一个自用的 Rust MCP 服务器，为 AI 编码代理在会话间持久的代码理解和内存提供。作为个人工具构建，用于探索当你赋予 LLM 对代码库符号图、会话状态和累积项目知识的结构化访问时会发生什么。

它的作用

Codegraph 作为 MCP 服务器（stdio transport）运行，并公开了 AI 代理可以调用的 26 个工具：

• 代码图 — 用树状分析源代码（Rust、TypeScript、JavaScript、Python、Go），提取符号及其关系（调用、导入、继承），并以有向图形式存储。代理可以搜索符号、遍历依赖关系并理解文件结构，而无需读取整个文件。

• 会话内存 — 跟踪代理当前任务、子任务、决策和工作上下文。经由上下文窗口压缩保存，代理可以从中断处继续。

• 学习系统 — 记录模式（有效的方法）、失败（需要避免的陷阱）和解决方案的谱系（尝试链与结果）。一个反映引擎将结果转化为可重复使用的知识。一个建议系统将这三者结合起来，为新任务推荐方法。

Codegraph MCP

A self-use Rust MCP server that gives AI coding agents persistent code understanding and memory across sessions. Built as a personal tool to explore what happens when you give an LLM structured access to a codebase’s symbol graph, session state, and accumulated project knowledge.

What It Does

Codegraph runs as an MCP server (stdio transport) and exposes 26 tools that an AI agent can call:

• Code Graph — Parses source code with tree-sitter (Rust, TypeScript, JavaScript, Python, Go), extracts symbols and their relationships (calls, imports, inherits), and stores them as a directed graph. The agent can search symbols, traverse dependencies, and understand file structure without reading entire files.

• Session Memory — Tracks the agent’s current task, subtasks, decisions, and working context. Survives context window compaction so the agent can resume where it left off.

• Learning System — Records patterns (things that worked), failures (gotchas to avoid), and solution lineage (attempt chains with outcomes). A reflection engine converts outcomes into reusable knowledge. A suggestion system combines all three to recommend approaches for new tasks.

• Skill Distillation — Generates a SKILL.md from accumulated patterns, failures, and conventions — a machine-readable summary of project-specific knowledge.

• Cross-Language Inference — Detects REST/GraphQL calls in frontend code and matches them to backend route definitions.

• Bash Compression — Compresses verbose command output (git status, test results, directory listings) to reduce token usage.

Tech Stack

Architecture

src/
├── main.rs              # Entry point
├── config.rs            # Project root detection, config.toml
├── mcp/                 # MCP protocol layer
│   ├── protocol.rs      # JSON-RPC 2.0 + MCP types
│   ├── transport.rs     # Stdio transport
│   ├── server.rs        # Request dispatch, lazy init
│   └── tools.rs         # Tool registry (26 tools)
├── store/               # Persistence
│   ├── db.rs            # SQLite CRUD
│   ├── graph.rs         # In-memory petgraph
│   └── migrations.rs    # Schema versioning
├── code/                # Code analysis
│   ├── parser.rs        # tree-sitter symbol extraction
│   ├── indexer.rs       # Incremental indexing + cross-file resolution
│   ├── languages.rs     # Language configs + grammars
│   └── cross_language.rs
├── session/             # Session state machine
│   └── state.rs         # Task, decisions, context tracking
├── learning/            # Learning system