Constraint Registry

Policy-as-code guardrails for AI-generated code. An MCP server that serves engineering constraints to coding agents (Claude Code, Cursor, Codex) at generation time and validates artifacts with OPA, Conftest, Checkov, and Semgrep.

A single, queryable source of engineering constraints (infrastructure, organizational, architectural) that coding agents (Claude Code, Cursor, Codex, …) consult at code-generation time, exposed over an MCP server. It does not enforce constraints itself — it provides guidance to agents and delegates deterministic validation to existing enforcement engines (OPA, Conftest, Checkov, Semgrep).

Constraints are authored in source repos, aggregated into an immutable, versioned bundle, and served over MCP so an agent can:

1. describe_scope — discover the valid selector vocabulary,
2. get_constraints — fetch the rules relevant to what it's building, and
3. validate — check a candidate artifact against the bound enforcement engines.

Authoritative requirements: constraint-registry-v0-spec.md. Requirement → component → test mapping: TRACEABILITY.md.

Contents

• Features
• Prerequisites
• Quick start
• Running the MCP server
• Integrating with coding agents
• Authoring constraints
• Hot reload
• Validation harness
• Adding an enforcement engine
• Repository layout
• Contributing
• License

Features

• Three constraint categories — infrastructure, organizational, architectural, including relationship-style selectors (e.g. "no synchronous calls across domain boundaries") and advisory (no-enforcement) constraints.
• Multi-source aggregation — import from many source repos; ids are namespaced per source; deterministic, content-hashed, immutable versioned bundles.
• Precedence & anti-drift — a configurable default policy (hard outranks weaker; a downstream source may not relax a higher-precedence rule on the same scope); fixture cross-checks keep guidance and enforcement from drifting.
• Pluggable engines — a stable adapter interface with four real adapters: OPA and Conftest (Rego policies), Checkov (IaC scanning), and Semgrep (application source code). SARIF-emitting engines share one normalization seam (adapters/sarif/), so adding a new SARIF engine is mostly wiring. Adding an engine = one adapter + one config line (see Adding an enforcement engine).
• Catalog importers — Checkov and Semgrep ship importers that turn an engine's rule catalog/ruleset into draft constraint stubs (with license/source provenance) for a human to enrich — a fast path to bootstrapping a source.
• MCP server — three tools (describe_scope, get_constraints, validate) over stdio or a shared HTTP endpoint. get_constraints fails open so an agent is never blocked.
• Hot reload — the server can periodically re-import so constraint changes are picked up without a restart.
• Validation harness — proves the registry and constraint set are internally consistent; machine-readable JSON, non-zero exit on failure.

Prerequisites

Install the external engines on macOS:

brew install opa conftest checkov   # semgrep is installed by `uv sync`

Each engine is optional and independent: any test or harness check whose engine binary is not on PATH is skipped, not failed. The registry and the get_constraints/describe_scope guidance work without any engine at all — an engine is only needed to run validate and the fixture cross-checks for constraints bound to it.

Quick start

git clone https://github.com/SureshKhemka/constraints-registry.git
cd constraints-registry

uv sync                      # create the venv + install deps (incl. semgrep)

uv run cregistry-harness     # run the validation harness against the bundled samples

The harness emits machine-readable JSON and exits non-zero on any failure. A green run looks like:

{ "passed": true, "summary": { "pass": 21, "fail": 0, "skip": 0, "total": 21 }, "checks": [ ... ] }

(skip is used only when an optional engine like conftest is not installed.)

Running the MCP server

Two transports — pick based on how you want tools to connect.

# stdio (default): each tool launches its own copy; nothing to manage
uv run cregistry-mcp

# one shared HTTP server every tool connects to (recommended for multiple tools)
uv run cregistry-mcp --http --port 8765 --reload-interval 60

Flags: --transport {stdio,http,sse}, --http (shorthand), --host (default 127.0.0.1), --port (default 8765), --config (or $CREGISTRY_CONFIG), --reload-interval SECONDS (0 = off).

Manage the shared HTTP server:

lsof -ti tcp:8765 | xargs kill     # stop
# restart = stop + start

Full operational guide (stop/restart, macOS launchd auto-start, the repo-sync/decoupling pattern): docs/RUNNING.md. Tool input/output contracts: docs/MCP_CONTRACT.md.

Integrating with coding agents

The server exposes three tools: describe_scope, get_constraints, validate.

Claude Code

# shared HTTP server (start it first, see above), available in every project:
claude mcp add --scope user --transport http constraint-registry http://127.0.0.1:8765/mcp

# OR stdio (no separate server to run; Claude launches it):
claude mcp add constraint-registry -- uv run --directory "$(pwd)" cregistry-mcp

claude mcp list   # should show: constraint-registry ... ✔ Connected

Cursor (~/.cursor/mcp.json)

{ "mcpServers": { "constraint-registry": { "url": "http://127.0.0.1:8765/mcp" } } }

Codex / other stdio-only tools

Configure an MCP server with command: uv, args: ["run","--directory","/abs/path/to/repo","cregistry-mcp"].

Make the agent actually consult it

Agents auto-discover the tools, but to get them to consult the registry before generating code, add an instruction to your project (or ~/.claude/CLAUDE.md):

Before writing AWS/infra code, call the constraint-registry MCP: describe_scope to learn valid selector values, then get_constraints with the right scope, and comply with every hard constraint as a non-negotiable downstream gate. Optionally validate the result.

Authoring constraints

A source is a directory with constraints/*.yaml (one constraint per file) and, optionally, policies/ (engine policies) and fixtures/ (sample artifacts). Register sources and engines in registry.config.yaml:

sources:
  - { name: platform-security, path: sources/platform-security, precedence: 100 }
  - { name: data-platform,     path: sources/data-platform,     precedence: 50  }
engines:
  - { name: opa,      adapter: "cregistry.engine.adapters.opa:OpaAdapter" }
  - { name: conftest, adapter: "cregistry.engine.adapters.conftest:ConftestAdapter" }
  - { name: checkov,  adapter: "cregistry.engine.adapters.checkov:CheckovAdapter", options: { min_level: warning } }
  - { name: semgrep,  adapter: "cregistry.engine.adapters.semgrep:SemgrepAdapter", options: { min_level: warning } }
precedence_policy: default

A constraint (see sources/platform-security/constraints/aws-s3-no-public-access.yaml):

id: aws.s3.no-public-access
title: "S3 buckets must not be publicly accessible"
intent: "Public buckets are the top source of data-exposure incidents."
category: infrastructure          # infrastructure | organizational | architectural
scope:
  providers: [aws]
  resource_types: [aws_s3_bucket] # Terraform resource ids (NOT "s3_bucket")
  environments: [all]
  repos: ["tag:data-plane"]
severity: hard                    # hard | soft | advisory
enforcement:                      # omit for an advisory (guidance-only) constraint
  - { engine: opa, policy: policies/s3_public.rego }
guidance:
  do:   ["Attach an aws_s3_bucket_public_access_block with all four flags true"]
  dont: ["Never set acl = 'public-read' or 'public-read-write'"]
  example_compliant: |
    {"resources": {"aws_s3_bucket": {"data": {"acl": "private", "public_access_block": true}}}}
owner: platform-security
version: 1.0.0
fixtures:                         # optional; cross-checked against the engine
  pass: fixtures/s3_private.json
  fail: fixtures/s3_public.json

Scoping notes (matters when agents query):

• resource_types use the target tooling's identifiers (Terraform: aws_s3_bucket). Call describe_scope to discover the exact vocabulary present.
• A query that omits a dimension matches broadly; a value that contradicts a constraint's selector excludes it. Relationship-scoped constraints are only returned for queries that supply a matching relationship.
• After authoring, run uv run cregistry-harness to validate schema, precedence, and fixtures.

Hot reload (no restart on constraint changes)

Run the server with --reload-interval N and it re-imports from disk every N seconds, publishing a new immutable bundle when content changes:

uv run cregistry-mcp --http --port 8765 --reload-interval 60

• No-op when nothing changed; previous bundle versions stay pinnable by id.
• A failed re-import (e.g. an unresolvable precedence conflict) keeps the last-good bundle serving and logs the reason — the server never goes dark.
• The server reads from the configured source paths, so wire your teams' constraint repos to sync/pull into those paths (a separate ops job, e.g. a cron git pull or CI publish). Code/dependency changes still need a restart.

Validation harness

uv run cregistry-harness runs end-to-end against the bundled, self-contained sample sources and proves: schema conformance, deterministic import, malformed- constraint isolation, namespacing & precedence, versioning & deprecation, engine- interface conformance (incl. a reusable suite any adapter can be run against), fixture cross-checks / broken-binding detection, the MCP contract / scoping / fail-open, and hot-reload behavior. It prints structured JSON and returns a non-zero exit on any failure — suitable for CI.

uv run cregistry-harness            # human-readable JSON to stdout, exit 0/1
uv run cregistry-harness --config path/to/registry.config.yaml

Adding an enforcement engine

Implement the EngineAdapter interface in a new module under src/cregistry/engine/adapters/, add one line under engines: in registry.config.yaml, and validate it against the existing conformance suite — no changes to the schema, importer, MCP server, or harness. Full walkthrough: docs/ADDING_AN_ENGINE.md.

Repository layout

src/cregistry/
  model.py            constraint schema (Pydantic)
  loader.py           load + per-field schema validation
  config.py           registry config (sources, engines)
  importer.py         import → aggregate → bundle
  precedence.py       namespacing precedence / conflict resolution
  scope.py            scope matching (query + conflict)
  bundle.py store.py  immutable versioned bundles + store
  query.py validate.py  scoped queries + artifact validation
  integrity.py        fixture cross-check / anti-drift
  service.py          transport-independent service (+ hot reload)
  mcp_server.py       MCP server (stdio / http) + CLI
  engine/             stable engine interface, registry, and adapters:
    adapters/opa, adapters/conftest, adapters/checkov, adapters/semgrep
    adapters/sarif    shared SARIF normalization seam (used by checkov + semgrep)
  harness/            the validation harness (checks/*)
sources/              bundled sample source repos (constraints, policies, fixtures)
scenarios/            self-contained fixtures for harness edge cases
tests/                pytest suite: adapter conformance, fixtures, import, e2e
docs/                 RUNNING.md, MCP_CONTRACT.md, ADDING_AN_ENGINE.md
deploy/               launchd template for auto-starting the HTTP server
CONTRACTS.md          the frozen engine-adapter seam new adapters code against

Run the adapter test suite directly with uv run pytest.

Contributing

Contributions are very welcome — bug reports, new engine adapters, constraint sources, and docs. Adding an engine is intentionally small: one adapter module plus one config line, with no changes to the schema, importer, MCP server, or harness.

• Read CONTRIBUTING.md for dev setup, conventions, and the engine-adapter checklist.
• Be a good neighbor — see the Code of Conduct.
• Found a vulnerability? Report it privately per SECURITY.md.

Before opening a PR, make sure uv run pytest and uv run cregistry-harness are green; CI runs both on every pull request.

License

Licensed under the Apache License 2.0. See NOTICE for attribution and the licenses of the external engines this project integrates with.