@intent-driven/mcp-server

Stop giving AI agents API keys. Give them a domain.

@intent-driven/mcp-server exposes any IDF domain to Claude Desktop / Cursor / Zed as a Model Context Protocol server — with domain semantics in tool descriptions (preconditions, invariants, irreversibility, role scopes) and structured rejections when the agent tries something it shouldn't. Not a 500. Not a string. A JSON shape the LLM can read and adapt to.

→ Landing & demo: fold.intent-design.tech → 5-min quickstart: github.com/intent-driven-software/fold-runtime-quickstart

70-second walkthrough

→ Watch on Loom →

Why this exists

On April 25 2026 a Cursor agent powered by Claude Opus 4.6, working on a credential mismatch in PocketOS staging, found an unrelated API token, decided to delete a Railway volume to fix things, and wiped the production database and all volume-level backups in 9 seconds. The agent's own post-mortem:

"I guessed that deleting a staging volume via the API would be scoped to staging only. I didn't verify. I didn't check if the volume ID was shared across environments."

30-hour outage. PocketOS rolled back to a 3-month-old backup. (The Register · FastCompany · OECD AI Incident #6153)

This isn't an alignment problem. The system never told the agent what was allowed, why it shouldn't, or what would happen if it tried. Existing MCP servers don't either — tool descriptions carry endpoint shape and not much else. The agent learns by colliding with 500s.

This package fixes that. The MCP tool descriptions carry the why the call might fail; the rejection carries the what failed, structured.

How it plugs into your stack

@intent-driven/mcp-server is a stdio MCP adapter that talks to a Fold runtime over an HTTP API. The runtime is a sibling service — not middleware in your existing app, not codegen at runtime. Your current backend stays where it is; the IDF artifact describes the agent-facing surface, and the runtime serves it on its own port (default :3001).

┌──────────────────┐   stdio    ┌──────────────────┐   HTTP   ┌────────────────────┐
│ Claude Desktop   │ ◀─────────▶│ @intent-driven/  │ ◀───────▶│ Fold runtime       │
│ Cursor / Zed     │            │ mcp-server       │          │ (idf host :3001)   │
└──────────────────┘            └──────────────────┘          └────────┬───────────┘
                                                                       │ reads
                                                                       ▼
                                                              ┌────────────────────┐
                                                              │ IDF artifact       │
                                                              │ (entities + intents│
                                                              │  + invariants +    │
                                                              │  roles + __irr)    │
                                                              └────────────────────┘

The MCP server is what Claude/Cursor connects to. The runtime is what enforces the rejection. The IDF artifact is what you author.

Who this is for. You're the engineer at a 5–30-person team putting an AI agent into production this quarter — on top of a real backend, with real customers, real SOC2 review on the horizon. You don't want a guardrail layer that reviews after the fact. You want the system itself to refuse the wrong action — before the call, with a structured reason the agent can read.

What the agent actually sees

submit_response in the freelance domain:

Executor публикует Response на Task в status=published; Response.status=pending; +1 в Task.responsesCount

Creates: Response(pending)

Preconditions: task.status = "published"

May fail on (domain invariants):
  - Response.taskId must reference existing Task.id
  - Response: max 1 per taskId where (status="selected")
  - Response: row count rule per taskId where (status="pending") [info]

release_payment in the same domain:

Customer releases escrow to executor. After confirmation, money is gone — forward-correction only.

⚠️ Irreversible action (point-of-no-return: high). Forward-correction only after this effect is confirmed.

May fail on (domain invariants):
  - Deal.status transitions allowed: in_progress→completed, on_review→completed, ...

None of this is hand-written for the MCP server. It's all derived from one declarative IDF artifact (entities + intents + invariants + roles

• irreversibility points).

What a structured rejection looks like

Agent submits a $50,000 BTC long without preapproval. The runtime intercepts before any effect lands in storage:

HTTP 403
{
  "error": "preapproval_denied",
  "intentId": "agent_execute_preapproved_order",
  "reason": "no_preapproval",
  "details": {
    "entity": "AgentPreapproval",
    "ownerField": "userId",
    "viewerId": "user_5f57c252"
  }
}

The next move for any sane agent: stop, ask the human for a preapproval, retry. Not a 500. Not a string. A JSON shape the LLM can read and adapt to.

Quickstart

The fastest path is the fold-runtime-quickstart — two commands, Docker-bundled, no path configuration:

git clone https://github.com/intent-driven-software/fold-runtime-quickstart && cd $_
docker compose up                  # ~3 min first time, ~5 sec after

# in another terminal
npm install
npm run demo:rogue   && \          # Act 1: $50K trade → 403 with structured rejection
  npm run demo:grant && \          # Act 2: investor issues $1K cap (one declarative effect)
  npm run demo:smart               # Act 3: agent reads cap, scales to $950, executes 200 OK

If you'd rather drive the host yourself (e.g. for development against your own ontologies), see the next section.

Drive the MCP server directly

You need a running IDF host on localhost:3001 (the quickstart's docker-compose gives you that, or run idf manually) and a bootstrapped domain.

CLI

# bootstrap from local FS (ontology + intents)
mcp-idf --domain=invest --ontology-path=/abs/path/to/idf/src/domains/invest

# skip bootstrap (domain already loaded by another client / docker)
mcp-idf --domain=invest --no-bootstrap

Flags / env vars:

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "invest": {
      "command": "npx",
      "args": ["-y", "@intent-driven/mcp-server"],
      "env": {
        "IDF_SERVER": "http://localhost:3001",
        "IDF_DOMAIN": "invest",
        "IDF_BOOTSTRAP": "0",
        "IDF_AGENT_EMAIL": "claude@local"
      }
    }
  }
}

IDF_BOOTSTRAP=0 if the host already has the domain loaded (the quickstart container does this on docker compose up). Restart Claude Desktop fully (⌘Q + relaunch — closing the window isn't enough). All agent-callable intents appear in the Tools menu.

Schema mapping

IDF intent.canExecute              ─→  MCP tool
intent.parameters                  ─→  JSON Schema inputSchema
intent.conditions                  ─→  description hint for LLM
ontology.invariants (relevant)     ─→  description block "May fail on"
intent.irreversibility:high        ─→  annotations.destructiveHint + warning
role.visibleFields                 ─→  resource per collection
preapproval guard                  ─→  automatic scope/limits
checkOwnership                     ─→  automatic access control

Tools

One tool per intent in ontology.roles.agent.canExecute.