CI-1T MCP Server

Version: 1.7.0
Last Updated: February 27, 2026
License: Proprietary

MCP (Model Context Protocol) server for the CI-1T prediction stability engine. Lets AI agents — Claude Desktop, Cursor, Windsurf, VS Code Copilot, and any MCP-compatible client — evaluate model stability, manage fleet sessions, and control API keys directly.

One credential. One env var. That's it.

Tools (20) + Resources (1)

Onboarding

New users get guided setup automatically. If no API key is configured:

• Startup log prints a hint: "Create a free account at collapseindex.org — 1,000 free credits on signup"
• onboarding tool returns a full welcome guide with account status, setup steps, config examples, available tools, and pricing
• Auth-guarded tools return a friendly error with specific setup instructions instead of a raw 401
• Utility tools (interpret_scores, convert_scores, generate_config) always work — no auth, no credits

Every new account gets 1,000 free credits (no credit card required), enough for 1,000 evaluation episodes.

Setup

Environment Variables

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "ci1t": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "collapseindex/ci1t-mcp"],
      "env": {
        "CI1T_API_KEY": "ci_your_key_here"
      }
    }
  }
}

Cursor / Windsurf

Add to .cursor/mcp.json or equivalent:

{
  "mcpServers": {
    "ci1t": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "collapseindex/ci1t-mcp"],
      "env": {
        "CI1T_API_KEY": "ci_your_key_here"
      }
    }
  }
}

VS Code (GitHub Copilot)

Add to .vscode/mcp.json:

{
  "servers": {
    "ci1t": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "collapseindex/ci1t-mcp"],
      "env": {
        "CI1T_API_KEY": "ci_your_key_here"
      }
    }
  }
}

Run from source (no Docker)

git clone https://github.com/collapseindex/ci1t-mcp.git
cd ci1t-mcp
npm install
npm run build

# Set env var and run
CI1T_API_KEY=ci_xxx node dist/index.js

Build Docker Image

docker build -t collapseindex/ci1t-mcp .

Example Usage

Once connected, an AI agent can:

"Evaluate these prediction scores: 45000, 32000, 51000, 48000, 29000, 55000"

The agent calls evaluate with scores: [45000, 32000, 51000, 48000, 29000, 55000] and gets back stability metrics per episode, including credits used and remaining.

"Create a fleet session with 4 nodes named GPT-4, Claude, Gemini, Llama"

"List my API keys"

"Probe this prompt for stability: What is the capital of France?"

"Probe my local Ollama llama3 model with: What is the meaning of life?"

The agent calls probe in BYOM mode — sends the prompt 3x to http://localhost:11434/v1 and scores the responses locally. No CI-1T credits used.

"Interpret these scores: 0.12, 0.45, 0.88, 0.03, 0.67"

The agent calls interpret_scores locally (no API call, no credits) and returns mean, std, min/max, and normalized values. For full stability classification, use evaluate.

"Convert these probabilities to Q0.16: 0.5, 0.95, 0.01"

"Generate a FastAPI integration for CI-1T with guardrail pattern"

CI-1T Quick Reference

Classification labels (Stable / Drift / Flip / Collapse) are determined by the engine. Use the evaluate tool to get exact classifications — thresholds are configurable via the API.

Architecture

┌──────────────────────┐     stdio      ┌───────────────────────┐
│   Claude Desktop /   │◄──────────────►│   ci1t-mcp server     │
│   Cursor / VS Code   │               │   (Node.js / Docker)  │
└──────────────────────┘               └──────────┬────────────┘
                                                   │ HTTPS
                                                   │ X-API-Key
                                    ┌──────────────┼──────────────┐
                                    │              │              │
                               ┌────▼───┐   ┌─────▼────┐  ┌─────▼─────┐
                               │Evaluate│   │Fleet API │  │Dashboard  │
                               │  API   │   │Sessions  │  │API Keys   │
                               │        │   │          │  │Billing    │
                               └────────┘   └──────────┘  └───────────┘
                               collapseindex.org

Changelog

v1.7.0 (2026-02-27)

• BYOM Probe: probe tool now supports Bring Your Own Model mode
• Provide base_url + model (+ optional model_api_key) to probe any OpenAI-compatible endpoint directly
• Works with local models (Ollama, LM Studio, vLLM) and remote APIs (OpenAI, Anthropic, Together, etc.)
• BYOM mode runs entirely locally — no CI-1T auth needed, no credits consumed
• Default mode unchanged (routes through CI-1T backend, costs 1 credit)
• Local similarity scoring: Jaccard, length ratio, and character fingerprint cosine similarity
• 20 tools + 1 resource

v1.6.1 (2026-02-27)

• SEC-01 (Critical): API key generation now uses crypto.randomBytes() instead of Math.random()
• SEC-02 (High): Visualization title is HTML-escaped to prevent XSS
• SEC-03 (Medium): toQ16() decimal heuristic prevents integer arrays [0, 1] from being misclassified as floats
• SEC-04 (Medium): Score arrays capped at 10,000 per stream, 16 nodes max on fleet tools
• SEC-05 (Medium): Source maps disabled in production build
• SEC-06 (Low): Fixed template literal bug in compare_windows severity message
• SEC-07 (Low): Visualization temp files auto-cleaned after 1 hour
• SEC-08 (Low): Header version comment updated

v1.6.0 (2026-02-27)

• AI Discoverability: All 20 tool descriptions now include response schemas and chaining hints
• tools_guide MCP resource (ci1t://tools-guide): comprehensive usage guide with response schemas, chaining patterns, fleet session workflow, classification thresholds, and example pipelines
• Agents can now read the resource for full context beyond individual tool descriptions
• 20 tools + 1 resource

v1.5.0 (2026-02-27)

• compare_windows tool: compare baseline vs recent episodes — drift delta, trend direction, degradation detection
• alert_check tool: check episodes against custom thresholds (CI, EMA, AL, ghost, fault) with severity levels
• Both tools are local computation — no API call, no auth, no credits
• 20 tools total

v1.4.0 (2026-02-27)

• visualize tool: generates self-contained interactive HTML with Canvas 2D bar charts
• Fixed sidebar layout matching CI-1T Lab dashboard style (KPIs, legend, stats in sidebar)
• EMA Trend + Authority Level charts side-by-side
• Adaptive bar sizing, hover tooltips, color-coded classifications
• Links to collapseindex.org in sidebar
• 18 tools total

v1.3.0 (2026-02-27)

• Single credential: All tools now use CI1T_API_KEY — no Bearer token needed
• Removed CI1T_TOKEN env var entirely
• Backend auth unified: all API routes accept X-API-Key (resolves user via key hash)
• Simpler config: one env var to set, one credential to manage
• 17 tools total

v1.2.0 (2026-02-27)

• onboarding tool: welcome guide with account status, setup steps, config examples, pricing, and available tools
• Auth guards on all credentialed tools — returns a structured onboarding message instead of failing at the API level
• Enhanced startup log: new-user hint when no credentials are configured
• 17 tools total

v1.1.0 (2026-02-27)

• 3 new utility tools: interpret_scores, convert_scores, generate_config (local, no auth, no credits)
• evaluate and fleet_evaluate now auto-detect floats (0–1) vs Q0.16 (0–65535) — no manual conversion needed
• Dashboard parity: all Ask AI tools now available via MCP

v1.0.0 (2026-02-25)

• Complete rewrite from Python to TypeScript
• 13 tools: evaluate, fleet_evaluate, probe, health, fleet session CRUD, API key CRUD, invoices
• Docker image distribution
• stdio transport for Claude Desktop, Cursor, VS Code
• Dual auth: API key (X-API-Key) for evaluate, Bearer token for dashboard

© 2026 Collapse Index Labs™ — Alex Kwon
collapseindex.org · ask@collapseindex.org