MCP server

MCP server

The MCP server is TapPass's control plane, exposed as a Model Context Protocol server.

An operator points their MCP host (Claude Code, Cursor, custom agent, …) at it and configures TapPass through tool calls instead of clicking the Admin UI or running shell commands.

This is the path to a headless TapPass: every action that today requires the UI eventually has an MCP tool. "Add an OPA rule that denies tool X for project Y" becomes a single tool call.

At a glance

Speaks	Model Context Protocol (stdio + SSE transports)
Exposes	tools (mutations) + resources (read-only) covering the full control-plane API
Auth	session token (operator SSO) or scoped admin API key
Where it lives	tappass-mcp (planned package) — wraps the existing /api/v1/admin/* routes
Status	future

What it exposes

The MCP server projects the existing control-plane routes as MCP primitives:

Surface	MCP primitive	Examples
Mutations	tool	create_agent, update_pipeline, add_policy_rule, attach_check_pack, revoke_session, rotate_api_key
Reads	resource	list_agents, get_pipeline, get_compiled_policy, get_audit_trail
Wizards	prompt	onboard_team, setup_compliance_pack, dry_run_policy_change

A pull request to the TapPass server that adds a control-plane route should also expose the equivalent MCP tool — same review, same release.

Why this concept exists

Three drivers:

1. Headless ambition. The product vision is to grow toward CLI + MCP-led configuration, with the UI as the on-ramp rather than the long-term primary surface.
2. Operators who live in their IDE. Many TapPass operators are platform engineers who prefer to drive infrastructure from Claude Code or Cursor, not a separate browser tab.
3. Agentic loops over governance. The Agentic CLI needs a tool surface to drive. The MCP server is that surface.

Surfaces

Persona	Surface	What they do
Operator	Claude Code with tappass-mcp registered	"Show me agents on project X with no compliance pack attached" → resource read; "Attach NIST AI RMF pack to all of them" → tool calls
Operator	Cursor + MCP integration	Same flows from a different host
Custom agent	Direct MCP client	Programmatic governance ops in agent loops (e.g., self-onboarding)

Lifecycle

[install]   Operator runs `tappass mcp install` → writes ~/.claude/mcp.json entry
   ↓
[auth]      First call exchanges session/SSO token for a scoped MCP session
   ↓
[discover]  Host (Claude Code) lists available tools + resources
   ↓
[use]       Operator + agent loop drive TapPass via tool calls
   ↓
[audit]     Every mutation emits an audit row, same as the UI / admin API

Related concepts

• wraps → existing control-plane API (the same routes the Admin UI uses)
• enables ↑ Agentic CLI — gives the LLM-driven config loop a tool surface
• complementary to ↔ SDK (admin) — same operations, different surface
• emits → Audit log — every mutation recorded