@happyvertical/smrt-dev-mcp

Development MCP server for the s-m-r-t framework providing code generation, project introspection, deterministic ecosystem knowledge, and portable review or architecture prompt bundles.

Installation

pnpm install @happyvertical/smrt-dev-mcp

Agent Plugins package

The published package root is a self-contained Agent Plugins 1.0.0 plugin. Compatible clients discover plugin.json, mcp.json, and skills/smrt-code-review/SKILL.md directly from the installed package root. The shipped mcp.json declares only the local stdio server:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
  "mcpServers": {
    "smrt-dev-mcp": { "type": "stdio", "command": "./dist/index.js" }
  }
}

The executable path is plugin-relative and must remain inside the resolved plugin root. Clients provide PLUGIN_ROOT (the resolved package root) and the client-managed, persistent PLUGIN_DATA directory; this package does not set or override either reserved environment variable. Portable configuration never contains credentials, secrets, authorization headers, or OAuth settings. Clients own authorization interaction and credential storage.

Agent Plugins 1.0.0 is a Working Draft. This package targets only the canonical 1.0.0 schema identifiers and includes pinned schema snapshots for offline validation; clients must not fetch schemas when loading the package.

streamable-http is not shipped: its endpoint is gated on #2147. A future, credential-free configuration could declare a streamable-http URL only after that endpoint exists; it must not be inferred from this package today.

Usage

Add to your project-local .mcp.json, Codex MCP config, or Claude Desktop config:

{
  "mcpServers": {
    "smrt-dev-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@happyvertical/smrt-dev-mcp"]
    }
  }
}

For global MCP client config, prefer a launcher that does not depend on the current working directory. Avoid command = "pnpm" with args = ["exec", "smrt-dev-mcp"] in user-level config: MCP clients can start servers from repositories that do not install this package, and pnpm can run dependency-status or build-approval checks before the MCP server starts.

Install the package in a stable location and point Codex at the built server:

[mcp_servers.smrt-dev-mcp]
command = "node"
args = ["/absolute/path/to/node_modules/@happyvertical/smrt-dev-mcp/dist/index.js"]

If your Node runtime is managed by a toolchain, use a tiny absolute wrapper instead:

#!/usr/bin/env sh
exec /absolute/path/to/node /absolute/path/to/node_modules/@happyvertical/smrt-dev-mcp/dist/index.js "$@"

[mcp_servers.smrt-dev-mcp]
command = "/absolute/path/to/smrt-dev-mcp-wrapper"
args = []

Set DEBUG=true in the environment to enable diagnostic logging.

Workspace Discovery And Coverage

Package discovery reads the workspace globs, so apps/* products are indexed the same as packages/* ones. Resolution order:

1. pnpm-workspace.yaml packages: — literals, dir/*, **, and ! negations
2. package.json#workspaces (array or { packages: [...] })
3. packages/* as a last-resort fallback

Workspace globs must remain relative to the declared root: absolute paths and .. segments are rejected, and matched directories are realpath-confined before their manifests are read. Positive globs share a 10,000 directory-entry traversal budget. Exceeding it stops discovery with an error diagnostic and no partial package set; deeply nested valid ** matches remain supported because the limit is based on work performed rather than directory depth. A separate 512-package cap prevents broad matches from fanning out package reads, package paths are revalidated immediately before those reads, and scanner fallbacks run with at most eight concurrent scanners.

The workspace root is also indexed when it has a package.json, which is how single-package repositories work; it is scanned with the member package directories excluded so it can own objects without absorbing theirs. Per package, objects resolve from a domain artifact, then a package-local manifest, then a source scan (@happyvertical/smrt-scanner) — the same fallback introspect-project uses, so both paths agree on one root. Every package records objectSource (domain-artifact | manifest | scanner | none) plus a reason.

Manifest objects that belong to another package are rejected rather than counted: a runtime .smrt/manifest.json is often an aggregate registering a package's dependencies too, and counting those inflates the relationship facts. A consuming package's artifact can also restate its dependency's objects under its own name, so Relationships-v2 collapses a shared className::tableName across a dependency edge — by connected component, keeping the copy the others depend on — and reports duplicate-object-identity. Unrelated packages that share a class name stay distinct.

The index therefore carries two extra blocks (schemaVersion: 2):

• coverage — workspaceGlobs, workspaceGlobSource, packageDirs, packagesWithObjects, and packagesWithoutObjects with a reason, the artifact paths checked, and a remedy for each
• diagnostics — discovery problems. Discovering zero objects is an error-grade diagnostic naming the roots and artifact paths checked plus the commands to fix it, so an unseen project is never reported as a project with no model. smrt-architecture, smrt-review, and the reflect-* tools surface it.

Response Budgets

Knowledge and introspection tools return a summary by default and accept detail: "full" for the complete payload.

• introspect-project summary returns one compact record per object (className, qualifiedName, extends, tableName, tenantScope, fieldCount, compact relationship strings, mcpOperations). A response that exceeds its character budget reports a truncated block with the omitted count and filter guidance instead of being silently cut. maxChars overrides the budget.
• smrt-architecture, smrt-review, and the build-*-context tools list authored AGENTS.md and module docs by path rather than embedding them, and return compact package records.

The smrt dev:knowledge-* CLI commands request detail: "full", so their output is unchanged.

Knowledge Boundary

smrt-dev-mcp is model-agnostic. Its review and architecture tools do not call Codex, Claude, or any other model provider directly. They return deterministic findings plus a reusable prompt bundle that can be sent to the local model plan or provider of your choice.

Downstream s-m-r-t packages/apps can publish their own scoped smrt-knowledge.json artifact. Discovery prefers local .smrt/smrt-knowledge.json, then dist/smrt-knowledge.json, then source manifest artifacts before falling back to raw manifest/doc scanning. The runtime manifest.json stays focused on object registration; smrt-knowledge.json is the agent/developer contract.

Cache Metadata

The deploy-static tools/list and prompts/list catalogs advertise a one-day private cache lifetime. Workspace knowledge resources (resources/list and resources/read) are also private, but use ttlMs: 0: they are rebuilt from the current workspace on each request and have no transport-visible invalidation signal that could make a positive freshness promise honest.

After using a model to update package docs or expertise, always run the deterministic checker again:

pnpm knowledge:check --strict --format markdown

Use --format json when another script needs machine-readable output.

Downstream Review Flow

1. Run the downstream app build or dev server so .smrt/smrt-knowledge.json exists.
2. Call reflect-domain-knowledge to confirm package and SDK coverage.
3. Call build-domain-review-context or smrt-review with changed files, scope, and optional package.
4. Send the returned prompt bundle to Codex, Claude, or another model.
5. Re-run check-domain-knowledge after edits.

Equivalent CLI commands are:

smrt knowledge:review-context --scope package --package content --format markdown
smrt knowledge:architecture-context "tenant-aware publishing workflow" --format json

Agent Skills

The package ships harness-agnostic agent skills under skills/, the fixed Agent Plugins discovery location.

Downstream agents should fetch the review procedure before starting a formal s-m-r-t review:

{
  "name": "get-agent-skill",
  "arguments": {
    "name": "smrt-code-review"
  }
}

The returned skillMarkdown is plain Markdown with YAML frontmatter (name and description) and a harness-neutral body. Skill-aware harnesses can parse the frontmatter; other MCP-capable harnesses can ignore it, call smrt-review for deterministic context, inspect the actual diff, and produce a findings-first review. Native MCP prompt/resource clients can also load the smrt-code-review prompt or smrt-dev-mcp://agent-skills/smrt-code-review resource.

Available Tools

The server exposes these MCP tools:

generate-smrt-class

Generate a complete s-m-r-t class with @smrt() decorator, fields, and imports.

Supported property types: text, integer, decimal, boolean, datetime, json.

introspect-project

Scan a project directory for s-m-r-t objects and return a manifest-equivalent class/field/relationship report. Discovery prefers .smrt/manifest.json, then dist/manifest.json, then src/manifest/manifest.json; when no artifact is available it falls back to @happyvertical/smrt-scanner.

review-smrt-project

Run an advisory downstream ecosystem review. The tool scans package manifests and source imports for missing HappyVertical dependencies, direct storage bypasses, custom HTTP shells, custom object manifest generation, local auth/tenancy/audit seams, and UI shell drift. It returns deterministic findings and suggested follow-up issue titles; it does not modify files.

reflect-knowledge

Return package coverage, SDK package coverage, relationship-v2 counts, and freshness status from the deterministic s-m-r-t + HappyVertical SDK knowledge index.

reflect-domain-knowledge

Return domain artifact coverage, missing exported artifacts, SDK package coverage, relationship-v2 counts, and freshness status.

check-knowledge-freshness

Run the same deterministic freshness checks exposed by pnpm knowledge:check.

check-domain-knowledge

Alias over the deterministic checker that emphasizes downstream smrt-knowledge.json artifact freshness.

build-review-context

Select relevant s-m-r-t and HappyVertical SDK package expertise for changed files, then return a model-ready prompt bundle.

build-domain-review-context

Domain-scoped alias for build-review-context. Accepts the same fields plus scope and package.

smrt-review

Return deterministic review findings, a prompt bundle, or both.

build-architecture-context

Select relevant s-m-r-t and SDK package expertise for an idea or documentation, then return a model-ready architecture prompt bundle.

build-domain-architecture-context

Domain-scoped alias for build-architecture-context. Accepts the same fields plus scope and package.

smrt-architecture

Return package recommendations, SDK recommendations, an object-model sketch, risks, questions, and the reusable architecture prompt bundle.

list-agent-skills

List bundled harness-agnostic agent skills. No parameters.

get-agent-skill

Return a bundled agent skill as Markdown, with optional referenced files.

MCP Resources And Prompts

Resources:

• smrt://knowledge/project — composed project knowledge index as JSON.
• smrt://knowledge/package/{name} — package-scoped knowledge as JSON.
• smrt-dev-mcp://agent-skills/smrt-code-review — portable review skill.

Prompts:

• domain-code-review — returns the review prompt bundle.
• domain-architecture — returns the architecture prompt bundle.
• smrt-code-review — returns the harness-agnostic review procedure.

MCP Tier Context

• Tier 1 (Runtime): auto-generated from @smrt() objects -- live data operations
• Tier 2 (Development): this package -- code generation and project analysis
• Tier 3 (Docs): framework documentation access; smrt-docs-mcp is no longer launched from this monorepo unless an external package/repo is installed and configured explicitly

Dependencies

• @modelcontextprotocol/server -- MCP server protocol
• @happyvertical/smrt-core -- manifest and object registry
• @happyvertical/smrt-types -- shared domain knowledge contract

License

MIT