Skip to main content
Glama

kernelee-mcp-tools

Architecture-index MCP tools for kernelee apps — a ts-morph static scan and the kernel's own runtime wiring projection, assembled into one JSON index and served to coding agents over the Model Context Protocol.

A kernelee app's control flow is data (symbols, pipes, buffer states), so it can be indexed and queried instead of grepped: "who drives this endpoint?", "who writes this state?", "what is one hop downstream of this symbol?".

What it does

Three cooperating pieces:

  • Foundation (kernelee-mcp-tools) — builds an IndexDocument from: the runtime wiring projection (projectWiringGraph over the app's live pipe catalog), a ts-morph static scan of the app's tsconfig program (input types, readsState / writesState, drive sites — dispatch/call/run and React useDispatch — per-stage flows, fork branch selectors, declaration/implementation addresses), and a git stamp. Written atomically to index.json.

  • MCP server (kernelee-mcp-tools/server) — runKernelIntrospectMCPServer(configuration): a stdio MCP server that re-reads index.json per query and prepends a staleness verdict (fresh / stale / unknown vs. the current working tree) to every answer.

  • CLI (kernelee-introspect) — kernelee-introspect <config-module> regenerates the index from a consumer-supplied config and prints a one-line summary. Wire it into the app's build so the index can never silently rot.

Contract: introspection output artifacts — outputPath and KERNEL_INTROSPECT_TRACE_PATH, if set — must live outside the repository, or be gitignored. Staleness works by comparing a content hash of the whole non-ignored working tree; if the index (or a trace dump) itself falls inside that domain, writing it changes the very tree the hash just described, and staleness can never converge on fresh — every arch_refresh reports stale again, forever. The CLI enforces this with a hard error before writing anything (see docs/mcp-server-gotchas.md #10); the server appends a diagnostic note to a stale hint (including arch_refresh's own success response) whenever it can tell this is the cause. A path inside a git repo nested under the repo root is a known, conservative false positive — gitignoring it there is harmless.

Related MCP server: Graft

The tools

Tool

Args

Answers

arch_overview

—

The map: ring × device symbol counts, every endpoint headline, state / shared-stage / part lists. Start here.

arch_endpoint

key

One endpoint in full: the stage tree (kinds, flows, notes, fork branches), its binding, and who drives / diverts-to / reads / writes around it.

arch_state

state

One buffer state: declaration, named mutations, writers, and readers split into circuit vs. presentation.

arch_walk

from, edges?, depth?

A typed-edge BFS neighborhood (drivenBy, divertsTo, callsSymbol, reads, writes, usesSharedStage) around any node.

arch_refresh

—

Runs the consumer's refresh command, then reloads the index.

Misses come back with nearest-key suggestions rather than bare errors.

Consumer integration

The package hardcodes nothing about your app — everything arrives by injection. A consumer supplies two small pieces:

  1. An IntrospectConfig module for the CLI — repo root, tsconfig path, catalog file/function, state files, output path, and optional CI gates (failOnWiringIssues, failOnOffBufferControlValues, each with an allowlist):

    // package.json
    "scripts": {
      "introspect": "tsx node_modules/.bin/kernelee-introspect scripts/introspect.config.ts",
      "build": "npm run introspect && vite build"
    }
  2. A thin MCP launcher that bakes its defaults and starts the server — registered in the MCP client (e.g. Claude Code's .mcp.json) as a stdio server:

    // scripts/introspect-mcp-server.mjs
    import { runKernelIntrospectMCPServer, configurationFromEnvironmentOverridingDefaults }
      from '@s-age/kernelee-mcp-tools/server';
    
    await runKernelIntrospectMCPServer(configurationFromEnvironmentOverridingDefaults({
      indexPath: '.claude/introspect/index.json',
      repoRoot: process.cwd(),
      refreshCommand: 'npm run introspect',
      refreshTimeoutSeconds: 300,
    }));

    Every default can be overridden via KERNEL_INTROSPECT_INDEX_PATH, KERNEL_INTROSPECT_REPO_ROOT, KERNEL_INTROSPECT_REFRESH_COMMAND, KERNEL_INTROSPECT_REFRESH_TIMEOUT_SECONDS.

The reference consumer is kernelee-lifegame — see its scripts/introspect.config.ts, scripts/introspect-mcp-server.mjs, and docs/kernel-introspect-mcp.md (an end-user guide with example outputs and client-registration recipes).

Development

npm run build      # tsc → dist/ (the bin needs this before first use)
npm test           # vitest run — unit + MCP end-to-end integration tests
npm run typecheck  # tsc --noEmit

Dependencies: @modelcontextprotocol/sdk, ts-morph, zod; kernelee is a peer dependency. fixtures/authoringShapes/ is a minimal kernelee app the scan tests run against; docs/ holds implementation gotcha notes.

License

MIT © s-age

Related MCP Connectors

  • Patterns for designing and reviewing AI skills, agents, and multi-agent workflows. Find guidance on context economy, delegation, verification, and tool design; inspect claims, worked examples, maturity labels, and source references. Five read-only tools let agents discover relevant patterns, compare concise cards, read specific sections, and explore relationships. Hosted Streamable HTTP at https://agentic-atlas.dev/mcp/ — no installation, account, or API key required. Browse the atlas at https://agentic-atlas.dev/.

  • Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.

  • The Cortex MCP server provides read-only access to real-time engineering context from the Cortex developer portal, allowing AI coding assistants to answer natural language questions about your organization's catalog (microservices, libraries, domains, teams, infrastructure), scorecards (engineering standards and best practices), initiatives (goals and deadlines), and Engineering Intelligence metrics. It includes tools for querying documentation, tracking personal entities, and accessing AI-assisted insights across the entire Cortex ecosystem.

  • Codebase graphs, caller impact analysis, and recorded project context for AI coding agents.

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Semantic code indexer with GraphRAG knowledge graph. Index your codebase, search in natural language, and expose everything via MCP so AI agents understand architecture — not just files.
    477
    Apache 2.0
  • A
    license
    B
    quality
    D
    maintenance
    Local-first codebase context engine that parses code into a ranked dependency graph and serves it to AI tools via MCP for deep structural understanding.
    5
    19 npm
    1
    MIT
  • A
    license
    Not graded
    quality
    D
    maintenance
    Provides AI-powered architecture analysis and visualization of codebases, exposing 17 MCP tools for querying components, dependencies, and generating interactive diagrams.
    47 npm
    5
    MIT
  • A
    license
    Not graded
    quality
    A
    maintenance
    Provides AI agents with a live architecture model of a codebase, enabling queries for root cause analysis, blast radius, and dependency traversal through MCP tools.
    181 npm
    20
    Apache 2.0

Appeared in Searches