pm-mcp-servers

126 MCP tools across 18 modules for UK government IPA Gate Review assurance.

Part of the PDA Platform. Connects Claude and other AI assistants to schedule data, risk registers, earned value, benefits realisation, gate readiness, portfolio health, calibration evaluation, four-tier escalation routing, and pre-loaded IPA benchmark data. Implements the nine-layer Verified Autonomy framework (Newman et al., May 2026, DOI 10.5281/zenodo.19096229) end-to-end — deterministic guardrails, groundedness checking, tamper-evident audit chains, conformal prediction intervals, and formally-verified RAG monotonicity.

Install

pip install pm-mcp-servers

Or install the meta-package which pulls in all dependencies:

pip install pda-platform

Connect to Claude Desktop

Add the unified server to claude_desktop_config.json:

{
  "mcpServers": {
    "pda-platform": {
      "command": "pda-platform-server",
      "args": [],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Or use the hosted SSE endpoint directly from Claude.ai:

https://pda-platform-i33p.onrender.com/sse

Modules

Module	Tools	Capability
pm-data	6	Schedule loading, querying, format conversion
pm-analyse	8	Risk identification, forecasting, health scoring, narrative divergence, calibration evaluation
pm-validate	4	Structural, semantic, and NISTA validation
pm-nista	5	GMPP reporting and NISTA submission
pm-assure	29	P1–P14 assurance framework, cross-module red flag scanner, four-tier escalation routing
pm-brm	12	Benefits Realisation Management, outturn forecasting, trajectory tracking
pm-portfolio	5	Cross-project health rollup
pm-ev	2	Earned Value metrics and dashboard
pm-synthesis	2	AI executive health summaries
pm-risk	9	Risk register, heat map, velocity, stale-risk detection
pm-change	5	Change control log and pressure analysis
pm-resource	5	Resource loading, conflicts, and capacity
pm-financial	5	Budget baseline, actuals, and EAC forecasting
pm-knowledge	8	IPA benchmarks, reference class forecasting with conformal bands, pre-mortem
pm-simulation	2	Monte Carlo schedule simulation with conformal P50/P80 intervals
pm-lessons	5	AI lessons extraction from gate reviews/PIRs, systemic pattern analysis
pm-reporting	6	IPA-format gate review summaries, SRO dashboards, board exception reports, PIR templates, UDS export
pm-assumptions	8	Assumption drift detection, confidence scoring, live external signals, AI executive reports
Total	126

Verified Autonomy framework

Every AI-authored response now carries trust-signal annotations alongside the original output:

• _groundedness — token-overlap verdict + ungrounded terms + provenance trail (L6)
• _quality — quality score + potential_hallucinations boolean (L3)
• _calibration — conformal prediction interval (L4, Monte Carlo and reference-class)
• L5 deterministic guardrails gate every AI-authored prose output before it reaches the consumer
• L8 cryptographic audit chains record every decision-producing handler invocation

See docs/verified-autonomy-overview.md for the consumer-facing summary.

Example questions Claude can answer

• "What is the current DCA rating for Project ALPHA and which gate conditions are outstanding?"
• "Run a reference class check on our £240m cost estimate — how does it compare to IPA benchmarks for IT projects, and what's the 80% confidence band?"
• "Which risks in the register are stale or accelerating? Generate pre-mortem questions for Gate 3."
• "Produce an earned value dashboard and interpret SPI and CPI trends."
• "Summarise the benefits realisation status and flag any benefits without an identified owner."
• "Evaluate the calibration of our last quarter's DCA forecasts against actual gate outcomes."
• "Should the assumption-confidence outputs for this project be routed to expert review?"

Documentation

Full documentation, practitioner guides, persona guides, and prompt library at github.com/antnewman/pda-platform/tree/main/docs.

UK Government Compliance

• Transparency: MIT open source, full code visibility
• Accountability: Evidence trails on all AI outputs with confidence scoring, groundedness verdicts, and tamper-evident audit chains
• Human oversight: All outputs are advisory; governance decisions require human review. Four-tier escalation router (route_outputs_to_review) surfaces which outputs need it
• Safety: Documented limitations and model cards for all AI-powered modules; deterministic L5 guardrails block overclaim and template-leak failure modes before they reach consumers

Licence

MIT