MCP FUSION

The TypeScript framework for secure, MCP 2.0-native servers.

MCP Fusion is a TypeScript framework that enforces security at the architectural level of every MCP server. Raw data never reaches the LLM without passing through a typed egress firewall. Tools are physically removed from the agent's namespace when the workflow state forbids them. Every behavioral surface is hashed, locked, and auditable in version control.

The framework ships with a SKILL.md — a machine-readable architectural contract. AI coding agents read the Skill and produce correct, governed servers on the first pass.

MCP 2.0 (2026-07-28) — Full Compliance

MCP Fusion is 100% compatible with MCP 2.0 (protocol revision 2026-07-28). Every feature the spec defines is implemented or handled via the MCP SDK v2. Every feature the spec deprecates is deprecated in MCP Fusion.

Implemented MCP 2.0 Features

Deprecated Features (mirrored from MCP 2.0)

All features deprecated by MCP 2.0 (SEP-2577 / SEP-2596) are deprecated in MCP Fusion with @deprecated markers and migration guidance. They remain functional during the deprecation window (earliest removal: 2027-07-28).

📄 Deprecation Registry — full compliance statement and migration guide.

The Skill — AI Writes the Server

MCP Fusion includes a SKILL.md that encodes the entire MVA architecture, security patterns, and governance rules into a format AI coding agents consume directly.

Open your project in Cursor, Claude Code, GitHub Copilot, or Windsurf and describe what you need:

"Build an MCP server for patient records with Prisma. Redact SSN and diagnosis from LLM output. Gate discharge tools until attending physician signs off."

The agent reads the Skill. It produces defineModel() declarations with m.hidden() for sensitive fields, definePresenter() with .redactPII(['*.ssn', '*.diagnosis']) for DLP compliance, FSM state gating via .bindState() for workflow enforcement, and file-based routing under src/tools/. You review the PR.

The Skill is not documentation. It is the security contract. Every server the AI produces inherits the governance stack because the Skill encodes Presenters, state machines, and lockfile generation as mandatory structural patterns.

📄 SKILL.md · llms.txt (complete API reference for LLM consumption)

Security Architecture

Egress Firewall — Presenter

The Presenter validates every response through a Zod schema compiled from defineModel(). Undeclared fields are stripped in RAM before serialization. PII is redacted via V8-optimized fast-redact compiled functions. Rules travel with data, not in the system prompt. The Late Guillotine pattern applies redaction after UI blocks render — charts and suggestions always see full data, the wire never does.

const PatientPresenter = createPresenter('Patient')
    .schema(PatientModel)
    .redactPII(['*.ssn', '*.diagnosis'])
    .rules((p) => [
        p.status === 'critical' ? 'PRIORITY: Patient is critical.' : null,
    ])
    .suggest((p) => p.status === 'admitted'
        ? [suggest('ward.discharge', 'Begin discharge protocol')]
        : []);

The Presenter also runs a PromptFirewall — an LLM-as-Judge that evaluates dynamically generated system rules for prompt injection before they reach the agent. Fail-closed by default.

State Gate — FSM

Tools bound to FSM states are physically removed from tools/list when the current state does not match. The LLM cannot call what does not exist in its namespace. Powered by XState v5 with manual fallback when XState is not installed.

const gate = f.fsm({
    id: 'discharge', initial: 'admitted',
    states: {
        admitted:    { on: { PHYSICIAN_SIGNOFF: 'approved' } },
        approved:    { on: { DISCHARGE: 'discharged' } },
        discharged:  { type: 'final' },
    },
});

export default f.mutation('ward.discharge')
    .bindState('approved', 'DISCHARGE')
    .handle(async (input, ctx) => ctx.db.patients.discharge(input.id));

Serverless-compatible: FsmStateStore persists state to Redis/KV across request boundaries. Each request gets an isolated gate.clone().

Governance Stack

Eight introspection modules that make behavioral changes visible and auditable:

Sandbox

SandboxEngine executes LLM-provided JavaScript in a sealed V8 isolate. No process, require, fs, or network access. One isolate per engine, fresh empty context per execution. Memory-limited, timeout-enforced, output-capped, abort-signal-compatible.

Three Pathways

1. YAML — Zero Code

version: "1.0"
server:
  name: "github-tools"

connections:
  github:
    type: rest
    base_url: "https://api.github.com"
    auth:
      type: bearer
      token: "${SECRETS.GITHUB_TOKEN}"

tools:
  - name: search_repos
    description: "Search GitHub repositories"
    instruction: "Use for finding projects by topic or keyword."
    rules:
      - "Max 10 results per query"
    parameters:
      query: { type: string, required: true }
    execute:
      connection: github
      method: GET
      path: "/search/repositories"
      query: { q: "{{query}}", per_page: "10" }
    response:
      extract: ["items[].{full_name, description, stargazers_count, html_url}"]

mcpfusion yaml dev

2. Typed MVA — Full Control

export const InvoiceModel = defineModel('Invoice', m => {
    m.casts({
        id:           m.string(),
        amount_cents: m.number('CRITICAL: in CEN