synaps

Codebase intelligence for AI coding agents. Analyzes TypeScript/JavaScript codebases and serves actionable intelligence via MCP.

What It Does

Gives AI coding agents deep understanding of your codebase — not just what files exist, but how they connect, what conventions to follow, and what breaks when you change something. Combines AST analysis, git history mining, and import graph intelligence to answer questions no other tool can:

• "What files should I also check when modifying src/types.ts?" → Import dependents, co-change partners from git history, implicit coupling (files that change together without import relationships), and affected execution flows.
• "Which specific files import Convention from src/types.ts?" → 25 files (not all 98 importers of the file — only the ones using that symbol).
• "What caused this test failure?" → Ranked suspect files with confidence assessment, validated on 95 real bug-fix commits across 10 repos.
• "What patterns should I follow?" → 13 convention detectors covering error handling, async patterns, state management, API patterns, hooks, testing, and more.

demo.mp4

Quick Start

# MCP server for Claude Code
claude mcp add synaps -- npx synaps serve

# Install Claude Code hooks (automatic search augmentation)
npx synaps setup-hooks

# Or generate a focused AGENTS.md (no API key needed)
npx synaps init

MCP Tools (16)

Tool	What It Returns
get_commands	Build, test, lint commands with exact flags
get_architecture	Directory structure, entry points, execution flows
get_conventions	DO/DON'T rules from 13 detectors (error handling, async, state management, API patterns, hooks, testing, file naming, import ordering, frameworks, databases, data fetching, build tools)
get_workflow_rules	File coupling and co-change patterns
get_contribution_guide	How to add code, with inline examples and registration steps
get_exports	Public API with resolved TypeScript types (parameter types, return types)
analyze_impact	Blast radius: importers, callers, co-change partners, implicit coupling, co-change clusters, git history metadata
plan_change	Full change plan: dependents, co-changes, implicit coupling, execution flows, registration/barrel updates, tests. With optional symbols parameter: narrows dependents to files importing specific symbols.
get_test_info	Test file path + exact per-file run command
auto_register	Exact code insertions for registration + barrel updates
review_changes	Pattern compliance: suffix, imports, registration, barrel, tests
diagnose	Root cause analysis with confidence level (high/medium/low), import path traces, and 7 scoring signals
search	Find symbols, files, and conventions by name or concept — searches public API, internal functions, file paths, and conventions with call graph enrichment
rename	Find all references to a symbol for safe renaming — definition, imports, re-exports, call sites with checklist
get_module_doc	Structured per-directory documentation: files, exports, dependencies, dependents, call graph, flows, co-change partners, contribution patterns
list_packages	Monorepo package inventory

Every tool response includes next-step hints guiding the agent to the logical next action.

MCP Resources (5)

Resource	URI	Content
Conventions	synaps://conventions	DO/DON'T rules from all detectors
Processes	synaps://processes	Execution flows with confidence scores
Clusters	synaps://clusters	Co-change file groups (cliques)
Packages	synaps://packages	Package inventory with types and entry points
Schema	synaps://schema	Analysis data model reference

MCP Prompts (2)

Prompt	Description
analyze-impact	Guided workflow: identify changed files → plan_change → check clusters → summarize blast radius
onboard	Guided workflow: commands → architecture → conventions → schema overview

Multi-Repo Support

Serve multiple repositories from a single MCP server:

claude mcp add synaps -- npx synaps serve /path/to/repo1 /path/to/repo2

All tools accept an optional repo parameter. Single-repo usage is unchanged.

What Makes It Different

Git Co-Change Intelligence

Mines commit history using Jaccard similarity to find files that frequently change together. Produces workflow rules ("when modifying X, also check Y") and implicit coupling — file pairs that co-change but have no import relationship. This catches the "forgotten file" that static analysis misses.

Convention Detection (13 Detectors)

Extracts real coding patterns from AST analysis, not configured rules:

• Error handling: custom error classes, Result/Either patterns, typed error hierarchies
• Async patterns: Promise.all usage, sequential-await-in-loops detection, AbortController
• State management: Redux, Zustand, Jotai, MobX, Signals, Context API (reports all, not just dominant)
• API patterns: Express, Fastify, Hono, NestJS, tRPC, GraphQL (framework-aware, not directory heuristics)
• Plus: hooks, testing, file naming, import ordering, web frameworks, databases, data fetching, build tools

Symbol-Level Filtering

plan_change({ files: ["src/types.ts"], symbols: ["Convention"] })

Narrows 98 dependents to 25 — only files that actually import Convention. Every other tool shows all dependents regardless of which symbol you're changing.

Type-Aware Analysis

Opt-in --type-checking creates a TypeScript Program for resolved parameter types and return types:

analyze(options: Partial<ResolvedConfig> & { packages: string[] }): Promise<StructuredAnalysis>

Not text extraction — actual TypeChecker resolution through re-export chains.

Execution Flow Tracing

Detects execution paths from entry points through the call graph:

runPipeline → analyzePackage → buildSymbolGraph → computeImpactRadius (4 steps, 4 files)

Scored by co-change confidence — flows through files that frequently change together are higher confidence.

Validated Diagnose Tool

Tested against 95 real bug-fix commits across 10 repos:

• Unit-test repos: 83% recall@3 (root cause in top 3 suspects)
• All repos: 47% recall@3, 33% precision@1
• Confidence assessment based on signal quality and score discrimination

Claude Code Hooks

npx synaps setup-hooks

Installs PreToolUse + PostToolUse hooks:

• PreToolUse: When you grep for "validateUser", automatically shows callers, co-change partners, and execution flows alongside results
• PostToolUse: After git commit, detects when analysis cache is stale

Claude Code Setup

# Add MCP server
claude mcp add synaps -- npx synaps serve

# Install hooks for automatic search augmentation
npx synaps setup-hooks

With --type-checking for resolved TypeScript types:

claude mcp add synaps -- npx synaps serve --type-checking

AGENTS.md Generation

For tools without MCP support:

npx synaps init                         # Focused (~300 tokens, no API key)
npx synaps init --full                  # Comprehensive (needs API key)

Research-backed: focused context files improve AI accuracy by +4% and reduce runtime by 29%. LLM-generated comprehensive files hurt by -2%.

CLI Reference

synaps init [--full]                 Generate AGENTS.md
synaps serve [path] [options]        Start MCP server
synaps setup-hooks                   Install Claude Code hooks
synaps check                         Staleness detection for CI
synaps visualize [path]              Interactive codebase topology graph
synaps analyze [paths...] [options]  Analyze specific packages

Options:
  --type-checking    Enable resolved TypeScript types (requires tsconfig.json)
  --minimal          Focused output (<500 tokens, default for init)
  --full             Comprehensive output (requires ANTHROPIC_API_KEY)
  --format, -f       json | agents.md | claude.md | cursorrules
  --verbose, -v      Timing and analysis details
  --dry-run          Print to stdout (no file writes)

How It Works

18-stage analysis pipeline:

1. File Discovery → git ls-files with .gitignore respect
2. AST Parsing → TypeScript Compiler API for exports, imports, call references
3. Symbol Graph → barrel resolution, re-export chains, call graph construction
4. Import Chain → file-to-file coupling with confidence scores
5. Tier Classification → public API (T1), internal (T2), test/generated (T3)
6. Type Enrichment → optional ts.Program for resolved parameter/return types
7. Convention Extraction → 13 detectors with structured confidence metrics
8. Git History Mining → co-change mining with Jaccard similarity, adaptive thresholds
9. Implicit Coupling → co-change pairs with no import relationship
10. Execution Flow Tracing → entry point scoring, spine-first BFS, co-change validation
11. Impact Classification → high/medium/low impact scoring on conventions
12. Workspace Auto-Detection → monorepo roots expand to workspace packages

5 production dependencies. No native bindings, no graph database, no WASM. Installs instantly.

Stats

• 906 tests across 62 files
• 16 MCP tools + 5 resources + 2 prompts
• 13 convention detectors
• 95-commit diagnose validation corpus (10 repos)
• 10 execution flows detected on medium codebases
• Zero type errors, zero technology hallucinations

Library API

import { analyze, generateMinimalAgentsMd } from 'synaps';

const analysis = await analyze({
  packages: ['./'],
  typeChecking: true,  // optional: resolved TypeScript types
});

const agentsMd = generateMinimalAgentsMd(analysis);

Contributing

git clone https://github.com/msiric/synaps.git
cd synaps
npm install
npm test          # 906 tests
npm run typecheck # Zero errors
npm run build

See CONTRIBUTING.md for development workflow.

License

MIT