MCP Markdown 任务与记忆服务

backlog-mcp 是一个本地优先、基于 Markdown 的智能体上下文与记忆管理服务,统一管理任务、决策与记忆,并通过混合搜索和渐进式信息展示服务于 MCP 客户端、CLI 及只读查看器中的开发者工作流。

  • by gkoreli
  • TypeScript
  • 本地部署
  • 5 星标
  • 更新于
在 GitHub 查看 gkoreli/backlog-mcp

使用说明

backlog-mcp

A local-first, Markdown-backed engine for agent context and memory. Tasks, decisions and memories share one store across MCP clients, the CLI and a live read-only viewer. Humans can read, edit and diff the files directly.

Entities are Markdown with YAML frontmatter. A project’s substrate declarations own its types, fields and workflows. Local hybrid search combines text and embeddings; progressive disclosure keeps orientation small and expands content when the agent needs it.

Install

Requires Bun 1.4.2 or newer.

Add this to .mcp.json or your MCP client configuration:

{
  "mcpServers": {
    "backlog": {
      "command": "bunx",
      "args": ["--bun", "backlog-mcp"]
    }
  }
}

For agent-assisted setup, ask your agent to read and follow the installation skill. The default command starts a persistent local HTTP server and bridges MCP stdio requests to it. See installation and server lifecycle.

Quick start

From your project directory, explicitly select its home (even without docs/ yet):

bunx --bun backlog-mcp --home project --project-root . wakeup
bunx --bun backlog-mcp --home project --project-root . create "Fix authentication flow" --content "Investigate the failing login."
bunx --bun backlog-mcp --home project --project-root . list --type task
bunx --bun backlog-mcp --home project --project-root . recall "authentication decisions"
bunx --bun backlog-mcp --home project --project-root . remember "Describe one verified durable login convention." --title "Login convention" --layer semantic

For MCP task work, discover and inspect the full schemas of backlog_create_work, backlog_start_task and backlog_complete_task. The Markdown body field is content; declarations may add required fields.

Wake up once at session start. Recall when prior knowledge matters, inspect its age and correction lineage, then expand selected entities with backlog_get (context: true includes relation stubs). Remember one durable fact per memory; correct an old fact with supersedes or state_key. See the memory protocol.

Viewer

Open localhost:3030 while the server is running. The Desk surfaces at most seven attention items with provenance and copy-ready instructions for your agent. The viewer provides search, entity detail, activity and live updates; edits flow through your agent or native Markdown files. See viewer usage.

Documentation

Local Markdown remains the source of truth. Legacy Workers/D1 hosting is retained and descoped; see the installation guide.

MIT licensed. npm package · Maintainer’s engineering story.