ai-safe2-framework

AI SAFE² v3.1 MCP Server

Live framework and CP.5.MCP profile access at the agent-to-tool boundary

AI SAFE² Surface Profile

Framework Home Cross-Pillar Governance AISM NEXUS Skills Dashboard

Profile discovery and technology review

The server implementation is MIT licensed. Framework methodology and taxonomy content retain their framework/documentation designation, CC BY-SA 4.0 where designated. Package metadata does not replace the served content’s notices. See Licensing by component.

The MCP server’s taxonomy/profile access does not itself establish runtime enforcement or product maturity. Use the Technology Contribution Profile to identify actual control effects and claim evidence at each plane. Assurance v1.0 is documentation guidance; no new MCP assessment tool or schema is exposed by this update.

What This Is

The AI SAFE² MCP server exposes the framework taxonomy and MCP security profile to MCP-compatible clients such as Claude Code, Codex, Cursor, and other agent runtimes.

AI SAFE² v3.1 separates two counts deliberately:

The server loads the stable 161-control core taxonomy and overlays the v3.1 MCP profile so MCP-1 through MCP-19 are directly queryable.

v3.1 Security Model

AI SAFE² v3.1 is aligned to MCP 2026-07-28 and does not use protocol-owned session state as the governance identity or authorization boundary.

Governance instead binds to:

The governing CP.5 rule is:

A CP.5 profile MUST NOT bind a control to a construct owned by the protocol it profiles.

server/discover is optional in MCP 2026-07-28 and is not required for AI SAFE² conformance.

Three Enforcement Planes

Plane Traffic Primary concern
North-south Agent to model provider Content, policy, consumption, spend
East-west Agent to agent Identity, delegation, lineage, authority
Agent-to-tool Agent to MCP server or tool Tool authorization, provenance, output trust, economic amplification

The MCP server primarily serves the agent-to-tool plane.

CP.5.MCP Profile

The current MCP profile contains 19 controls:

Range Coverage
MCP-1 through MCP-13 Core server/tool hardening, trust, audit, state, provenance, delegation, and revocation
MCP-14 Extension capability negotiation
MCP-15 Header/body assertion integrity
MCP-16 State-handle binding and lifecycle
MCP-17 MRTR round-trip integrity and replay resistance
MCP-18 Catalog-cache integrity and provenance revalidation
MCP-19 Authorization-chain integrity, intended resource/audience validation, and SSRF boundaries

Canonical profile: ../../00-cross-pillar/cp5_mcp_server_security.md

Machine-readable profile: data/mcp-profile-v3.1.json

MCP-19 and Authentication

Legacy static bearer tokens remain supported for entitlement compatibility, but an opaque token does not contain an audience claim. Static-token possession therefore does not by itself demonstrate MCP-19 audience-validation conformance.

A deployment claiming MCP-19 must validate the intended resource or audience, or establish equivalent resource binding with independently reconstructable evidence.

Use MCP_AUTH_AUDIENCE for deployments that add JWT/OAuth audience validation.

Quick Start

Local stdio

Local stdio runs without a remote bearer token and is scoped to the local process boundary.

cd skills/mcp
pip install -e .
MCP_TRANSPORT=stdio python -m mcp_server.app

Example Claude Code configuration:

{
  "mcpServers": {
    "ai-safe2": {
      "command": "python",
      "args": ["-m", "mcp_server.app"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "PYTHONPATH": "/absolute/path/to/ai-safe2-framework/skills/mcp/src"
      }
    }
  }
}

Example Codex configuration:

[mcp_servers.ai-safe2-local]
command = "python"
args = ["-m", "mcp_server.app"]
env = { MCP_TRANSPORT = "stdio", PYTHONPATH = "/absolute/path/to/ai-safe2-framework/skills/mcp/src" }

Remote streamable HTTP

For remote deployment, terminate TLS in front of the service and configure authentication appropriate to the environment.

MCP_TRANSPORT=streamable-http
MCP_HOST=127.0.0.1
MCP_PORT=8000
TOKENS=legacy_token:free
MCP_AUTH_AUDIENCE=https://mcp.example.com/

Treat TOKENS as a legacy entitlement mechanism. Do not describe it as MCP-19-complete authorization.

Operational settings

Variable Default Effect
MCP_STDIO_TIER pro Tier granted to local stdio sessions (pro or free).
AUTH_FAIL_RATE_LIMIT 20 Failed-auth attempts per client IP per hour. When exhausted, every request from that IP gets 429 before its token is evaluated, so the lockout cannot be used as a guessing oracle.
CODE_REVIEW_MAX_CHARS 100000 Largest code_review submission accepted.
ALLOWED_STDIO_MODULE_PATTERNS (none) Comma-separated entrypoint patterns added to the built-in stdio allowlist. Defaults are never removed.

The server requires mcp>=1.12,<2. MCP SDK 2.x renamed FastMCP and changed other v1 APIs; the pin keeps a fresh pip install startable.

See the v1-to-v2 migration matrix before changing this compatibility boundary. A dependency-only range expansion is not a migration.

What the Server Provides

The server supports framework-oriented capabilities including:

ControlsDB.count() keeps framework and profile counts separate to prevent accidental claims that the framework has 180 controls.

NEXUS Relationship

AI SAFE² specifies what must be governed and enforced. NEXUS is Cyber Strategy Institute’s first-party reference implementation for agent-to-agent and agent-to-tool enforcement.

Organizations may use NEXUS or another implementation that demonstrably satisfies the applicable AI SAFE² controls and required evidence.

The v3.1 NEXUS MCP adapter is currently a fail-closed interface contract. Unimplemented methods raise rather than silently allowing traffic. It should not be represented as production-ready until the enforcement implementation is connected and tested.

See ../../NEXUS/adapters/mcp/.

Compatibility

MCP 2025-11-25 is retained as a legacy compatibility binding for twelve months after the v3.1 release. Legacy Mcp-Session-Id values are treated only as principal-scoped state handles, not identity or authorization boundaries.

Data Model

The MCP server intentionally uses a layered data model:

data/ai-safe2-controls-v3.0.json  -> stable 161-control core taxonomy
data/mcp-profile-v3.1.json       -> CP.5.MCP 19-control protocol profile

The release version is v3.1 even though the unchanged core taxonomy source remains v3.0. This prevents duplicated control databases and makes profile-level expansion explicit.

Testing Expectations

Before a v3.1 release or MCP security change is considered complete, verify at minimum:


Framework Home Cross-Pillar Governance AISM NEXUS Skills Scanner Dashboard

AI SAFE² v3.1 · Cyber Strategy Institute