local-knowledge-suite
Provides MCP tools for workspace initialization, health checks, context recall, and durable memory capture, enabling project-specific memory and shared company knowledge with citations.
README
Local Knowledge Suite
Tiếng Việt · Documentation · Security · Support
Local Knowledge Suite is a free, open-source Codex plugin for Windows. It gives each project a private memory vault and can optionally connect many projects to one citation-backed company knowledge base.
What it does
- Skills tell Codex how to handle memory and company knowledge safely.
- MCP tools let Codex initialize a workspace, check its health, recall context, and save durable project memory without repeated copy and paste.
- Project memory keeps preferences and decisions inside one project.
- Company knowledge lets several projects reuse verified facts with source excerpts and citations.
The plugin runs locally. It includes no telemetry or analytics. Windows is the supported release platform for v1.
Five-minute start: memory only
Requirements: Windows 10/11, PowerShell 5.1+, Git, Codex, and
uv.
Install the tagged marketplace directly from GitHub:
codex plugin marketplace add luutru433-coder/local-knowledge-suite --ref v1.0.0
codex plugin add local-knowledge-suite@john-idk
Restart Codex and open a new task. In your project, ask:
Use $manage-local-knowledge to initialize this project as memory-only,
then check its Local Knowledge status.
No company repository is needed. The v1 default profile is memory-only.
Normal users can keep asking Codex and do not need a plugin path. To run the CLI yourself, resolve and verify the installed path once:
$Plugin = (codex plugin list --json | ConvertFrom-Json).installed |
Where-Object { $_.pluginId -eq "local-knowledge-suite@john-idk" } |
Select-Object -First 1
if (-not $Plugin) { throw "Local Knowledge Suite is not installed." }
$PluginRoot = [IO.Path]::GetFullPath([string]$Plugin.source.path)
$Launcher = Join-Path $PluginRoot "scripts\Invoke-LocalKnowledge.ps1"
if (-not (Test-Path -LiteralPath $Launcher)) {
throw "Installed Local Knowledge Suite launcher was not found."
}
$ProjectRoot = (Resolve-Path ".").Path
& $Launcher `
-Tool local-knowledge `
-ToolArguments @(
"project", "init", "--root", $ProjectRoot,
"--profile", "memory-only", "--install-mode", "global"
)
Always use an absolute project path. Run project doctor after setup and open a
new Codex task so the installed skills and MCP server are reloaded.
In the user's application repository (not this product source repository), review and commit the stable project identity plus managed ignore rules:
git add .local-knowledge/project.toml .gitignore
git diff --cached -- .local-knowledge/project.toml .gitignore
git commit -m "chore: initialize local knowledge workspace"
Without the committed project manifest, a clone gets a different project ID and cannot import the original project's Memory snapshot.
Add company knowledge
Create a separate repository for a fictional example company:
$CompanyRoot = "D:\Knowledge\Northstar-Bicycle-Labs"
$AdminUser = "admin"
& "$PluginRoot\scripts\Invoke-LocalKnowledge.ps1" `
-Tool local-knowledge `
-ToolArguments @(
"company", "init", "--root", $CompanyRoot,
"--name", "Northstar Bicycle Labs", "--admin", $AdminUser
)
The command returns a stable company_id. Initialize a project with profile
combined to use both its own memory and the shared company evidence. V1 does
not rebind an existing Memory-only project; choose combined at its first
bootstrap. Put
approved originals under the ignored COMPANY_ROOT\.local-sources\ directory;
commit only the company manifest and canonical knowledge/.
Four MCP tools
| Tool | Simple meaning |
|---|---|
bootstrap_workspace |
Prepare one project safely and repeatedly. |
knowledge_status |
Check compatibility, stores, warnings, and pending work. |
recall_context |
Search skills, project memory, and verified company evidence separately. |
capture_memory |
Save one durable project-specific fact after redaction and distillation. |
knowledge_status and recall_context are non-destructive and idempotent, but
they are not declared read-only. Status may acquire transient local locks;
recall may distill already-captured pending Memory events and rebuild or cache
derived indexes.
Company answers must have evidence_sufficient: true and verified citations.
Memory and skill text never replace company evidence.
Memory snapshots are opt-in plain text for a user's own project repository.
This product repository ignores .local-knowledge/memory-snapshot/ and never
ships real user Memory.
Documentation
- English user guide
- Vietnamese user guide
- Windows migration checklist
- Architecture
- Platform roadmap
- Distribution and commercial use
- Release checklist
The guides cover global and vendored installation, company ingestion, multi-project reuse, Memory export/import, privacy, security, troubleshooting, uninstalling, MCP/CLI reference, and development with pnpm.
Development
Node.js is not required to use a released plugin. Maintainers use Node 22 and the repository-pinned pnpm 11.9.0 to build the portal and VitePress docs:
pnpm install --frozen-lockfile
pnpm --filter @local-knowledge-suite/docs build
pnpm --filter @local-knowledge-suite/portal build
uv sync --project .\plugins\local-knowledge-suite\runtime `
--frozen --extra build --extra dev --python 3.12
uv run --project .\plugins\local-knowledge-suite\runtime `
--frozen --extra build --extra dev --python 3.12 python -m pytest tests
pnpm is used because one lockfile can reproduce both JavaScript workspaces, its content-addressed store avoids duplicate downloads, and frozen installs fail when a manifest and lockfile disagree.
License
Copyright © 2026 John-IDK. Released under the MIT License. Third-party components keep their own licenses; see THIRD_PARTY_NOTICES.md.
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。