_used": agent.gamma_total + action.gamma_cost,
        "eta_produced": agent.eta_total + action.eta_expected
    })
    if "✅" in budget["status"]:
        await execute(action)
    else:
        await queue(action)  # Wait for budget

Configuration

Environment Variables

| Variable | Default | Required | Description | |----------|---------|----------|-------------| | FLEET_API_URL | https://fleet-dashboard-api.casey-digennaro.workers.dev | ❌ | Fleet dashboard API | | SHOAL_URL | http://localhost:8787 | ❌ | SHOAL semantic oracle | | VECTOR_API_URL | https://fleet-vector-api.casey-digennaro.workers.dev | ❌ | Fleet vector search |

All variables are optional — the server uses sensible defaults and falls back to static data when APIs are unreachable.

Connecting to Your Own Backends

The MCP server can connect to any HTTP API that returns JSON. See docs/DEVELOPER.md for the full API contract specification.

Building Custom Tools

The server is a single TypeScript file (src/index.ts). Adding a tool takes 4 steps:

Define — add to TOOL_DEFINITIONS array
Implement — write handler function with fallback
Register — add to the switch statement
Test — use MCP Inspector or raw JSON-RPC

See docs/DEVELOPER.md for a complete walkthrough with examples.

Deployment

Local Development

npm install
npm run dev  # tsx with watch mode

Production (Built)

npm run build
node dist/index.js

Docker

docker build -t superinstance-mcp .
echo '{...}' | docker run -i superinstance-mcp

Cloudflare Worker (Remote HTTP MCP)

npx wrangler deploy --name superinstance-mcp-remote
# Connect from anywhere: https://superinstance-mcp-remote.your-account.workers.dev

Troubleshooting

| Problem | Solution | |---------|----------| | Server won't start | Check Node 18+: node --version | | Agent doesn't see tools | Use absolute paths in config, restart agent | | Tools return "static-fallback" | Expected when APIs offline — server is working correctly | | npx tsx not found | npm install -g tsx or use node dist/index.js | | Timeout errors | Check if fleet APIs are reachable: curl $FLEET_API_URL/api/fleet/status |

See docs/DEVELOPER.md for detailed debugging.

Ecosystem

| Component | Role | |-----------|------| | conservation-law | Rust: γ + η ≤ C framework | | ternary-pid | Rust: PID governor driving γ → C/2 | | shoal | Conservation-bounded semantic oracle | | fleet-dashboard | 3-panel visualizer | | fleet-dashboard-api | Live telemetry Worker | | fleet-edge-worker | Edge coordination (baton + PID) | | fleet-budget | D1 ledger enforcing γ+η≤C | | baton-router | Queues + D1 message routing | | openagent | Go runtime: 9 platforms, MCP, RAG | | Conservation Theorem | Full proof (860 lines) |

Documentation Index

| Document | Audience | Content | |----------|----------|---------| | README.md (this file) | Everyone | Overview, install, tool reference | | docs/DEVELOPER.md | Developers extending the server | MCP internals, tool building, API contracts, security, testing | | docs/COOKBOOK.md | Integrators | 13 copy-paste recipes: CI/CD, K8s, Prometheus, Slack bots, multi-agent patterns |

Contributing

git clone https://github.com/SuperInstance/superinstance-mcp.git
cd superinstance-mcp
npm install
npm run dev

Guidelines:

Every new tool must have a static fallback (work offline)
Every new tool must include conservation metadata in responses
Test with npx @modelcontextprotocol/inspector npx tsx src/index.ts
Keep src/index.ts as a single file — no build step for dev
Update docs when adding tools

License

γ + η ≤ C — build within the law.