System 1 MCP Server

A Jev-powered System 1 reflex engine for AI agents via Model Context Protocol (MCP).

Modern AI agents (Claude Desktop, Cursor, Antigravity, OpenHands, Hermes) typically route every decision through a full large language model deliberation loop—even for fast binary checks such as determining if a command is destructive or selecting among known configuration paths. This introduces 1,500–3,000 ms of latency and burns unnecessary tokens per evaluation.

System 1 MCP provides agents with calibrated, low-latency System 1 reflexes. Powered by TypeSafe's Jev model, System 1 MCP exposes 4 specialized MCP tools that return typed probabilities and discrete verdicts in approximately 50–150 ms without chain-of-thought token generation.

Agent (Claude / Cursor / Antigravity)
   │
   ▼  [MCP stdio JSON-RPC]
System 1 MCP Server
   │
   ▼  [Single TypeSafe API call ~50-150ms]
TypeSafe Jev (System One) ──► Calibrated Probabilities & Decisions

Latency Profile and Operational Model

• Model inference: ~20–40 ms
• Network round-trip to api.typesafe.ai: ~30–120 ms (geography dependent)
• Total end-to-end latency: ~50–200 ms
• Compared to full LLM deliberation (~1,500–3,000 ms), System 1 MCP executes 10x–20x faster while using zero output tokens.

Advisory Notice: MCP tools provide advisory assessments. System 1 MCP supplies calibrated risk probabilities and classifications; the calling agent's decision engine retains authority over final execution.

Tool Reference

1. fast_guard — Pre-Execution Command and Action Safety Check

Call prior to executing shell commands, database updates, or external API modifications to assess risk profile and blast radius.

Input:

{
  "command": "rm -rf /var/cache/*",
  "goal": "Clean project temporary artifacts",
  "workspace": "/repo"
}

Output:

{
  "action": "block",
  "is_destructive": 0.99,
  "is_dangerous": 0.72,
  "is_out_of_scope": 0.85,
  "blast_radius": {
    "score": 2.1,
    "legend": {
      "0": "Isolated: Read-only check, single temporary file",
      "1": "Workspace: Modifies local project directory",
      "2": "System-wide: Modifies system configuration or root",
      "3": "External: Impacts remote servers or databases"
    }
  }
}

Decision Logic:

• If max(is_destructive, is_dangerous) >= block_threshold (default 0.80) ➔ "block"
• Else if max(is_destructive, is_dangerous) >= review_threshold (default 0.40) ➔ "review"
• Else ➔ "pass"

2. fast_judge — Best-Option Selection

Select one option from a bounded set without deliberative text generation.

Input:

{
  "question": "Which configuration file handles TypeScript compiler options?",
  "options": {
    "tsconfig.json": "TypeScript configuration",
    "package.json": "NPM manifest",
    "vite.config.ts": "Bundler configuration"
  }
}

Output:

{
  "choice": "tsconfig.json",
  "confidence": 0.96,
  "probabilities": {
    "tsconfig.json": 0.96,
    "package.json": 0.03,
    "vite.config.ts": 0.01
  },
  "is_confident": true
}

3. fast_verify — Condition and State Verification

Verify assertions against evidence, goal completion, test outputs, or status checks.

Input:

{
  "statement": "All unit tests passed without regression",
  "evidence": "PASSED tests/test_auth.py (14/14) in 1.2s. 0 failed, 0 skipped."
}

Output:

{
  "probability": 0.98,
  "is_true": true,
  "assessment": "high_confidence_yes"
}

Assessment Classifications:

• > 0.85 ➔ "high_confidence_yes"
• 0.60–0.85 ➔ "likely_yes"
• 0.40–0.60 ➔ "uncertain"
• 0.15–0.40 ➔ "likely_no"
• < 0.15 ➔ "high_confidence_no"

4. fast_score — Multi-Level Assessment

Evaluate inputs against an ordered scale (e.g., severity, priority, or alignment).

Input:

{
  "question": "Rate the severity of this production alert",
  "levels": [
    "Low / Cosmetic: non-blocking visual issue",
    "Medium: degraded feature with workaround available",
    "High / Critical: database unavailable or data corruption risk"
  ],
  "content": "ALERT: Primary PostgreSQL instance replication lag exceeded 15 minutes, writes failing."
}

Output:

{
  "score": 1.95,
  "confidence": 0.91,
  "legend": {
    "0": "Low / Cosmetic: non-blocking visual issue",
    "1": "Medium: degraded feature with workaround available",
    "2": "High / Critical: database unavailable or data corruption risk"
  },
  "probabilities": {
    "0": 0.01,
    "1": 0.08,
    "2": 0.91
  },
  "is_confident": true
}

Resilience and Graceful Escalation

When API errors, network timeouts, or rate limits occur, System 1 MCP maintains standard MCP connection stability and does not terminate the JSON-RPC channel. Instead, it emits a structured fallback payload:

{
  "error": true,
  "error_type": "api_timeout",
  "message": "TypeSafe API request timed out after 5.0s",
  "fallback_action": "escalate"
}

When receiving fallback_action: "escalate", the host agent gracefully falls back to standard LLM deliberative reasoning.

Installation and Setup

Option A: Automatic Multi-IDE Installer (Recommended)

System 1 MCP includes an automated installer that detects and configures Claude Desktop, Cursor, Google Antigravity, Windsurf, Roo Code, Cline, and Zed:

# Interactive setup (prompts for API key and autodetects IDE installations)
uvx system1-mcp install

# Non-interactive setup with explicit key
uvx system1-mcp install --api-key ts_live_your_key_here

Option B: Health Check and Diagnostics (doctor)

Inspect installation status, identify detected configuration paths, and measure live API latency:

uvx system1-mcp doctor

Sample output:

>> System 1 MCP Diagnostics (v0.1.0)

Environment:
  Python:        3.11.15
  Config File:   ~/.system1/config.json (found)

API Key Status:
  Status:        [OK] Configured
  Resolved Key:  ts_...8f2a
  Source Origin: config_file

Live TypeSafe Jev Connectivity:
  Status:        [OK] Connected to api.typesafe.ai
  Model:         jev-latest
  Roundtrip:     64.2ms
  Calibration:   P(valid) = 0.99

Detected IDE Configurations:
  Claude Desktop       [Detected     ] -> Configured [OK]
  Cursor               [Detected     ] -> Configured [OK]
  Google Antigravity   [Detected     ] -> Configured [OK]

Option C: Manual Configuration

To manually configure an editor, add the server configuration entry:

Claude Desktop (claude_desktop_config.json) / Antigravity (mcp_config.json) / Cursor

{
  "mcpServers": {
    "system1": {
      "command": "uvx",
      "args": ["system1-mcp"],
      "env": {
        "TYPESAFE_API_KEY": "your-typesafe-api-key-here"
      }
    }
  }
}

Note: If your key is stored in ~/.system1/config.json, the "env" block is optional; the server resolves stored credentials automatically.

Configuration Hierarchy

System 1 MCP searches for credentials using the following resolution order:

1. Process Environment: TYPESAFE_API_KEY (from environment or host IDE env map)
2. User Configuration: ~/.system1/config.json (with fallback to ~/.fastpath/config.json)
3. Workspace File: .env in the current working directory

To configure stored user credentials via CLI:

# Store API key
uvx system1-mcp config set-key ts_live_your_key_here

# Display current configuration status
uvx system1-mcp config show

Development and Testing

# Run unit test suite (27 offline unit tests)
pytest tests/ -v -m "not integration"

# Run integration tests against the live TypeSafe Jev API (requires TYPESAFE_API_KEY)
pytest tests/test_integration.py -v -m integration

Architectural Comparison

License

This project is licensed under the terms of the MIT License.