MCP Engine · Execution layer

The engine
that runs your skills.

The MMC MCP Server reads your SKILL.md files from GitHub, executes the business logic, and connects to your tools — all without sending your data to any MMC database.

Runs in your environment · Your logic stays in your repository

Three layers. One standard.

The MMC architecture separates authoring (proprietary Workbench), logic storage (open SKILL.md standard), and execution (MCP Server) into three independent, auditable layers.

Proprietary / SaaS (Workbench)
Standard (SKILL.md)
Execution (MCP Server)
AI Model

Every layer is
independently auditable.

Layer 1 · Proprietary SaaS
MMC Workbench
The visual modeling environment where business teams design context models and generate SKILL.md files. Governed, versioned, human-in-the-loop.
Outcome-driven context modeling canvas
SOP / process document import
SKILL.md generation & GitHub publish
Human-in-the-loop governance controls
Layer 2 · Open Standard
SKILL.md Format
A plain-text, open specification for encoding business context, decision rules, and event flows. Stored in your own GitHub — readable by any AI agent that implements MCP.
YAML frontmatter + markdown structure
Ordered decision-rule encoding (first-match)
Event bus publishing specification
Version-controlled in your own repo
Layer 3 · Execution
MMC MCP Server
The execution engine. Syncs SKILL.md from GitHub and executes from a local copy, parses business rules, runs the event-slicing logic, and calls your tool connectors.
SKILL.md parser & executor
Event-slicing logic
Secure tool connectors (Xero, Slack, Jira…)
Runs in your own infrastructure

What happens when
your AI calls a skill.

1
AI Agent
Requests skill execution via MCP
Claude, Gemini, or any MCP-compatible agent sends a tool call to the MMC MCP Server — either triggered by user input or by the event bus dispatching to the next skill in the flow.
# AI sends tool call
tool: slice-1-receive-request
role: claims-processor
2
MCP Server
Executes from a locally synced copy
The server syncs SKILL.md from your repository — on startup and on change — and caches it locally with per-file SHA tracking. Runtime execution reads that local copy, so there's no per-slice GitHub call and no rate-limit exposure. Nothing is cached on MMC infrastructure — the sync target is storage you control.
# Executes from local synced copy
source: your-org/mmc-skills
file: slice-1-receive-request.SKILL.md
3
Parser & Executor
Presents interface, selects the first matching scenario
The parser presents the required user interface, collects facts, then evaluates scenarios in authored order and selects the first whose conditions hold — exactly one outcome per slice. Evaluation runs in parallel for latency, but the winner is chosen by order, so conflicting conditions (e.g. invoice > $10k AND overseas vendor) resolve deterministically.
# First-match evaluation (authored order)
Scenario A: invalid inputs → error
Scenario B: valid → log event
4
Event Bus
Publishes outcome, dispatches next skill
The selected outcome is logged to the event bus via log-event-to-bus. The bus then dispatches to the next skill in the flow — creating a fully traceable, step-by-step audit trail of every AI decision.

Your tools. Your infrastructure.

The MCP Server connects to your existing business systems via secure connectors. All reads and writes happen from infrastructure you control — never routed through MMC servers.

📊
Xero
Finance
Log transactions, create audit entries, and write financial outcomes directly to your Xero instance.
💬
Slack
Communications
Send notifications, alerts and approval requests to the right channels and people at the right moment.
🐙
GitHub API
Repo / Users
Read SKILL.md files, manage versions, and integrate with your existing developer workflows and CI/CD.
🎯
Jira
Tickets
Raise, update and resolve tickets automatically as business events are logged through the skill flow.
🤖
Claude
AI Agent
Anthropic's Claude connects natively via MCP. Skills run within the Claude system prompt context.
✨
Gemini
AI Agent
Google Gemini connects via MCP. The same SKILL.md runs on any compatible model without modification.
🔌
Custom MCP
Your Systems
Any MCP server you or your implementation partner build becomes a connector. Any system your team already uses can integrate.
+
More coming
Roadmap
Salesforce, HubSpot, ServiceNow and more. Or add your own — see below.
Add and maintain your own connectors.

A connector is an ordinary MCP server, written in any language by your team or your implementation partner. Register it in your deployment’s configuration and every step in your workflow can call its tools. You own its code: version it, test it and update it on your own schedule by pointing your deployment at the new version — no change to the engine, and no MMC release to wait for.

Built-in audit trail.

Every execution is logged to a local event store in your deployment, queryable via get-session-events and the events dump. You get bottleneck, error and outcome visibility — it lives in your infrastructure, not an MMC database.

Runs on
your infrastructure.

The MCP Server is a standalone process that runs where you decide — your cloud, your VPC, your region. Your business logic stays in your own GitHub repository, readable and version-controlled. You keep your models, your data and your repository.

Your procurement team already approved GitHub. There is no MMC database to review. No sensitive data touches our infrastructure.

1
Deployed in your environment
The MCP Server runs as your own deployment: in your cloud account, or your implementation partner’s, in the region you choose.
2
Pointed at your GitHub repo
It syncs your SKILL.md files from your own repository with a read-only token, and executes from a local copy.
3
Connected to your AI agent
Your skills are live — fully governed, fully auditable.
Works with
  • Claude Desktop
  • Gemini CLI
  • OpenAI Codex
  • Any MCP-compatible agent
Your agent connects to the engine over the Model Context Protocol. We set up the deployment with you during the pilot.
No vendor database · Your infrastructure · Your repository

Your logic. Your infrastructure.
Keep your models and data.

The MCP Engine runs in your environment and executes the SKILL.md files in your own repository. Start with the Workbench to generate your first SKILL.md, then connect the engine to any AI agent.

Works with Claude, Gemini & any MCP-compatible agent · Your GitHub, your data