promptspeak-mcp-server

Pre-execution governance for AI agents. Blocks dangerous tool calls before they execute.

AI agents call tools (file writes, API requests, shell commands) with no validation layer between intent and execution. A prompt injection, hallucinated argument, or drifting goal can trigger irreversible actions. PromptSpeak intercepts every MCP tool call, validates it against deterministic rules, and blocks or holds risky operations for human approval — in 0.1ms, before anything executes.

When to use this

• You run AI agents that call tools (MCP servers, function calling, tool use) and need a governance layer between the agent and the tools.

• You need human-in-the-loop approval for high-risk operations (production deployments, financial transactions, legal filings).

• You want to detect behavioral drift — an agent gradually shifting away from its assigned task.

• You need an audit trail of every tool call an agent attempted, whether it was allowed or blocked.

• You operate in a regulated domain (legal, financial, healthcare) where agent actions must be deterministically constrained.

Install

Claude Code

Add to ~/.claude/settings.json (or project-level .claude/settings.json):

{
  "mcpServers": {
    "promptspeak": {
      "command": "npx",
      "args": ["promptspeak-mcp-server"]
    }
  }
}

Restart Claude Code. All 56 governance tools are immediately available.

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "promptspeak": {
      "command": "npx",
      "args": ["promptspeak-mcp-server"]
    }
  }
}

As a library

npm install promptspeak-mcp-server

From source

git clone https://github.com/chrbailey/promptspeak-mcp-server.git
cd promptspeak-mcp-server
npm install && npm run build
npm start

Usage examples

1. Validate a governance frame before execution

// Request: validate that a frame is structurally and semantically correct
{
  "name": "ps_validate",
  "arguments": {
    "frame": "Ξ⊕●▶α",
    "validationLevel": "full"
  }
}

// Response: validation report with confidence score
{
  "valid": true,
  "parseConfidence": 1.0,
  "summary": { "errors": 0, "warnings": 3, "passed": 14 }
}

2. Execute a tool call under governance

// Request: execute a file write under PromptSpeak governance
{
  "name": "ps_execute",
  "arguments": {
    "agentId": "code-assistant",
    "frame": "Ξ⊕●▶α",
    "action": {
      "tool": "write_file",
      "arguments": { "path": "/tmp/output.txt", "content": "Hello world" }
    }
  }
}

// Response: execution result with governance metadata
{
  "allowed": true,
  "driftScore": 0.02,
  "executionTime": "0.12ms",
  "auditId": "exec_a1b2c3"
}

3. Hold a risky operation for human approval

// When an agent attempts a dangerous action, it's automatically held:
{
  "name": "ps_hold_list",
  "arguments": {}
}

// Response: pending holds awaiting human review
{
  "holds": [{
    "holdId": "hold_001",
    "agentId": "devops-agent",
    "tool": "modify_infrastructure",
    "severity": "critical",
    "reason": "human_approval_required",
    "state": "pending"
  }],
  "count": 1
}

// Approve or reject:
{ "name": "ps_hold_approve", "arguments": { "holdId": "hold_001", "reason": "Reviewed and safe" } }
{ "name": "ps_hold_reject",  "arguments": { "holdId": "hold_001", "reason": "Too risky" } }

4. Halt a drifting agent immediately

// Request: emergency halt — trips circuit breaker, blocks all future calls
{
  "name": "ps_state_halt",
  "arguments": {
    "agentId": "rogue-agent",
    "reason": "Drift score exceeded threshold (0.85)"
  }
}

// Resume after investigation:
{
  "name": "ps_state_resume",
  "arguments": {
    "agentId": "rogue-agent",
    "reason": "Root cause identified and fixed",
    "resetMetrics": true
  }
}

How it works: 9-stage validation pipeline

Every tool call passes through this pipeline. If any stage fails, execution is blocked.

Agent calls tool
  │
  ├─ 1. Circuit Breaker ──── Halted agents blocked instantly (no further checks)
  ├─ 2. Frame Validation ─── Structural, semantic, and chain rule checks
  ├─ 3. Drift Prediction ─── Pre-flight behavioral anomaly detection
  ├─ 4. Hold Check ────────── Risky operations held for human approval
  ├─ 5. Interceptor ───────── Final permission gate (confidence thresholds)
  ├─ 6. Security Scan ─────── Scans write actions for vulnerabilities (see below)
  ├─ 7. Tool Execution ────── Only reached if all 6 pre-checks pass
  ├─ 8. Post-Audit ────────── Confirms behavior matched prediction
  └─ 9. Immediate Action ──── Halts agent if critical drift detected post-execution

Stages 1-6 are pre-execution — the tool never runs if any check fails. Stages 8-9 are post-execution — they detect drift and can halt the agent for future calls.

Security scanning

When an agent writes code (write_file, edit_file, create_file, patch_file), the content is scanned against 10 detection patterns before execution. Severity determines enforcement:

What works (tested)

All claims below are backed by passing tests (104 tests across 5 test files):

• Pattern detection works. Each of the 10 patterns is tested for true positives AND false positives. Example: api_key = "sk-1234567890abcdef" is caught; API_KEY = process.env.API_KEY is not. SQL injection catches \SELECT * FROM users WHERE id = ${id}` but not parameterized queries (db.query("SELECT * FROM users WHERE id = ?", [id])`).

• Severity enforcement works. Critical findings block execution. High findings hold for review. Medium findings warn but allow. Tested end-to-end through the interceptor pipeline.

• Only write actions are scanned. read_file and other non-write actions pass through without scanning, even if their arguments contain vulnerable code. Tested.

• Runtime configuration works. Patterns can be enabled/disabled and severity can be changed at runtime via ps_security_config. A disabled pattern stops firing immediately. Changing a pattern from medium to critical makes it block instead of warn. Tested end-to-end.

• Performance is fine. 100-line file scans complete in under 10ms. Tested.

• Multiple findings in one file work. A file with 6 different vulnerability types correctly classifies each into the right severity bucket. Tested.

What does NOT work yet

• No hold queue integration for HIGH findings. FIXED. HIGH-severity security findings now create real holds in HoldManager via security_finding HoldReason. They appear in ps_hold_list and can be approved/rejected through the normal hold flow.

• No auto-scan on ps_execute. The security scan only triggers in the interceptor's intercept() method for direct tool calls. If an agent uses ps_execute (the governed execution path), the scan runs only if the inner tool is a write action AND the content is passed as a top-level arg. Nested argument structures may bypass scanning. Why: ps_execute wraps tool calls in its own argument schema; the scanner checks proposedArgs.content, not deeply nested fields.

• No file-path-based scanning. The scanner only examines content passed as arguments. It cannot scan files already on disk — it doesn't read from the filesystem. Why: The scanner is a pure function that takes a string. Adding filesystem access would change the security model.

• Patterns are regex-based, not AST-aware. The patterns use regular expressions, which means they can't understand code structure. A hardcoded secret inside a test fixture or a SQL injection in a comment will still trigger. False positive rates range from 25-70% depending on the pattern (documented per-pattern). Why: AST parsing would add dependencies and complexity. Regex is fast and good enough for a governance layer that holds for human review rather than silently blocking.

• Partial persistence. Holds and circuit breaker state now persist to SQLite (data/governance.db) and survive server restarts. However, pattern configuration changes (enable/disable, severity changes via ps_security_config) are still in-memory only and reset on restart. Why: Pattern config is lightweight and rarely changed; full config persistence would need a separate config store.

MCP tools (56)

Core governance

Human-in-the-loop holds

Agent lifecycle

Delegation

Configuration

Symbol registry (entity tracking)

Security enforcement

Audit

Architecture

src/
├── gatekeeper/       # 8-stage validation pipeline (core enforcement)
│   ├── index.ts      #   Pipeline orchestrator + agent eviction policy
│   ├── validator.ts  #   Frame structural/semantic/chain validation
│   ├── interceptor.ts#   Permission gate with confidence thresholds
│   ├── hold-manager.ts#  Human-in-the-loop hold queue
│   ├── resolver.ts   #   Frame resolution with operator overrides
│   └── coverage.ts   #   Coverage confidence calculator
├── drift/            # Behavioral drift detection
│   ├── circuit-breaker.ts  # Per-agent halt/resume
│   ├── baseline.ts         # Behavioral baseline comparison
│   ├── tripwire.ts         # Anomaly tripwires
│   └── monitor.ts          # Continuous monitoring
├── security/         # Security vulnerability scanning
│   ├── patterns.ts   #   10 detection patterns (regex-based)
│   └── scanner.ts    #   Scanner engine + severity classification
├── persistence/      # SQLite governance persistence
│   └── database.ts   #   Holds, decisions, circuit breakers (WAL mode)
├── symbols/          # SQLite-backed entity registry (11 CRUD tools)
├── policies/         # Policy file loader + overlay system
├── operator/         # Operator configuration
├── tools/            # MCP tool implementations
│   ├── registry.ts   #   29 core tools
│   ├── ps_hold.ts    #   5 hold tools
│   └── ps_security.ts#   3 security tools
├── handlers/         # Tool dispatch + metadata registry
├── core/             # Logging, errors, result patterns
└── server.ts         # MCP server entry point (stdio transport)

Performance

Requirements

• Node.js >= 20.0.0

• TypeScript 5.3+ (build from source)

• No external services required — SQLite for symbols and governance persistence

Related Projects

• deeptrend — Structured AI trend feed for autonomous agents. Curated from 14+ sources, synthesized via LLM Counsel, published every 6h as JSON Feed, RSS, and llms.txt. Designed as a data source for agent monitoring pipelines.

Privacy Policy

https://promptspeak.admin-as-a-service.com/privacy

PromptSpeak does not collect personal data, has no telemetry, and stores all governance data locally in SQLite. See the full policy at the link above.

License

MIT