CodeSynapse

CodeSynapse indexes a repository into one property graph and answers small, precise questions from it. A coding assistant asks the graph ("who calls ReplaceRepo?", "what does this diff break?") instead of opening whole files. That cuts tokens and gives better review context.

On this repository the context an agent needs per question drops by a median 92%, measured by codesynapse stats.

What goes into the graph

Source	What is extracted
Go	Packages, files, functions, methods, types, exact calls and interface dispatch (go/types)
Python, JS/TS, Vue, Java, Kotlin, Rust, C#, PHP, Ruby	Structural symbols, imports, name-based call edges, tests
Markdown	Sections with line ranges; MENTIONS edges to exported symbols
YAML, Helm, Go templates	Kubernetes resources, key paths, {{define}} blocks, .Values links
Git history	Commits, authors (e-mail hashed), per-file churn
Dependencies	Go, npm, PyPI, Rust, PHP, Ruby, .NET and Gradle modules with licences and CVEs (via osv.dev)
Security rules	Findings per rule, linked to the owning function

Big picture

flowchart LR
    subgraph Sources
        R[Repository checkout]
        G[Git remote]
    end
    subgraph Ingest
        P[Pipeline<br/>Go, lang, docs, config, git]
        V[Vuln and licence lookup]
    end
    DB[(Graph store<br/>HugeGraph + Milvus)]
    subgraph Readers
        MCP[MCP server<br/>:8090]
        API[JSON API<br/>:8080]
        CLI[codesynapse CLI]
    end
    A[AI agents]
    W[Web dashboard<br/>:3000]
    CI[CI / developer shell]

    R --> P
    G --> P
    V --> P
    P --> DB
    DB --> MCP --> A
    DB --> API --> W
    DB --> CLI --> CI

Where to start

• I want to understand the system: Architecture overview.
• I want to know what is stored: Graph schema.
• I am wiring up an agent: Agent context (MCP), then the Agent guide. For Claude Code, start with the Claude Code bridge.
• I am reviewing security posture: Security review and Security model.
• I am deploying it: Running the stack.

Design principles

1. Deterministic answers. Graph tools never call an LLM; the agent using them is the LLM.
2. Read-only by default. Only review tools and opt-in agent memory write, and only to their own stores.
3. No source bodies in the graph. Hashes, signatures, key paths, line ranges and redacted summaries only.
4. Untrusted input everywhere. Repositories, model output and agent text are treated as hostile.
5. Tenant isolation in the data. Every API route and MCP tool is scoped to the caller's organisation.