Archrad Deterministic Mcp
@archradhq
About Archrad Deterministic Mcp
stdio MCP server from @archrad/deterministic—same engine as the archrad CLI. Validates architecture IR (structural + IR-LINT), merges local PolicyPack YAML, checks export drift vs generated files on disk, and returns static archrad_suggest_fix guidance per built-in rule code. No
Config
Add this server to your MCP-compatible client using the configuration below.
{
"mcpServers": {
"archrad": {
"command": "npx",
"args": [
"-y",
"--package=@archrad/deterministic",
"archrad-mcp"
]
}
}
}Tools
6Architecture-as-code validation: run this when you need to check whether an IR graph is valid or to list violations before export or drift checks. Keywords: validate IR, architecture lint, IR-STRUCT, IR-LINT, policy pack, blueprint graph, nodes and edges. Runs in one call: 1) Structural validation — graph shape, references, IR-STRUCT-* errors. 2) Architecture lint — design rules (auth, dead nodes, DB access, sync chains, etc.). 3) Optional PolicyPack rules — pass policiesDirectory to load YAML/JSON packs from disk. Returns irStructuralFindings, irLintFindings, and combined (sorted by severity). ok is false when any finding has severity "error". After results: call archrad_suggest_fix with a finding code for remediation text; use archrad_lint_summary for a short human-readable digest. Input: provide exactly one of ir (inline JSON object) or irPath (path to .json). Large graphs: prefer irPath.
Human-readable summary of validation results: error/warning counts and up to 20 top findings (plain text). Keywords: summary, PR comment, explain violations, readable lint output. Use when you need a short narrative or comment, not structured JSON. For machine-actionable findings, use archrad_validate_ir instead. Same inputs as archrad_validate_ir: ir or irPath, optional policiesDirectory. Provide only one of ir or irPath.
Look up curated remediation steps and documentation URL for one built-in rule code (e.g. IR-LINT-MISSING-AUTH-010, IR-STRUCT-*, DRIFT-*). Keywords: remediation, how to fix, rule code, docs link, IR-LINT, IR-STRUCT. Does not return generated code patches or IR edits — only static guidance. PolicyPack and org-specific rule ids are not covered; see your YAML packs. Call archrad_list_rule_codes to list codes that have static guidance.
Returns the sorted list of built-in IR-STRUCT-*, IR-LINT-*, and DRIFT-* codes that archrad_suggest_fix can explain. Keywords: catalog, all rules, rule list, documentation index. Use before suggest_fix to confirm a code exists. Excludes PolicyPack custom ids. No arguments.
Compare the architecture IR to generated code under exportDir and report drift (files that no longer match deterministic export). Keywords: drift, CI, codegen diff, FastAPI, Express, Node, Python, validate export, architecture vs implementation. Requires: ir or irPath, exportDir (absolute path to the export tree), and target. target must be "python" or "nodejs" (use "nodejs" for Node/TypeScript; do not use "node"). Optional: policiesDirectory, skipIrLint (true to skip IR-LINT and only check drift). Returns driftFindings plus IR structural and lint findings from the same engine as CLI validate-drift.
Validate PolicyPack YAML/JSON without running against a graph: syntax, rule ids, and compilation. Keywords: policy pack, YAML rules, validate policies, org rules, offline check. You usually do not need this before archrad_validate_ir or archrad_validate_drift — those accept policiesDirectory and load packs internally. Use this tool to debug pack files in isolation. Provide either directory (folder path) or files (array of { name, content }), not both.
Overview
What is Archrad Deterministic Mcp?
Archrad Deterministic Mcp gives AI coding agents (Claude Desktop, Cursor, and other MCP-capable hosts) deterministic architecture governance. It validates an intermediate representation (IR) against structural rules and built-in IR-LINT with no LLM in the validation loop, using the same engine as the archrad CLI. Licensed under Apache-2.0.
How to use Archrad Deterministic Mcp?
Install globally with npm install -g @archrad/deterministic; the binary archrad-mcp ships with the package. Alternatively, use npx without a global install by adding the server configuration to your MCP host (e.g., Claude Desktop's claude_desktop_config.json). The server exposes six MCP tools: archrad_validate_ir, archrad_lint_summary, archrad_validate_drift, archrad_policy_packs_load, archrad_suggest_fix, and archrad_list_rule_codes.
Key features of Archrad Deterministic Mcp
- Deterministic IR validation without LLM involvement
- Built-in IR-LINT rules for common architecture issues
- Drift detection between IR and on-disk export directories
- Optional PolicyPack support for custom governance rules
- Stateless per-call design; policies must be passed on each invocation
- Curated, static fix suggestions for individual rule codes
Use cases of Archrad Deterministic Mcp
- AI agents validating system graphs before code generation
- Pre-commit structural checks in CI/CD pipelines
- Comparing architecture IR against existing code exports for drift
- Smoke-checking PolicyPack YAML or JSON configurations
- Listing all built-in rule codes with static remediation guidance
FAQ from Archrad Deterministic Mcp
Does Archrad Deterministic Mcp use an LLM for validation?
No. The validation is fully deterministic, using the same engine as the archrad CLI. No AI is involved in rule evaluation.
What transport protocol does the server use?
The server uses standard MCP transport. For detailed transport options, IR size limits, and testing recipes, refer to the separate MCP.md documentation.
How are policies loaded and applied?
Policies are passed per-tool call via the policiesDirectory parameter. The archrad_policy_packs_load tool compiles and smoke-checks them but does not register them; they must be re-provided for each validation or lint call.
Can I use this with enterprise policy management?
Yes. For teams using ArchRad Cloud, a separate package (@archrad/remote-mcp) provides an HTTP MCP gateway that fetches org PolicyPack context using a bearer token. Deployment and token issuance are product-specific.
What are the runtime dependencies?
The server requires Node.js and the npm package @archrad/deterministic. It is tested with standard MCP hosts like Claude Desktop and Cursor. No additional databases or services are needed for the open-source version.
Frequently asked questions
Does Archrad Deterministic Mcp use an LLM for validation?
No. The validation is fully deterministic, using the same engine as the `archrad` CLI. No AI is involved in rule evaluation.
What transport protocol does the server use?
The server uses standard MCP transport. For detailed transport options, IR size limits, and testing recipes, refer to the separate `MCP.md` documentation.
How are policies loaded and applied?
Policies are passed per-tool call via the `policiesDirectory` parameter. The `archrad_policy_packs_load` tool compiles and smoke-checks them but does not register them; they must be re-provided for each validation or lint call.
Can I use this with enterprise policy management?
Yes. For teams using ArchRad Cloud, a separate package (`@archrad/remote-mcp`) provides an HTTP MCP gateway that fetches org PolicyPack context using a bearer token. Deployment and token issuance are product-specific.
What are the runtime dependencies?
The server requires Node.js and the npm package `@archrad/deterministic`. It is tested with standard MCP hosts like Claude Desktop and Cursor. No additional databases or services are needed for the open-source version.
Basic information
More Developer Tools MCP servers
AptiBuild AI — Career Intelligence
ParklandBuildsAptiBuild AI — Career Intelligence MCP Server Structured career decision scaffolding for AI agents. Provides personality-conditioned business idea generation, official BLS OEWS wage data, FRED labor market signals, and

Apiosk
obcraftAI-native payments for tools and APIs. Discover, pay for, execute, and publish monetized APIs directly from your agent, settled per call in USDC over x402, through the Model Context Protocol. payments · finance · x402 ·
CodeSentinel
icohangar-opsCodebase health as MCP tools: dead code, circular dependencies, coupling, architectural drift.

Reelier
Maxime HouleAgents make claims. Reelier writes receipts — record an agent's tool-call workflow once, replay it deterministically at 0 tokens, and diff runs to catch drift.

endoflife.ai
endoflife.aiFree MCP server for software end-of-life intelligence: EOL dates, support status, and the 0-100 EOL Risk Score across 480+ tracked products. Agents can check a single version, pull a product's full lifecycle schedule, or
Comments