pi-cmp
Codebase Memory for pi. Provides cbm_* tools that answer structural questions about a codebase (symbol search, call chains, architecture, git change blast radius, and more) by querying a local tree-sitter knowledge graph, avoiding slow grep/read fallbacks.
README
pi-cmp · Codebase Memory
Codebase Memory tools (cbm_*) for pi. They answer structural questions about a codebase from a local tree-sitter knowledge graph — before pi falls back to slow grep/read loops.
The extension proxies each tool call to the codebase-memory-mcp server, which maintains a graph of symbols, call chains, and relationships per project.
Quick start
cd /path/to/this/repo
pi install "$(pwd)"
Then /reload in pi (or restart pi) and verify:
pi list
What's included
Extension tools only — no MCP configuration to maintain. The extension spawns and talks to the local codebase-memory-mcp server itself.
| Tool | Description |
|---|---|
cbm_search_graph |
Symbol search by name pattern or natural-language query |
cbm_trace_path |
Call chains: who calls a function (inbound), what it calls (outbound) |
cbm_get_code_snippet |
Full source for one symbol with metrics |
cbm_get_architecture |
High-level codebase overview (languages, entry points, hotspots) |
cbm_search_code |
Literal text / exact identifier search over indexed files |
cbm_get_graph_schema |
Node/edge labels and relationship patterns |
cbm_query_graph |
Read-only openCypher queries |
cbm_detect_changes |
Map git changes to affected symbols with blast radius |
cbm_list_projects |
All indexed projects with node/edge counts |
cbm_index_status |
Index health (indexed, stale, or missing) with git-HEAD freshness proof |
cbm_index_repository |
Index a repository (one-time; auto-sync keeps it fresh) |
cbm_delete_project |
Remove a project from the knowledge graph (requires UI confirmation) |
cbm_manage_adr |
Create, update, or read Architecture Decision Records |
cbm_ingest_traces |
Ingest runtime traces to validate service-to-service edges |
Requirements
- Node.js 22.19.0 or newer
- The
codebase-memory-mcpserver binary onPATH, or point theCBM_BINARYenvironment variable at it
codebase-memory-mcp is a separate project; install it from its own source before using these tools.
Usage
1. Start pi inside a project
cd /path/to/project
pi
2. Index the project (once)
Ask pi:
Index this project with cbm_index_repository.
Subsequent queries use the local .codebase-memory/ index in the project.
3. Ask structural questions
Good prompts:
Use Codebase Memory. Explain how authentication reaches the request handler.
Use Codebase Memory. What calls PlanBoostSession?
Use Codebase Memory. What would break if I change UserRepository?
Use Codebase Memory. Show files under internal/services and important symbols.
Tool choice
cbm_get_architecture/cbm_search_graphfor broad "how does this work?" questionscbm_trace_pathwhen you already know a function name and want callers/calleescbm_search_graphfor declarations and symbol names, not arbitrary textcbm_search_codefor literal text, exact identifiers, or distinctive stringscbm_detect_changesbefore refactors — blast radius of uncommitted changes
How it works
pi extensions are not MCP configuration files. This package registers native pi tools, and each tool proxies one JSON-RPC request to the local codebase-memory-mcp process over stdio:
pi agent
-> pi-cmp extension tool
-> codebase-memory-mcp process (cwd = your project)
-> .codebase-memory/ graph database in the current project
-> structured result back to pi
Additional behavior:
- On
session_startthe project cache is reset;before_agent_startinjects tool-choice guidance into the system prompt. - Tools default to the project matching pi's current working directory; pass
projectexplicitly (seecbm_list_projects) to query others. cbm_index_statuscompares the indexed git HEAD against the repository's current HEAD for real freshness proof (no fabricated timestamps).cbm_delete_projectrequires interactive confirmation when pi has a UI.- Long-running tools (
cbm_index_repository) use a 10-minute timeout; others default to 20 seconds. Errors surface as diagnostics with secrets redacted.
Development
npm install
npm run typecheck # tsc --noEmit
npm test # vitest run
Install the local checkout into pi:
pi install /path/to/pi-cmp
Repository layout
pi-cmp/
├── extensions/ # extension source only (logic lives here)
│ ├── index.ts # entry point: registers tools, injects guidance
│ ├── tools.ts # cbm_* tool definitions (typebox schemas)
│ ├── mcp-client.ts # spawns codebase-memory-mcp, JSON-RPC over stdio
│ ├── project.ts # project resolution from list_projects output
│ └── sync-check.ts # git-HEAD freshness annotation for index_status
├── __tests__/ # vitest tests (mirror the files above)
├── package.json # pi manifest: "pi": { "extensions": ["./extensions"] }
└── tsconfig.json
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。