Stele

dev.stele-aiv1.0.0更新于 Sep 30, 2026

Shared project memory for AI coding agents: decisions, lessons, risks and tasks in one graph.

已验证Streamable HTTP可网页运行Knowledge & MemoryProductivity & Workflow

概览

AI 生成的概览

为 AI 编程代理提供共享的项目知识图谱，用于回溯决策、记录经验与风险，并跨会话协调任务。

功能
Stele 是面向 AI 代理及其协作者的托管式共享项目记忆。通过该连接器，代理可以回溯项目已知内容、追踪某项决策的原因及其替代对象、随时记录决策、经验、风险与缺口，并创建、认领和完成任务，避免两个代理重复劳动。它还支持把工作干净地交接给下一个代理、会话或同事。工具分为只读、写入和单一 destructive 层级，后者在执行前始终会提示确认。
适用场景
当多个代理或人员长期协作同一项目、且上下文在会话之间不断丢失时，值得添加。它适合希望自动浮现既往决策、风险和未完成任务，而不必每次重新说明的团队。
运行要求
需要一个 Stele 账号和至少一个项目；通过连接器登录时若没有账号会自动创建，连接不需要付费方案。它是托管在远端的 MCP 服务器，通过 Streamable HTTP 访问提供方端点，使用 OAuth 2.1 与动态客户端注册；客户端首次连接时自行注册并打开 Stele 登录页。插件仓库中不存储任何凭据或密钥。另有可选的命令行工具用于本地仓库，但并非必需。
安装前请注意
该连接器会写入共享的项目图谱：记录决策、经验与风险，以及创建、认领或完成任务，都会改变其他代理和同事看到的数据。单一 destructive 工具涵盖取消、强制认领、解除关联、停用和合并等重大操作，且执行前始终提示确认。访问范围限于你拥有或所属的项目，项目内容存储在托管服务而非本地。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Stele，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "stele": {
      "type": "http",
      "url": "https://app.stele-ai.dev/api/mcp"
    }
  }
}

README

Stele — MCP connector plugin

Stele is shared project memory and work tracking for AI agents and the people who work with them. It keeps decisions, lessons, risks, documents, components, and tasks in one durable knowledge graph, so work can continue across agent clients without losing the project's history.

This repository is the plugin that connects an agent client to Stele. It follows the Agent Plugins open standard and contains configuration and documentation only.

What's in here

plugin.json          Agent Plugins manifestmcp.json             Stele's hosted MCP server, over Streamable HTTPskills/stele/        Guidance on when and how an agent should use Stele

There is no build step, no bundled binary, and no executable code. Stele itself runs as a hosted service; this plugin only points at it.

What the connector does

Once connected, an agent can:

  • Recall what a project already knows before it acts, instead of starting cold.
  • Trace why a decision was made, what caused it, and what it replaced.
  • Capture decisions, lessons, risks, and gaps as they happen, so they survive the conversation that produced them.
  • Coordinate tasks — create, claim, and complete them — so two agents don't duplicate each other's work.
  • Hand off cleanly to the next agent, session, or teammate.

The bundled skill tells the agent when reaching for these tools is appropriate and when it isn't, which matters on a surface where the agent chooses for itself.

Working locally

This connector is the whole product on a chat or IDE surface. If you also work in a local repository, Stele's command-line tool adds automatic recall — relevant prior decisions, risks, and open work surface as you go, rather than only when someone thinks to ask. See the documentation.

The tool is optional. Nothing in this plugin requires it, downloads it, or runs it for you.

Requirements

A Stele account and at least one project. Signing in through the connector creates the account if you don't already have one, and no paid plan is required to connect.

Authentication

OAuth 2.1 with dynamic client registration (RFC 7591). Your client registers itself on first connection and opens Stele's sign-in page. Access is limited to the projects you own or belong to.

No credentials, tokens, or secrets are stored in this repository, and none are needed to install it.

Tool safety

Stele's tools are split so a client can tell safe operations from unsafe ones. Read-only tools are annotated as such and never modify data. Write tools are separate. Every consequential operation — cancelling, force-claiming, unlinking, retiring, merging — lives in a single destructive tool that always prompts before acting. No tool mixes these tiers.

Links

License

MIT — see LICENSE.

来源：README.md，提交 7d6a9f4

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v1.0.0最新Sep 30, 2026