​

Authoring Policies with dmint-cli

​

The Policy Authoring Workflow

Human Security Intent (access.md)
        ↓
dmint create-policy (Multi-Turn Dialogue with LLM)
        ↓
Candidate Policy JSON
        ↓
Deterministic Core Validation (Policy.from_mapping)
        ↓
Verified Authoritative policy.json

​

1. General Policy Authoring (create-policy)

dmint create-policy

1. Requirements file: Path to your access.md (or defaults to typing requirements).
2. Local capability discovery: Statically parses local Python files using Python’s ast.parse() to discover @dmint.protected functions with zero code execution.
3. LLM Provider: Choose OpenAI, Gemini, Groq, OpenRouter, or a local Ollama instance.
4. Interactive refinement: Review the generated rules in terminal, request adjustments, and confirm.

​

Non-Interactive Batch Mode

dmint create-policy \
  -f access.md \
  -o policy.json \
  --provider openai \
  --model gpt-4o-mini \
  --non-interactive \
  -y

​

2. Specialized MCP Policy Authoring (create-mcp-policy)

1. Spawns or connects to the live downstream MCP server.
2. Dispatches tools/list to retrieve exact tool names, schemas, and descriptions.
3. Automatically namespaces the capabilities under the chosen --integration-id.
4. Generates both policy.json and the corresponding mcp_protection.json.

dmint create-mcp-policy \
  --command npx \
  --args "-y @modelcontextprotocol/server-postgres postgresql://localhost/mydb" \
  --integration-id db \
  -f access.md \
  -o policy.json \
  --config-output mcp_protection.json \
  -y

​

3. Offline Verification (verify-policy)

dmint verify-policy policy.json

✓ Verified policy.json: 4 rule(s)
  [ALLOW] tool='db', action='select', resource='*'
  [APPROVAL_REQUIRED] tool='db', action='update', resource='users'
  [APPROVAL_REQUIRED] tool='db', action='insert', resource='users'
  [DENY] tool='db', action='drop', resource='*'

​

Supported LLM Providers

​

Invariants Enforced During Authoring

• Least Privilege Default: Unmentioned capabilities default to DENY.
• No Wildcard Expansion: The compiler rejects attempts to replace specific resource scopes with unvetted wildcards (*).
• Capability Pinning: The authoring wizard validates all generated rules against the real discovered tools; hallucinated tool names are rejected and retried automatically.