Sane & Controlled AI Agent Access
An MCP client is a planned, separate thin client over Nuncio's gRPC contract — the same one the CLI uses today. It does not exist yet. This page describes the design: aggregate your accounts, control exactly how agents touch them, and audit everything.
Nothing on this page is usable yetWhy a Gateway, Not Direct Access
The design principle: never grant an AI agent uncontrolled credentials to your real accounts.
1. One Local Vault, Not Scattered Credentials
The plan: aggregate every account you connect into one local SQLite vault an agent can query, instead of handing separate credentials to separate AI tools.
2. A Gateway, Not Direct Access
Local LLMs and AI agents are designed to never receive raw passwords or direct, uncontrolled access to your provider accounts. Every action would pass through Nuncio's security enclave under an RBAC policy.
3. Every Action, Audited
The goal is that every agent tool call, read, draft, and send lands in the same immutable, hash-chained WORM audit ledger that already exists for other engine actions today.
What is the Model Context Protocol (MCP)?
The Model Context Protocol (MCP) is an open standard from Anthropic that lets AI applications safely read data and invoke actions on local services over stdio JSON-RPC. Nuncio plans to expose a governed MCP surface once its client architecture is built — think of the examples below as the target experience, not something you can do today.
Eventually: "Find urgent emails from today across all my accounts and summarize them."
Eventually: "Schedule a sync tomorrow afternoon and create the event." Depends on CalDAV, which is also planned.
Eventually: "Create a rule to auto-archive newsletters." Rule creation works today via the CLI; wiring it to an agent is future work.
Target Configuration Shape
This is what an MCP client config would look like once it exists — it will not connect to anything today.
Target Config Shape (`claude_desktop_config.json`)
No `nuncio-mcp` binary exists to run this yet{
"mcpServers": {
"nuncio": {
"command": "nuncio-mcp",
"args": ["--stdio"]
}
}
}Planned Security Enclave
You would control exactly what data and permissions any connected AI agent gets. None of this is built yet.
Toggle fine-grained capabilities per agent: read_mail, send_mail, read_calendar, write_calendar, and manage_filters. Unauthorized operations would be rejected.
Before message text reaches an LLM context window, sensitive identifiers would be masked automatically. This redaction pipeline does not exist yet.
Planned MCP Tool Surface
The tool names an eventual MCP client would expose — none of these are callable by an AI agent today.
List messages with folder, account, and unread filters.
Draft or send email with human-in-the-loop confirmation.
Full-text search across synced mail.
Export MBOX, EML, or JSON data programmatically.
Query the immutable audit log.
Verify the HMAC-SHA256 hash-chain ledger.
Query calendar events, once calendar sync exists.
Create calendar events with conflict detection.
Inspect NSQL filter rules and execution logs.
Compile, validate, and save a new NSQL rule.
Dry-run test an NSQL rule against a sample message.
Inspect connected accounts and protocol status.