synaptiq 2.0.3

Graph-powered code intelligence engine — indexes codebases into a knowledge graph, exposed via MCP tools for AI agents and a CLI for developers.

Graph-powered code intelligence engine — indexes your codebase into a knowledge graph and exposes it via MCP tools for AI agents and a CLI for developers.

$ synaptiq analyze .

Index ready.                             3,514 files — 22,689 symbols, 115k edges
Encoding 26,909 embeddings in the background — run synaptiq status to check.

Done in 16.1s   # 790k-LOC monorepo, cold start

$ vim src/api/billing.py                 # ...edit one file...
$ synaptiq analyze .

Index ready.
Incremental: 2 file(s) re-analyzed (1 changed + 1 dependent), 14 symbol(s) updated

Done in 0.5s

Most code intelligence tools treat your codebase as flat text. Synaptiq builds a structural graph — every function, class, import, call, type reference, and execution flow becomes a node or edge in a queryable knowledge graph. AI agents using Synaptiq don't just search for keywords; they understand how your code is connected.

---

Why Synaptiq?

For AI agents (Claude Code, Cursor, agent swarms):

• "What breaks if I change this function?" — blast radius via call graph + type references + git coupling
• "What code is never called?" — dead code detection with framework-aware exemptions
• "Show me the login flow end-to-end" — execution flow tracing from entry points through the call graph
• "Which files always change together?" — git history change coupling analysis

For developers:

• Instant answers to architectural questions without grepping through files
• Find dead code, tightly coupled files, and execution flows automatically
• Raw Cypher queries against your codebase's knowledge graph
• Watch mode that re-indexes on every save

Zero cloud dependencies. Everything runs locally — parsing, graph storage, embeddings, search. No API keys, no data leaving your machine.

---

⚡ Performance

All numbers measured on a 790k-LOC production monorepo (3,514 files / 22,689 symbols / ~27k embeddings), 10-core Apple Silicon laptop, v2.0.0 vs v1.5.1. Full methodology and raw data: docs/plans/benchmarks/2026-07-baseline.md.

Scenario	v1.5.1	v2.0.0	Change
Full analyze → usable index	726s	16.1s	45× faster (embeddings continue in background)
Re-analyze, nothing changed	726s	1.4s	534× faster (incremental engine)
Re-analyze, one file changed	726s	~0.5s + background delta encode	incremental + vector reuse
Storage load (no-embeddings full build)	5.99s*	2.58s*	−57% — Arrow COPY on LadybugDB
Watcher save-time cost	O(whole repo) FTS rebuild	O(file)	per-save full-text rebuild eliminated
Incremental insert path	13.9s / 1k nodes	0.29s	~47× — transactions + prepared statements
Embedding encode (optional fast tier)	237 texts/s	43,385 texts/s	183× — model2vec static vectors
Index size on disk	385 MB	239 MB	−38% (single-file format)

_{*measured on the synaptiq repo itself (122 files); other rows on the 790k-LOC monorepo.}

Correctness guarantees behind the speed: incremental results are property-tested equivalent to a full rebuild (50-seed randomized edit-script soak, strict-core graph equality); 1,599-test suite; every optimization landed with a frozen-reference equivalence test against the code it replaced.

---

Table of Contents

• Performance
• Features
• Supported Languages
• Installation
• Quick Start
• CLI Reference
• MCP Integration
• Multi-Agent Concurrency
• Knowledge Graph Model
• Architecture
• Example Workflows
• How It Compares
• Development

---

🔬 Features

11-Phase Analysis Pipeline

Synaptiq doesn't just parse your code — it builds a deep structural understanding through 11 sequential analysis phases:

Phase	What It Does
1. File Walking	Walks repo respecting .gitignore, filters by supported languages
2. Structure	Creates File/Folder hierarchy with CONTAINS relationships
3. Parsing	tree-sitter AST extraction — functions, classes, methods, interfaces, enums, type aliases
4. Import Resolution	Resolves import statements to actual files (relative, absolute, bare specifiers)
5. Call Tracing	Maps function calls with confidence scores (1.0 = exact match, 0.5 = fuzzy)
6. Heritage	Tracks class inheritance (EXTENDS) and interface implementation (IMPLEMENTS)
7. Type Analysis	Extracts type references from parameters, return types, and variable annotations
8. Community Detection	Leiden algorithm clusters related symbols into functional communities
9. Process Detection	Framework-aware entry point detection + BFS flow tracing
10. Dead Code Detection	Multi-pass analysis with override, protocol, and decorator awareness
11. Change Coupling	Git history analysis — finds files that always change together

Hybrid Search (BM25 + Vector + RRF)

Three search strategies fused with Reciprocal Rank Fusion:

• BM25 full-text search — fast exact name and keyword matching via LadybugDB FTS
• Semantic vector search — conceptual queries via embeddings; two selectable model tiers (see "Embedding model tiers" below), 384-dim quality (default) or 256-dim fast
• Fuzzy name search — Levenshtein fallback for typos and partial matches

Test files are automatically down-ranked (0.5x), source-level functions/classes boosted (1.2x).

Dead Code Detection

Finds unreachable symbols with intelligence — not just "zero callers" but a multi-pass analysis:

1. Initial scan — flags functions/methods/classes with no incoming calls
2. Exemptions — entry points, exports, constructors, test code, dunder methods, __init__.py public symbols, decorated functions, @property methods
3. Override pass — un-flags methods that override non-dead base class methods (handles dynamic dispatch)
4. Protocol conformance — un-flags methods on classes conforming to Protocol interfaces
5. Protocol stubs — un-flags all methods on Protocol classes (interface contracts)
6. Ruby exemptions — initialize constructors, metaprogramming hooks (method_missing, inherited, ...), attr_*/Rails-callback macro methods, and Rails framework base classes (ApplicationRecord/ApplicationController/...)
7. Go exemptions — main/init runtime entries, exported (upper-case) identifiers, and _test.go files (Test*/Benchmark*/Fuzz*/Example* helpers)

Impact Analysis (Blast Radius)

When you're about to change a symbol, Synaptiq traces upstream through:

• Call graph — every function that calls this one, recursively (batched BFS — one query per depth level)
• Type references — every function that takes, returns, or stores this type
• Git coupling — files that historically change alongside this one

Community Detection

Uses the Leiden algorithm (igraph + leidenalg) to automatically discover functional clusters in your codebase. Each community gets a cohesion score and auto-generated label based on member file paths.

Execution Flow Tracing

Detects entry points using framework-aware patterns:

• Python: @app.route, @router.get, @click.command, test_* functions, __main__ blocks
• JavaScript/TypeScript: Express handlers, exported functions, handler/middleware patterns
• Ruby: Rails controller/job/mailer actions, Sinatra/Rails route blocks, RSpec *_spec.rb / *_test.rb, Rakefile/config.ru
• Go: main/init functions, exported package symbols, Test*/Benchmark*/Fuzz* functions

Then traces BFS execution flows from each entry point through the call graph, classifying flows as intra-community or cross-community.

Change Coupling (Git History)

Analyzes 6 months of git history to find hidden dependencies that static analysis misses:

coupling(A, B) = co_changes(A, B) / max(changes(A), changes(B))

Files with coupling strength >= 0.3 and 3+ co-changes get linked. Surfaces coupled files in impact analysis.

Watch Mode

Live re-indexing powered by a Rust-based file watcher (watchfiles):

$ synaptiq watch
Watching /Users/you/project for changes...

[10:32:15] src/auth/validate.py modified -> re-indexed (0.3s)
[10:33:02] 2 files modified -> re-indexed (0.5s)

• File-local phases run immediately on change — parsing happens without holding any lock
• Global phases (communities, processes, dead code) batch every 30 seconds
• The CPU-intensive parse step and the I/O-intensive storage write are split so locks are held for the shortest possible window

Branch Comparison

Structural diff between branches using git worktrees (no stashing required):

$ synaptiq diff main..feature

Symbols added (4):
  + process_payment (Function) -- src/payments/stripe.py
  + PaymentIntent (Class) -- src/payments/models.py

Symbols modified (2):
  ~ checkout_handler (Function) -- src/routes/checkout.py

Symbols removed (1):
  - old_charge (Function) -- src/payments/legacy.py

---

🌐 Supported Languages

Language	Extensions	Parser
Python	.py	tree-sitter-python
TypeScript	.ts, .tsx	tree-sitter-typescript
JavaScript	.js, .jsx, .mjs, .cjs	tree-sitter-javascript
Ruby	.rb, .rake, .gemspec, .ru, .rbi (+ Rakefile, Gemfile, Guardfile, Capfile, Vagrantfile, Brewfile, Podfile)	tree-sitter-ruby
Go	.go	tree-sitter-go

---

📦 Installation

# With uv (recommended) — installs as a global CLI tool
uv tool install synaptiq

# With pip
pip install synaptiq

# As a project dependency
uv add synaptiq

# With Neo4j backend support
pip install synaptiq[neo4j]

# With the fast embedding tier (--embedding-model fast)
pip install synaptiq[fast-embeddings]

Requires Python 3.11+.

From Source

git clone https://github.com/scanadi/synaptiq.git
cd synaptiq
uv sync --all-extras
uv run synaptiq --help

---

🚀 Quick Start

1. Index Your Codebase

cd your-project
synaptiq analyze .

2. Query It

# Search for symbols
synaptiq query "authentication handler"

# Get full context on a symbol
synaptiq context validate_user

# Check blast radius before changing something
synaptiq impact UserModel --depth 3

# Find dead code
synaptiq dead-code

# Run a raw Cypher query
synaptiq cypher "MATCH (n:Function) WHERE n.is_dead = true RETURN n.name, n.file_path"

3. Keep It Updated

# Watch mode — re-indexes on every save
synaptiq watch

# Or re-analyze manually
synaptiq analyze .

---

🖥 CLI Reference

synaptiq analyze [PATH]          Index a repository (default: current directory)
    --full                       Force full rebuild (skip incremental)
    --embeddings MODE            lazy (default): index ready in seconds, then
                                  encode vectors in the background; sync: encode
                                  inline before returning; off: skip vectors.
                                  (--no-embeddings is a deprecated alias for off.)
    --embedding-model TIER       quality (default): BAAI/bge-small-en-v1.5, 384-dim,
                                  fastembed/ONNX. fast: minishlab/potion-base-8M,
                                  256-dim, model2vec static embeddings — ~180x faster
                                  to encode; requires `synaptiq[fast-embeddings]`. See
                                  "Embedding model tiers" below.
    --jobs / -j N                Cap LadybugDB threads, ONNX embedding threads,
                                  and walk/parse worker pools to N (0 = explicit
                                  all-cores). See "CPU usage" below.

synaptiq status                  Show index status (incl. background embedding
                                  progress: encoding N/M, complete, deferred)
synaptiq list                    List all indexed repositories
synaptiq clean                   Delete index for current repo
    --force / -f                 Skip confirmation prompt

synaptiq query QUERY             Hybrid search the knowledge graph
    --limit / -n N               Max results (default: 20)

synaptiq context SYMBOL          360-degree view of a symbol
synaptiq impact SYMBOL           Blast radius analysis
    --depth / -d N               BFS traversal depth (default: 3)

synaptiq dead-code               List all detected dead code
synaptiq cypher QUERY            Execute a raw Cypher query (read-only)

synaptiq watch                   Watch mode — live re-indexing on file changes
synaptiq diff BASE..HEAD         Structural branch comparison

synaptiq setup                   Print MCP configuration JSON
    --claude                     For Claude Code
    --cursor                     For Cursor
    --http                       Show HTTP transport config
    --port N                     Port for HTTP config (default: 8080)

synaptiq mcp                     Start the MCP server (stdio transport)
synaptiq serve                   Start MCP server with daemon features
    --watch, -w                  Enable live file watching with auto-reindex
    --transport, -t TRANSPORT    Transport: stdio (default) or http
    --host HOST                  Host for HTTP transport (default: 127.0.0.1)
    --port PORT                  Port for HTTP transport (default: 8080)
synaptiq --version               Print version

CPU usage during analyze

By default, analyze keeps LadybugDB at its library default (all cores) but caps ONNX embedding threads to a polite max(2, cores - 2), so a foreground index doesn't lock up the rest of the machine. --jobs N overrides this:

Setting	Engine threads	ONNX embed threads	Walk/parse pools
default (no flag, no env)	all cores (library default)	max(2, cores - 2)	min(8, cores)
SYNAPTIQ_DB_THREADS / SYNAPTIQ_EMBED_THREADS env vars	as set	as set	unaffected
--jobs N (N > 0)	N	N	N
--jobs 0	all cores (library default)	all cores (library default)	min(8, cores)

SYNAPTIQ_DB_MEMORY_MB caps the engine buffer pool. SYNAPTIQ_KUZU_THREADS / SYNAPTIQ_KUZU_MEMORY_MB still work as deprecated aliases (they log a one-time warning) for one release.

Precedence is --jobs flag > environment variables > profile defaults. serve/mcp/watch are unaffected — those long-running daemons always use the stricter "server" profile regardless of --jobs.

Background embeddings (--embeddings lazy, default)

Encoding vectors dominates a cold index — on a large repo it can be ~98% of the total time. By default analyze therefore commits the graph first and returns a fully queryable index in seconds (keyword + fuzzy search work immediately), then encodes vectors in a detached background worker that fills in semantic search behind you.

synaptiq analyze .            # index ready in seconds; vectors fill in later
synaptiq status               # → Embeddings: encoding 12,431/26,203
                              #   …then → Embeddings: 26,203 (complete)

• --embeddings sync blocks until vectors are stored (the pre-1.x behavior).
• --embeddings off skips vectors entirely (--no-embeddings is a deprecated alias).
• The worker stays polite (max(2, cores // 2) ONNX threads) and never fights a running serve/watch daemon for the database — if the index is busy it defers, and the next analyze/rebuild encodes. serve/watch themselves always embed synchronously.
• Repeat analyze runs reuse unchanged vectors instead of re-encoding everything: only symbols whose text actually changed hit the ONNX model, so a small edit re-encodes a handful of vectors in the background — or none at all, in which case no background worker even starts.

Embedding model tiers (--embedding-model)

Two named tiers trade encode speed for representation quality — pick whichever fits the machine you're indexing on:

Tier	Model	Dim	Backend	Measured encode rate*	Dependency
quality (default)	BAAI/bge-small-en-v1.5	384	fastembed/ONNX	~235 texts/sec	core (fastembed)
fast	minishlab/potion-base-8M	256	model2vec (static embeddings)	~43,000 texts/sec (~180x)	synaptiq[fast-embeddings]

* CPU, single-threaded, measured on ~2k synthetic code-description texts of the shape embed_graph generates. fast has no transformer forward pass at all (token lookup + mean-pool, no ONNX) — it trades some retrieval quality for an order-of-magnitude speedup, which matters most on CI runners and low-power machines where even a lightweight ONNX session is expensive.

# Install the optional fast-tier dependency (pure Python + numpy — no torch,
# no onnxruntime):
uv sync --extra fast-embeddings          # or: pip install 'synaptiq[fast-embeddings]'

synaptiq analyze . --embedding-model fast

• The tier is recorded per index (meta.json's stats.embedding_model) — every later analyze, background worker cycle, and query automatically uses the same model the index was built with. A query is never encoded with a different tier than the one that produced the stored vectors (they have different vector widths — LadybugDB raises a clear error, telling you to re-analyze, rather than silently returning wrong or zero results).
• Omitting --embedding-model always defaults to quality, even on a repo previously indexed with fast — pass the flag explicitly on every analyze call if you want to stay on fast.
• Switching tiers always forces a full re-encode on that run: vectors from one tier are never reused for another (they're different-width, differently distilled models), so cached-vector reuse (see above) intentionally treats a tier switch like every symbol's text changed at once.
• serve --watch has no per-cycle flag — its routine rebuilds and the analyze --embeddings lazy background worker always keep using whatever tier the index's meta.json already records.

---

🤖 MCP Integration

Synaptiq exposes its full intelligence as an MCP server, giving AI agents like Claude Code and Cursor deep structural understanding of your codebase.

Setup for Claude Code

Add to your .claude/settings.json or project .mcp.json:

{
  "mcpServers": {
    "synaptiq": {
      "command": "synaptiq",
      "args": ["serve", "--watch"]
    }
  }
}

This starts the MCP server with live file watching — the knowledge graph updates automatically as you edit code. To run without watching, use "args": ["mcp"] instead.

Or run the setup helper:

synaptiq setup --claude

HTTP Transport

For remote or browser-based MCP clients, Synaptiq supports Streamable HTTP transport:

# Start with HTTP transport
synaptiq serve --watch --transport http --port 8080

# Get the MCP config snippet
synaptiq setup --http --port 8080

{
  "synaptiq": {
    "url": "http://127.0.0.1:8080/mcp"
  }
}

Claude Code Skill (Optional)

For richer Claude Code integration, copy the bundled skill into your project:

cp -r "$(python -c 'import synaptiq; import pathlib; print(pathlib.Path(synaptiq.__file__).parent.parent.parent)')"/.claude/skills/synaptiq .claude/skills/

Or manually copy .claude/skills/synaptiq/SKILL.md from the Synaptiq repo. This teaches Claude when and how to use Synaptiq's tools, the knowledge graph schema, Cypher query patterns, and investigation workflows.

Claude Code Plugin

For a batteries-included Claude Code experience, install the Synaptiq Claude Plugin — it bundles the MCP server, skill, setup command, and a session hook in a single install:

claude plugins add scanadi/synaptiq-claude-plugin

Then run /synaptiq:setup to install Synaptiq, index your codebase, and configure the MCP connection automatically.

Setup for Cursor

Add to your Cursor MCP settings:

{
  "synaptiq": {
    "command": "synaptiq",
    "args": ["serve", "--watch"]
  }
}

Or run:

synaptiq setup --cursor

MCP Tools

Once connected, your AI agent gets access to these tools:

Tool	Description
synaptiq_list_repos	List all indexed repositories with stats
synaptiq_query	Hybrid search (BM25 + vector + fuzzy) across all symbols
synaptiq_context	360-degree view — callers, callees, type refs, community, processes
synaptiq_impact	Blast radius — all symbols affected by changing the target
synaptiq_dead_code	List all unreachable symbols grouped by file
synaptiq_detect_changes	Map a git diff to affected symbols in the graph
synaptiq_cypher	Execute read-only Cypher queries against the knowledge graph

Every tool response includes a next-step hint guiding the agent through a natural investigation workflow:

query   -> "Next: Use context() on a specific symbol for the full picture."
context -> "Next: Use impact() if planning changes to this symbol."
impact  -> "Tip: Review each affected symbol before making changes."

MCP Resources

Resource URI	Description
synaptiq://overview	Node and relationship counts by type
synaptiq://dead-code	Full dead code report
synaptiq://schema	Graph schema reference for writing Cypher queries

Index Freshness

Every tool response (and the synaptiq://overview resource) ends with a compact freshness trailer, so an agent can tell at a glance whether the graph it just queried might be stale — without a separate round trip to synaptiq status:

[index: 4m old · embeddings: complete]
[index: 12s old · embeddings: encoding 12431/26203]
[index: 2h old · embeddings: failed]
[index: age unknown]

It reads last_indexed_at from meta.json for the index age and .synaptiq/embeddings_state.json (written by the background embedding worker — see "Background embeddings" above) for the vector state, falling back to the stored vector count when no worker has run. The trailer is omitted entirely when there's no index yet, so the existing "no index found" errors are unaffected. Set SYNAPTIQ_MCP_FRESHNESS=0 to disable it.

---

🔄 Multi-Agent Concurrency

Synaptiq is designed for environments where multiple AI agents hit the same codebase simultaneously — Claude Code with sub-agents, Cursor with background indexing, or any swarm-of-agents setup.

Primary/Proxy Architecture

flowchart TB
    A1(["🤖 Agent 1"])
    A2(["🤖 Agent 2"])
    A3(["🤖 Agent 3"])
    AN(["🤖 Agent N"])

    subgraph PRIMARY ["🟢 Primary Instance"]
        MCP["📡 MCP Server"]
        FW["👁️ File Watcher<br/>watchfiles"]
        RWL{{"🔒 AsyncRWLock<br/>concurrent reads · exclusive writes"}}
        DB[(🗄️ LadybugDB + Pool<br/>8 read connections)]
        SOCK["🔌 Unix Socket Server<br/>16 concurrent slots"]
        MCP --> RWL
        FW --> RWL
        RWL --> DB
        SOCK --> RWL
    end

    P1["🔵 Proxy 1<br/>MCP + Socket Client"]
    P2["🔵 Proxy 2<br/>MCP + Socket Client"]
    PN["🔵 Proxy N ···"]

    A1 -->|MCP stdio| MCP
    A2 --> P1
    A3 --> P2
    AN --> PN
    P1 -->|unix socket| SOCK
    P2 -->|unix socket| SOCK
    PN -->|unix socket| SOCK

    style PRIMARY fill:#ecfdf5,stroke:#10b981,stroke-width:2px
    classDef agent fill:#fbbf24,stroke:#d97706,color:#000,stroke-width:2px
    classDef server fill:#6ee7b7,stroke:#10b981,color:#000,stroke-width:2px
    classDef lock fill:#c4b5fd,stroke:#8b5cf6,color:#000,stroke-width:2px
    classDef database fill:#67e8f9,stroke:#06b6d4,color:#000,stroke-width:2px
    classDef socket fill:#fda4af,stroke:#f43f5e,color:#000,stroke-width:2px
    classDef proxy fill:#93c5fd,stroke:#3b82f6,color:#000,stroke-width:2px

    class A1,A2,A3,AN agent
    class MCP,FW server
    class RWL lock
    class DB database
    class SOCK socket
    class P1,P2,PN proxy

How it works:

1. The first synaptiq serve --watch becomes the primary — it owns the database, runs the file watcher, and opens a Unix domain socket
2. Subsequent instances detect the lock and become proxies — they forward all queries through the socket to the primary
3. Role detection is automatic via fcntl.flock() — no configuration needed
4. If the primary crashes, the OS releases the lock and the next instance takes over

Concurrency Model

Layer	Mechanism	Purpose
Storage reads	Connection pool (8 ladybug.Connection objects)	Multiple threads query in parallel
Storage writes	Dedicated write connection + AsyncRWLock	Exclusive access during index updates
MCP dispatch	AsyncRWLock.reader() with 120s timeout	Concurrent tool/resource calls
Socket server	asyncio.Semaphore(16) backpressure	Limits concurrent proxy dispatches
Socket client	Request-ID multiplexing + asyncio.Future	Concurrent calls on a single connection
File watcher	Split parse/write pattern	CPU work runs lock-free; lock held only for I/O
Process lock	fcntl.flock() on .synaptiq/synaptiq.lock	Primary election, auto-released on crash
Proxy recovery	Auto-reconnect (3 attempts, exponential backoff)	Handles primary restarts gracefully

Lock Strategy: Write-Preferring RWLock

The AsyncRWLock allows multiple readers to query simultaneously while giving priority to writers:

• Reads (agent queries): Run concurrently — multiple agents can search, inspect context, and analyze impact at the same time
• Writes (file watcher updates): Get exclusive access, and pending writes block new reads to prevent writer starvation
• Timeout: All lock acquisitions have a 60-second timeout to prevent deadlocks
• Split processing: The watcher parses files without the lock (CPU-intensive), then acquires the write lock only for the storage I/O step

macOS Compatibility

Unix domain socket paths are limited to 104 bytes on macOS. When the .synaptiq/synaptiq.sock path exceeds 100 bytes (common with deep directory nesting), Synaptiq automatically falls back to /tmp/synaptiq-<hash>.sock using a truncated SHA-256 hash of the data directory.

---

📊 Knowledge Graph Model

Nodes

Label	Description
File	Source file
Folder	Directory
Function	Top-level function
Class	Class definition (incl. Go struct)
Method	Method within a class (incl. Go receiver funcs)
Module	Ruby module (namespace / mixin) or Go package
Interface	Interface / Protocol definition (incl. Go interface)
TypeAlias	Type alias (incl. Go type X = Y / type X Y)
Enum	Enumeration
Community	Auto-detected functional cluster
Process	Detected execution flow

Relationships

Type	Description	Key Properties
CONTAINS	Folder -> File/Symbol hierarchy	—
DEFINES	File -> Symbol it defines	—
CALLS	Symbol -> Symbol it calls	confidence (0.0-1.0)
IMPORTS	File -> File it imports from	symbols (names list)
EXTENDS	Class -> Class it extends (incl. Go struct/interface embedding)	—
IMPLEMENTS	Class -> Interface it implements	—
MIXES_IN	Class/Module -> Ruby module it mixes in (include/extend/prepend)	—
USES_TYPE	Symbol -> Type it references	role (param/return/variable)
EXPORTS	File -> Symbol it exports	—
MEMBER_OF	Symbol -> Community it belongs to	—
STEP_IN_PROCESS	Symbol -> Process it participates in	step_number
COUPLED_WITH	File -> File that co-changes with it	strength, co_changes

Node ID Format

{label}:{relative_path}:{symbol_name}

Examples:
  function:src/auth/validate.py:validate_user
  class:src/models/user.py:User
  method:src/models/user.py:User.save
  file:src/auth/validate.py:

---

🏗 Architecture

flowchart TB
    SRC(["📁 Source Code<br/>.py · .ts · .js · .tsx · .jsx · .rb · .go"])

    PIPELINE["⚙️ Ingestion Pipeline — 12 Phases<br/>walk → structure → parse → imports → calls<br/>rest_linking → heritage → types → communities<br/>processes → dead_code → coupling"]

    KG(["🧠 KnowledgeGraph<br/>in-memory during build"])

    KUZU[(LadybugDB<br/>graph)]
    FTS[(FTS<br/>BM25)]
    VEC[(Vector<br/>HNSW)]

    PROTO["⬡ StorageBackend Protocol"]

    MCP_OUT["📡 MCP Server<br/>stdio + HTTP transport"]
    CLI_OUT["⌨️ CLI<br/>Typer + Rich"]

    CLAUDE(["🤖 Claude Code · Cursor"])
    DEV(["👤 Developer Terminal"])

    SRC --> PIPELINE --> KG
    KG --> KUZU
    KG --> FTS
    KG --> VEC
    KUZU --> PROTO
    FTS --> PROTO
    VEC --> PROTO
    PROTO --> MCP_OUT
    PROTO --> CLI_OUT
    MCP_OUT --> CLAUDE
    CLI_OUT --> DEV

    classDef source fill:#d8b4fe,stroke:#a855f7,color:#000,stroke-width:2px
    classDef pipeline fill:#fed7aa,stroke:#f97316,color:#000,stroke-width:2px
    classDef kg fill:#6ee7b7,stroke:#10b981,color:#000,stroke-width:2px
    classDef store fill:#93c5fd,stroke:#3b82f6,color:#000,stroke-width:2px
    classDef proto fill:#cbd5e1,stroke:#64748b,color:#000,stroke-width:2px
    classDef iface fill:#5eead4,stroke:#14b8a6,color:#000,stroke-width:2px
    classDef user fill:#fbbf24,stroke:#d97706,color:#000,stroke-width:2px

    class SRC source
    class PIPELINE pipeline
    class KG kg
    class KUZU,FTS,VEC store
    class PROTO proto
    class MCP_OUT,CLI_OUT iface
    class CLAUDE,DEV user

Tech Stack

Layer	Technology	Purpose
Parsing	tree-sitter	Language-agnostic AST extraction
Graph Storage	LadybugDB	Embedded graph database (KuzuDB successor) with Cypher, FTS, and vector support
Graph Algorithms	igraph + leidenalg	Leiden community detection
Embeddings	fastembed (quality) / model2vec (fast, optional)	ONNX 384-dim vectors, or static (no-ONNX) 256-dim vectors ~180x faster to encode — see "Embedding model tiers"
MCP Protocol	mcp SDK	AI agent communication via stdio and HTTP
CLI	Typer + Rich	Terminal interface with progress bars
File Watching	watchfiles	Rust-based file system watcher
Gitignore	pathspec	Full .gitignore pattern matching

Storage

Everything lives locally in your repo:

your-project/
+-- .synaptiq/
    +-- kuzu           # LadybugDB database (graph + FTS + vectors); the path
    |                  # name is retained so an index from the former KuzuDB
    |                  # backend is detected on open and rebuilt in place
    +-- meta.json      # Index metadata and stats
    +-- synaptiq.lock  # Process lock (present while serve is running)

Add .synaptiq/ to your .gitignore.

The storage layer is abstracted behind a StorageBackend Protocol — LadybugDB (the actively-maintained KuzuDB successor) is the default, with an optional Neo4j backend available via pip install synaptiq[neo4j].

---

💡 Example Workflows

"I need to refactor the User class — what breaks?"

# See everything connected to User
synaptiq context User

# Check blast radius
synaptiq impact User --depth 3

# Find which files always change with user.py
synaptiq cypher "MATCH (a:File)-[r:CodeRelation]->(b:File) \
  WHERE a.name = 'user.py' AND r.rel_type = 'coupled_with' \
  RETURN b.name, r.strength ORDER BY r.strength DESC"

"Is there dead code we should clean up?"

synaptiq dead-code

"What are the main execution flows in our app?"

synaptiq cypher "MATCH (p:Process) RETURN p.name, p.properties ORDER BY p.name"

"Which parts of the codebase are most tightly coupled?"

synaptiq cypher "MATCH (a:File)-[r:CodeRelation]->(b:File) \
  WHERE r.rel_type = 'coupled_with' \
  RETURN a.name, b.name, r.strength \
  ORDER BY r.strength DESC LIMIT 20"

"What changed in this PR and what does it affect?"

# Structural diff
synaptiq diff main..feature

# Or pipe a git diff through detect_changes (via MCP)
git diff main | synaptiq detect-changes

---

⚖ How It Compares

Capability	grep/ripgrep	LSP	Synaptiq
Text search	Yes	No	Yes (hybrid BM25 + vector)
Go to definition	No	Yes	Yes (graph traversal)
Find all callers	No	Partial	Yes (full call graph with confidence)
Type relationships	No	Yes	Yes (param/return/variable roles)
Dead code detection	No	No	Yes (multi-pass, framework-aware)
Execution flow tracing	No	No	Yes (entry point -> flow)
Community detection	No	No	Yes (Leiden algorithm)
Change coupling (git)	No	No	Yes (6-month co-change analysis)
Impact analysis	No	No	Yes (calls + types + git coupling)
AI agent integration	No	Partial	Yes (full MCP server)
Multi-agent concurrency	No	No	Yes (RWLock + connection pool)
Structural branch diff	No	No	Yes (node/edge level)
Watch mode	No	Yes	Yes (Rust-based, 500ms debounce)
Works offline	Yes	Yes	Yes

---

🛠 Development

git clone https://github.com/scanadi/synaptiq.git
cd synaptiq
uv sync --all-extras

# Run tests (581 tests)
uv run pytest

# Run only fast unit tests (skip e2e)
uv run pytest tests/core/ tests/cli/ tests/mcp/

# Lint
uv run ruff check src/ tests/

# Format
uv run ruff format src/ tests/

# Run from source
uv run synaptiq --help

Project Structure

src/synaptiq/
+-- cli/
|   +-- main.py                    # Typer CLI (analyze, query, serve, watch, ...)
|   +-- update_check.py            # Non-blocking PyPI update notifier
+-- mcp/
|   +-- server.py                  # MCP server (tools + resources + dispatch)
|   +-- http_transport.py          # Streamable HTTP MCP transport (Starlette + uvicorn)
|   +-- tools.py                   # Tool handler implementations
|   +-- resources.py               # Resource handler implementations
+-- core/
|   +-- ingestion/
|   |   +-- pipeline.py            # 11-phase pipeline orchestrator
|   |   +-- walker.py              # File discovery respecting .gitignore
|   |   +-- structure.py           # Folder/File hierarchy
|   |   +-- parser_phase.py        # tree-sitter AST extraction
|   |   +-- imports.py             # Import resolution
|   |   +-- calls.py               # Call tracing with confidence
|   |   +-- heritage.py            # Inheritance and implementation
|   |   +-- types.py               # Type reference extraction
|   |   +-- community.py           # Leiden clustering
|   |   +-- processes.py           # Entry point + flow detection
|   |   +-- dead_code.py           # Multi-pass dead code analysis
|   |   +-- coupling.py            # Git co-change analysis
|   |   +-- watcher.py             # File watcher with split parse/write
|   +-- storage/
|   |   +-- base.py                # StorageBackend protocol
|   |   +-- ladybug_backend.py     # LadybugDB implementation + connection pool
|   +-- daemon/
|   |   +-- rwlock.py              # Write-preferring AsyncRWLock
|   |   +-- lock.py                # fcntl.flock() based process lock
|   |   +-- socket_server.py       # Unix socket server (primary)
|   |   +-- socket_client.py       # Unix socket client (proxy)
|   +-- search/hybrid.py           # BM25 + vector + fuzzy fusion
|   +-- parsers/                   # Language-specific tree-sitter parsers
+-- config/
    +-- languages.py               # Language registry
    +-- ignore.py                  # Gitignore loading

See CONTRIBUTING.md for detailed contribution guidelines.

---

Acknowledgments

Synaptiq was originally inspired by and built upon axon by @harshkedia177. This project has since evolved into an independent codebase with its own architecture, features, and direction.

---

License

MIT — see LICENSE.