MCP 服务器

ABIS MCP Server

Enables real-time AI behavioral drift monitoring and detection for Claude Code, tracking shifts in model behavior across 11 major models with tools for scorecards, version changes, and compliance reports.

README

<h1 align="center">ABIS MCP Server</h1>

AI Behavioral Drift Monitoring for Claude Code

37 tools · 11 models · 3 tiers · Real-time drift detection · Patent-backed

What is ABIS?

ABIS (Adaptive Behavioral Intelligence System) monitors LLM behavioral drift in real-time. It detects when AI models silently change — shifts in tone, reasoning depth, safety calibration, or capability — even when providers don't announce updates.

This MCP server gives Claude Code native access to:

Live drift scores across 11 major models (GPT-4o, Claude, Gemini, DeepSeek)
9-score behavioral scorecards (drift, coherence, entropy, alignment, and more)
Silent version change detection — catches updates providers don't announce
Immune threat assessment — prompt injection, coordinated attacks, organic drift
Correction simulation — test fixes before applying them
Compliance reporting — EU AI Act, SOC 2, and investor briefs

Backed by Patent GB2521226.7 · Built by CIJ Labs

Quick Start

One-line install (Claude Code)

/plugin install abis-drift-monitor

Manual setup

Add to your .mcp.json:

{
  "mcpServers": {
    "abis": {
      "type": "sse",
      "url": "https://abis-mcp-production.up.railway.app/sse"
    }
  }
}

No API key required for the Free tier. Connects to the ABIS cloud instantly.

Access Tiers

Tier	Tools	Key Required	What you get
Free	15	None	Drift scores, scorecards, version changes, reports
Pro	27	`MCP_PRO_KEY`	+ Scoring pipeline, memory analysis, immune system, correction simulation
Admin	37	`ABIS_ADMIN_KEY`	+ Benchmarking, stress tests, config changes, pipeline stats

<details> <summary>Pro/Admin connection</summary>

{
  "mcpServers": {
    "abis": {
      "type": "sse",
      "url": "https://abis-mcp-production.up.railway.app/sse",
      "headers": {
        "Authorization": "Bearer YOUR_KEY_HERE"
      }
    }
  }
}

</details>

37 Tools

Free Tier (15 tools)

Tool	Description
`abis_ambient_status`	Quick snapshot of all models — call at session start
`abis_get_drift_scores`	Live drift scores for all 11 monitored models
`abis_get_scorecard`	Full 9-score breakdown (drift, coherence, entropy, alignment, etc.)
`abis_whisper`	Contextual status line (clear / monitor / alert / critical)
`abis_drift_gate`	Pre-response drift interceptor (Layer 3)
`abis_observe`	Post-response observer — how ABIS learns (Layer 3)
`abis_historical_drift`	Drift time-series by model or date
`abis_version_changes`	All detected silent version changes across models
`abis_memory_versions`	Silent version change detection history per model
`abis_memory_compare`	Cross-model stability ranking
`abis_corrector_strategies`	List all 8 correction strategies with severity levels
`abis_nervous_status`	Event bus health, registered handlers, recent event log
`abis_monitoring_config`	Current phase, tier assignments, cost estimate
`abis_ears_status`	EARS webhook sources: Datadog, PagerDuty, Zendesk, Intercom
`abis_generate_report`	5 report types: daily_summary, drift_alert, migration_report, compliance_doc, investor_brief

Pro Tier (+12 tools)

Tool	Description
`abis_score_message`	Score any text through the behavioral pipeline
`abis_conversation_history`	Scored conversation log with drift per turn
`abis_memory_profile`	Full 30-day behavioral memory for a model
`abis_memory_trend`	Time-series trend for any score dimension
`abis_immune_assess`	Classify threat: CLEAN / ORGANIC_DRIFT / PROMPT_INJECTION / COORDINATED_ATTACK
`abis_immune_scan`	Historical attack scan for a model
`abis_simulate`	Test a correction before applying it
`abis_corrector_history`	Which strategies work best per model
`abis_cas_trends`	What features drive drift (energy distance, entropy, coherence)
`abis_cost_report`	Estimated monthly API cost breakdown by model and tier
`abis_cost_vs_drift`	30-day cost vs drift matrix
`abis_run_benchmark`	Single or full benchmark run

Admin Tier (+10 tools)

Tool	Description
`abis_api_cost_analysis`	Most-called models with drift trajectory over 7 days
`abis_ears_correlations`	Drift vs business impact correlations
`abis_sdk_sessions`	View SDK customer session history
`abis_notification_history`	Recent alert delivery log
`abis_notification_config`	Get/set alert preferences (email, Slack, webhook)
`abis_pipeline_stats`	Ingestion statistics, vector counts
`abis_training_status`	XGBoost AUC scores, retraining history
`abis_baseline_recalibrate`	Recompute drift baselines for all or one model
`abis_monitoring_phase`	Switch phase (BASELINE / AGGRESSIVE / SUSTAIN / SCALE)
`abis_stress_benchmark`	30 adversarial prompts across all models

Included Skills

After installing, you get two built-in skills:

/abis-ops — Daily health check: drift snapshot, version changes, stability ranking, cost summary
/abis-alerts — Investigate a drift alert: scorecard deep-dive, CAS trends, threat classification, correction recommendation

Models Monitored

Provider	Models
OpenAI	gpt-4o, gpt-5.2, gpt-5.2-instant
Anthropic	claude-opus-4.6, claude-sonnet-4.5, claude-haiku-4.5
DeepSeek	deepseek-v3.2, deepseek-r1
Google	gemini-2.5-pro, gemini-2.5-flash, gemini-3-flash

Try It

"What is the current drift score for GPT-4o?"
"Run ABIS daily check"
"Has Claude Sonnet drifted in the last 7 days?"
"Generate an EU AI Act compliance report"
"Which model is most stable right now?"
"Have any models had silent version updates this week?"
"Show me the cost vs drift matrix"

Self-Hosted Setup

git clone https://github.com/cijlabs/abis-mcp
cd abis-mcp
npm install
cp .env.example .env
# Set ABIS_ADMIN_KEY, MCP_PRO_KEY, and RAILWAY_URL in .env
node index.js

Then add to Claude Code:

{
  "mcpServers": {
    "abis": {
      "command": "node",
      "args": ["/path/to/abis-mcp/index.js"],
      "env": {
        "ABIS_ADMIN_KEY": "your_admin_key_here"
      }
    }
  }
}

Architecture

abis-mcp/
├── index.js          # MCP server — SSE + stdio dual transport
├── config.js         # Environment config
├── db.js             # MongoDB persistence
├── api.js            # Railway API proxy with circuit breaker
├── tools/            # 37 tool implementations (one per file)
└── skills/           # /abis-ops and /abis-alerts

Dual transport: SSE (HTTP for cloud) + stdio (Claude Desktop/Code)
Rate limiting: 60 req/min global; per-tier limits (free=30, pro=120, admin=300)
Backend: Proxies to ABIS FastAPI on Railway
Persistence: MongoDB Atlas (sessions, scores, conversations, benchmarks)

Contributing

Contributions welcome! Please open an issue first to discuss what you'd like to change.

npm run dev    # Start with --watch on port 3100

License

MIT — see LICENSE for details.

About

CIJ Labs Ltd · Company No. 16984759 · London Patent: GB2521226.7 · abis.cijlabs.com