Paichart MCP Server

pAIchart — Intent-Driven, Human-Gated Change Synthesis

Across network, cloud, and Kubernetes — on an open MCP hub, tracked as structured delivery. Give pAIchart a one-line objective and it synthesizes a reviewed, approved change package your team applies idempotently — out-of-band, with a rollback. pAIchart designs and reviews the change; it never applies it.

State the intent in plain language; pAIchart's delivery engine harvests your live systems, designs the change, authors the per-vendor config + validation + rollback, and an independent reviewer gates it. What comes back is a change you approve — you don't author it, and nothing is applied until a human says so. The expertise a multi-vendor change used to demand shifts from authoring across every system to approving one reviewed result.

Three layers make that work, and they're one product, not three:

• The engine — synthesizes the change from intent (the Pipeline Harness, below).
• The MCP Hub — how the engine reaches your live systems safely: every harvest is a Hub call, authenticated per-service with no shared secrets. It's also an open registry anyone can self-register a service with, and orchestrate.
• Delivery management (POVs → Phases → Tasks) — where the objectives, tasks, artifacts, and change packages are organized, analyzed, and governed.

Agents and AI clients reach all of it through a single Hub with trust-level authentication and per-user OAuth passthrough.

What pAIchart Does

Change synthesis from intent — the delivery engine

Give pAIchart a one-line objective and it orchestrates a team of specialist agents into a reviewed, decision-grade change — decompose into typed tasks, wire dependencies, chain each agent's full output to the next, quality-gate every step, synthesize the result. You provide direction; the agents provide labor.

A single pipeline handles one domain. A program is a pipeline of pipelines — how one intent spans multiple domains and hands a real value from one to the next. The flagship shape is a sequenced, cross-domain program: e.g. "the switches export from new dedicated addresses; the cloud archive bucket authorises exactly the range covering them" runs the network pipeline first, then feeds its actual derived address range into a cloud-IaC pipeline — the policy authorises exactly what the network design produced, with no human reconciling the two by hand. One intent, configuration in several vendors' languages, none of it hand-written.

Every run ends the same way: a reviewed change package you approve, then apply on your own terms. pAIchart never applies it for you — device/cluster/state access is read-only, output is sanitized before any reasoner reads it, secrets are redacted from the artifact, and apply stays a separate, human-gated step.

• Network Provisioning — turn "add a Loopback0 per switch and advertise it into BGP" into an approved change package the provisioning team applies idempotently: the pipeline self-provisions a read-only device service from a descriptor, harvests the device's real running state, designs the change, authors per-device config + validation + rollback, and an independent reviewer gates it. → example change report
• Kubernetes / GitOps — turn "add an HPA and resource requests/limits to the orders-api Deployment" into a declarative GitOps change package (a kustomize overlay) from live cluster state, with offline validation (kubeconform / kustomize build / OPA — never kubectl diff) and rollback. Read-only + RBAC-scoped; secret names surface, values never leave the cluster. Apply is a GitOps-reconcile / human-gated step. → example change report (includes an earned NEEDS-REVISION — the reviewer refusing to approve what it couldn't verify)
• Terraform / Cloud IaC — turn "add versioning and a public-access-block to the acme-app-logs S3 bucket" into an approved HCL change package (a PR) the team applies from real Terraform state (a scoped state pull — no providers launched, no state lock), with terraform validate / plan / tflint / OPA expected-facts and rollback. Apply is the team's governed terraform apply. → example change report (shows the layered defense: a secret-shaped tag redacted, a prompt-injection tag refused)
• Artifact Synthesis — turn source material (git history, execution logs, a POV's own delivery history, external MCP services) into a publishable deliverable (case study, post-mortem, quarterly recap) via a harvest → author → review pipeline. → example case study

Every kind runs on the same harness — for the full how-to, see the in-product HOWTO-use-pipeline-harness guide (run list_prompts() in your AI client to find it). For a narrative walkthrough of a real cross-domain program and how its correctness is machine-checked, see the coordinated-infrastructure-change case study.

Reaching your live systems safely — the MCP service hub

The engine's harvest step is a Hub call, and the same machinery is open to anyone: register a service, discover it by capability, and orchestrate it — with per-user identity and no shared API keys in URLs.

• Free Service Registration — Comprehensive guides available via list_prompts() or as MCP resources
• Service Discovery — AI agents find services by capability, not by name
• Multi-Service Workflows — Chain services sequentially, in parallel, or conditionally with variable passing
• Per-User Authentication — Each user's operations run as themselves via External OAuth (validated with Snowflake)
• Trust Level System — 6-tier security model controls token forwarding (INTERNAL → TRUSTED → OWNER → TEAM_MEMBER → SCOPED → ANONYMOUS)
• JWKS Token Validation — RS256 asymmetric cryptography, public-key verification, no shared secrets
• Per-Service Audience Scoping — Hub-minted access tokens carry a per-service audience (RFC 8707 resource indicators): each service receives a short-lived credential scoped to only itself, so a token leaked from one service can't be replayed against another. Services that validate it via JWKS can accept pAIchart-issued identity instead of static API keys in URLs.
• Trustworthy Error-Recovery Signals — When a service call fails, the Hub returns facts an AI client can act on — the honoured timeout, the service's recent success rate, and recovery guidance that never points at a blind health check — rather than unvalidated verdicts that can mislead. Built so the client recovers on its own; see the Error Recovery Signals case study.

Organizing and governing the work — POVs → Phases → Tasks

Where the engine's work lives and is governed: programs and pipelines are typed tasks inside a structured, AI-readable delivery plan, and their change packages, analytics, and history hang off it.

• POVs → Phases → Tasks — run proof-of-value engagements as structured, AI-readable delivery plans
• Natural-language operation — ask "Which of my POVs are at risk?" or "show open tasks for BlackEye" — no UI required
• AI agents on your work — configure, assign, and execute agents against delivery tasks
• Portfolio analytics — health, insights, and execution metrics across your POVs

Get Started

pAIchart is a hosted MCP hub — nothing to install. Point your AI client at the endpoint, authenticate, and start asking in natural language.

• Hub access: https://paichart.app/mcp
• Connect with: Claude Desktop (GitHub OAuth) or ChatGPT (Microsoft OAuth)
• First thing to say: "Help me get started with paichart" — or run list_prompts() to see every guided workflow
• Privacy: PRIVACY-DEMO.md — what a demo account holds, what it can do, 30-day auto-deletion

Once you're connected, try:

• "Provision a Loopback0 per switch and advertise it into BGP" — the delivery engine harvests, designs, authors, and returns a reviewed change package to approve (see HOWTO-use-pipeline-harness)
• "Which of my POVs are at risk?" — delivery analytics, answered directly
• "Discover services" — browse the registry by capability
• "Run the prompt energy_operations_optimizer" — correlates weather forecasts with energy data into operational recommendations, a multi-service workflow across two live services

Under the Hood

Every request is either answered directly, synthesized into a change package by the engine, or composed into a workflow across services — and every external call runs as you, never as a shared platform account:

You (Claude Desktop / ChatGPT)
  → authenticate to the pAIchart Hub
  → state intent in natural language, e.g.
      • "Provision a Loopback0 per switch, advertise into BGP"  → engine: harvest → design → author → review → change package (your team applies)
      • "Which of my POVs are at risk?"                         → project / analytics tools answer directly
      • "Texas energy mix + this week's weather"                → Hub composes a multi-service workflow
  → for external / live-system access (incl. the engine's harvest):
      → Hub discovers services by capability, determines trust level, mints a per-service JWT
      → the external service validates it via JWKS — no shared API keys
  → operations execute as the authenticated user; pAIchart designs and reviews — your team applies, idempotently and out-of-band

Live Services

Register Your MCP Service

New to this? Run the HOWTO-register-service guide (list_prompts() in your AI client to find it) — a step-by-step walkthrough from a basic registration to Grade-A tool schemas, access control, and trust levels.

Any MCP service can register with the Hub in one command:

registry(action: "register", {
  name: "my-service",
  description: "What your service does",
  endpoint: "https://my-service.com/mcp",
  category: "data-services"
})

Services that support External OAuth (like Snowflake, Databricks) get per-user authentication automatically.

Learn

• Case studies — three narrative walkthroughs of the same real cross-domain program (network → cloud): Coordinated Infrastructure Change, Checked by Machine (can you trust it — how correctness is machine-checked, not trusted), You Approve; You Don't Author (what it buys you — one intent, configs in several vendors' languages, none hand-written), and Inside a Multi-Domain Program (how it's built — the DAG, the review tiers, and how far it scales). See case-studies/README.md.
• MCP Tool Excellence — a 12-chapter tutorial series on building MCP tools AI clients can call without external documentation, extracted from pAIchart's own production audits: tutorials/README.md

Links

• Hub (connect here): paichart.app/mcp — Claude Desktop, GitHub OAuth
• JWKS: https://paichart.app/api/auth/jwks
• Documentation: provided as an MCP resource (or run list_prompts()) in your AI client
• Demo User Privacy: PRIVACY-DEMO.md — what a demo account holds, what it can do, 30-day auto-deletion

Keywords

mcp mcp-hub mcp-server mcp-orchestration model-context-protocol ai-native change-synthesis intent-driven human-gated network-provisioning kubernetes gitops terraform infrastructure-as-code multi-domain-automation pipeline-harness autonomous-agents change-management delivery-management proof-of-value pov ai-services service-discovery external-oauth jwks per-service-audience rfc8707 per-user-authentication workflow-orchestration error-recovery mcp-tutorials claude-desktop chatgpt snowflake