MCP 服务器

CI-1T Prediction Stability Engine

Prediction stability engine for AI agents. Evaluate model stability, detect ghosts, probe any LLM for instability, monitor fleet drift. 20 tools + 1 resource. Works with Claude, Cursor, VS Code.

README

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)

Tool	Description	Auth
`evaluate`	Evaluate prediction stability (floats or Q0.16)	API key
`fleet_evaluate`	Fleet-wide multi-node evaluation (floats or Q0.16)	API key
`probe`	Probe any LLM for instability (3x same prompt). BYOM mode: bring your own model via OpenAI-compatible API	API key or BYOM
`health`	Check CI-1T engine status	API key
`fleet_session_create`	Create a persistent fleet session	API key
`fleet_session_round`	Submit a scoring round	API key
`fleet_session_state`	Get session state (read-only)	API key
`fleet_session_list`	List active fleet sessions	API key
`fleet_session_delete`	Delete a fleet session	API key
`list_api_keys`	List user's API keys	API key
`create_api_key`	Generate and register a new API key	API key
`delete_api_key`	Delete an API key by ID	API key
`get_invoices`	Get billing history (Stripe)	API key
`onboarding`	Welcome guide + setup instructions	None
`interpret_scores`	Statistical breakdown of scores	None
`convert_scores`	Convert between floats and Q0.16	None
`generate_config`	Integration boilerplate for any framework	None
`compare_windows`	Compare baseline vs recent episodes for drift detection	None
`alert_check`	Check episodes against custom thresholds, return alerts	None
`visualize`	Interactive HTML visualization of evaluate results	None

Resource	URI	Description
`tools_guide`	`ci1t://tools-guide`	Full usage guide: response schemas, chaining patterns, fleet workflow, thresholds, example pipelines

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

Variable	Required	Description
`CI1T_API_KEY`	Yes	Your `ci_...` API key — single credential for all tools
`CI1T_BASE_URL`	No	API base URL (default: `https://collapseindex.org`)

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

Metric	Description
CI (Collapse Index)	Primary stability metric (Q0.16: 0–65535). Lower = more stable
AL (Authority Level)	Engine trust level for the model (0–4)
Ghost	Model appears stable but may be silently wrong
Warn / Fault	Threshold and hard-failure flags

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