Aegis Mcp

io.github.cleburn由 GitHub 账号 cleburn 发布v0.1.15更新于 Nov 10, 2026

Runtime governance enforcement for AI agents. Zero token overhead.

已验证STDIO仅桌面开发者工具安全与监控

概览

AI 生成的概览

一个本地 MCP 服务器,在 AI 代理执行文件写入、删除和命令执行之前,按项目治理策略进行校验和拦截。

可以做什么
Aegis MCP 将项目的 .agentpolicy/ 治理文件读入自身进程内存,并在运行时校验代理操作。工具包括用于查看角色边界的 aegis_policy_summary、aegis_select_role、aegis_check_permissions,受治理的 aegis_write_file、aegis_read_file、aegis_delete_file 和 aegis_execute,以及用于质量门禁的 aegis_complete_task 和用于人工确认例外的 aegis_request_override。被拦截的操作会返回原因,并在允许时给出覆盖令牌;执行决策会记录到仅追加的 overrides.jsonl。
适用场景
适用于已有 Aegis 治理文件、希望代理的文件与命令操作接受路径权限、角色边界、敏感内容模式和质 量门禁检查的项目。也适合需要记录被拦截操作和人工确认覆盖的团队。
运行要求
以 npm 包 aegis-mcp-server 通过 stdio 在本地运行(需全局安装,因此需要 Node.js)。项目需包含 .agentpolicy/ 目录和 .mcp.json 连接配置,可由 aegis-cli 生成。未声明认证、环境变量或请求头。
安装前请注意
它管控写入、删除和执行操作,策略配置错误可能拦截正常操作,或在 construction 模式下授予整个仓库的写权限。覆盖需要人工确认并会被记录;不可变策略无法覆盖,只能通过 aegis init 修改。它会读取项目策略文件并写入仅追加的 overrides.jsonl 日志。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Aegis Mcp,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。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

aegis-mcp-server

[npm version] [License: MIT] [MCP Registry]

MCP enforcement layer for the Aegis agent governance specification.

Policy at the root. Enforcement at runtime. Accountability on every action.

What It Does

[Aegis Lifecycle Overview]

aegis-mcp-server is an MCP server that validates every agent action against your .agentpolicy/ files before it happens. Path permissions, content scanning, role boundaries, quality gates — all enforced at runtime with zero token overhead to the agent.

The agent never loads your governance files. The MCP server reads them into its own process memory and validates silently. The agent calls governed tools and gets back either a success or a blocked response with the specific reason.

Quick Start

bash
# Install globallynpm install -g aegis-mcp-server

If you generated your policy with aegis-cli, the .mcp.json connection config is already in your project root. Just install the MCP and open your agent — it connects automatically.

First Prompt

When starting a new agent session in a governed project, use this as your first prompt:

Call aegis_policy_summary now. This is your governance contract — it defines yourrole, your boundaries, and which tools to use. Do not take any action until you havecalled this tool and received confirmation from the user to proceed.

For initial builds, the Aegis CLI generates a custom handoff prompt tailored to your project — use that instead.

How It Works

Universal Mode (Default)

The MCP starts without a pre-assigned role. When the agent calls aegis_policy_summary, it receives the list of available roles — including the built-in construction role and all specialist roles from .agentpolicy/roles/. The agent presents them to the user, the user picks, and the agent calls aegis_select_role to lock in. All enforcement uses the selected role for the rest of the session.

This is the default behavior — no configuration needed beyond the .mcp.json that aegis init creates automatically.

Construction Mode

The construction role is always available for initial builds and major restructuring. When selected:

  • The agent has full repository access (all paths writable and readable)
  • The .agentpolicy/ files serve as the blueprint — the agent reads constitution, governance, and role files to understand the project's architecture, conventions, and quality standards
  • File operations run through native tools rather than governed tools, for speed
  • The MCP logs the construction session start to state/overrides.jsonl with a timestamp and human_confirmed: true
  • When the build is complete, the agent calls aegis_complete_task to run quality gates and close construction mode — the closing timestamp is logged alongside the opening entry
  • All future sessions after construction should select a specialist role for governed operations

Construction mode is not a bypass — the agent still follows the governance files as its blueprint. It's a speed optimization for greenfield builds where enforcing write restrictions on every file would be counterproductive.

Fixed Mode

If you know which role to assign at startup:

json
{  "mcpServers": {    "aegis": {      "command": "aegis-mcp",      "args": ["--project", ".", "--role", "backend"]    }  }}

The MCP locks to that role immediately. aegis_policy_summary returns the role's boundaries directly, skipping role selection.

Tools

ToolWhat it doesToken cost
aegis_policy_summaryRole boundaries and governance summary (or available roles in universal mode)~200 tokens
aegis_select_roleSelect a role (including construction) in universal modeTiny
aegis_check_permissionsPre-check if an operation is allowedTiny
aegis_write_fileGoverned write with path + content validationSame as a normal write
aegis_read_fileGoverned read with path validationSame as a normal read
aegis_delete_fileGoverned delete with path validationTiny
aegis_executeGoverned command executionCommand output only
aegis_complete_taskRun quality gates and close construction mode if activeGate results only
aegis_request_overrideExecute a blocked action after human confirmationTiny

Zero Token Overhead

Traditional approach: load governance files into the agent's context window. Token cost scales with policy complexity.

Aegis MCP approach: the server loads policy into its own process memory. The agent calls tools and gets structured results. A project with 200 lines of governance has the same token cost as one with 20 lines. The complexity is absorbed by the server, not the agent.

Enforcement

[Aegis Runtime Enforcement]

  • Governance boundaries — writable, read_only, forbidden path lists
  • Role scoping — agents confined to their role's writable and readable paths
  • Sensitive pattern detection — content scanned against governance-defined regex patterns (content only, not path-based)
  • Cross-domain boundaries — imports validated against shared interface rules
  • Quality gate validation — pre_commit flags mapped to build_commands and executed
  • Override logging — every blocked action logged to append-only overrides.jsonl
  • Immutable policies — designated rules that cannot be overridden, ever
  • Construction session logging — start and end timestamps for initial builds

Override Protocol

[Aegis Completion]

When an action is blocked and the governance override behavior is warn_confirm_and_log:

  1. The blocked response includes an override_token and the specific policy violated
  2. The agent presents the violation to the user
  3. If the user confirms, the agent calls aegis_request_override with the token and the user's rationale
  4. The action proceeds — the override is logged with human_confirmed: true
  5. Normal governance resumes immediately — the override is a one-time exception

Immutable policies (e.g., HIPAA compliance, ITAR data sovereignty, audit completeness) return override_available: false and cannot be overridden. The user must modify the governance through aegis init.

Consent-Based Governance

The Aegis MCP does not override the agent's native directives. It introduces itself through tool descriptions, explains why governance is active, and asks the agent to seek user permission to route write operations through Aegis tools. The user's authority is the enforcement mechanism.

Native tools for reading, searching, and exploring the codebase work fine without governance gating. Only write, delete, and execute operations benefit from routing through Aegis.

Architecture

Agent ──→ aegis-mcp-server ──→ File System              │              ├── Loads .agentpolicy/ into process memory (once)              ├── Watches for policy changes (auto-reload)              ├── Validates every tool call against policy              ├── Returns success or blocked with override option              └── Logs all enforcement decisions to overrides.jsonl

Three artifacts, one governance framework:

  • aegis-spec — The governance standard
  • aegis-cli — Generates the governance
  • aegis-mcp-server — Enforces the governance

License

MIT

来源:README.md,提交 1741ab2

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.1.15最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分