Aegis

io.github.getaegis由 GitHub 账号 getaegis 发布v1.0.5更新于 Oct 24, 2026

Credential isolation for AI agents. Inject secrets at the network boundary.

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

本地优先的凭据隔离代理,让 AI 助手通过 Aegis 调用外部 API,由它在网络边界注入已保存的密钥。

可以做什么
向已配置的服务发起带认证的 API 调用,由 Aegis 注入保存的凭据,助手本身看不到密钥(aegis_proxy_request)。列出可用服务名称,绝不返回密钥内容(aegis_list_services)。查询 Aegis 的运行状态与健康状况(aegis_health)。请求与 HTTP 网关走同一套流程:域名白名单、代理身份认证、请求体检查、限流和审计日志。
适用场景
当助手需要调用 Slack、GitHub、Stripe 等服务,但你不希望把原始 API 密钥写进 MCP 配置或让模型看到,并希望对出站请求做域名限制和审计时使用。
运行要求
以 npm 包 @getaegis/cli 在本地通过 stdio 运行,需要 Node.js 20 或更高版本。需先执行 aegis init 初始化,并提供密钥 AEGIS_MASTER_KEY 和 AEGIS_SALT;凭据须先加入 vault。可选配置包括 AEGIS_DATA_DIR、AEGIS_POLICIES_DIR、AEGIS_POLICY_MODE、AEGIS_PORT、AEGIS_REQUIRE_AGENT_AUTH、AEGIS_VAULT、AEGIS_METRICS 及日志选项。需要访问目标 API 的网络。
安装前请注意
保险库的安全性取决于 AEGIS_MASTER_KEY 和 AEGIS_SALT;丢失可能导致已存凭据无法恢复,泄露则所有密钥暴露。代理会发起真实的外部调用,可能在第三方服务中修改或发送数据,并可能产生费用(例如 Stripe)。代理身份认证默认关闭(AEGIS_REQUIRE_AGENT_AUTH 为 false),建议开启并按代理分别授权。审计日志和指标可能记录请求详情。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Aegis,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

[CI] [npm version] [Docker] [License]

Stop putting API keys where AI agents can read them.

Aegis is a local-first credential isolation proxy for AI agents. It sits between your agent and the APIs it calls — injecting secrets at the network boundary so the agent never sees, stores, or transmits real credentials.

[Aegis demo]

How It Works

[How Aegis works — agent sends request through Gate, credentials injected at the network boundary]

Why?

AI agents (Claude, GPT, Cursor, custom bots) increasingly call real APIs — Slack, GitHub, Stripe, databases. The current pattern is dangerous:

  1. Agents see raw API keys — one prompt injection exfiltrates them
  2. No domain guard — a compromised agent can send your Slack token to evil.com
  3. No audit trail — you can't see what an agent did with your credentials
  4. No access control — every agent can use every credential

Aegis solves all four. Your agent makes HTTP calls through a local proxy. Aegis handles authentication, enforces domain restrictions, and logs everything.

Quick Start

bash
# Installnpm install -g @getaegis/cli
# Initialize (stores master key in OS keychain by default)aegis init
# Add a credentialaegis vault add \  --name slack-bot \  --service slack \  --secret "xoxb-your-token-here" \  --domains slack.com
# Start the proxyaegis gate --no-agent-auth
# Test it — Aegis injects the token, forwards to Slack, logs the request# X-Target-Host tells Gate which upstream server to forward to (optional if credential has one domain)curl http://localhost:3100/slack/api/auth.test \  -H "X-Target-Host: slack.com"

Production Setup (with agent auth)

bash
# Create an agent identityaegis agent add --name "my-agent"# Save the printed token — it's shown once only
# Grant it access to specific credentialsaegis agent grant --agent "my-agent" --credential "slack-bot"
# Start Gate (agent auth is on by default)aegis gate
# Agent must include its tokencurl http://localhost:3100/slack/api/auth.test \  -H "X-Target-Host: slack.com" \  -H "X-Aegis-Agent: aegis_a1b2c3d4..."

MCP Integration

Aegis is a first-class MCP server. Any MCP-compatible AI agent can use it natively — no HTTP calls needed.

Before (plaintext key in config):

json
{  "mcpServers": {    "slack": {      "command": "node",      "args": ["slack-mcp-server"],      "env": { "SLACK_TOKEN": "xoxb-1234-real-token-here" }    }  }}

After (Aegis — no key visible):

json
{  "mcpServers": {    "aegis": {      "command": "npx",      "args": ["-y", "@getaegis/cli", "mcp", "serve"]    }  }}

Generate the config for your AI host:

bash
aegis mcp config claude   # Claude Desktopaegis mcp config cursor   # Cursoraegis mcp config vscode   # VS Codeaegis mcp config cline    # Clineaegis mcp config windsurf # Windsurf

The MCP server exposes three tools:

ToolDescription
aegis_proxy_requestMake an authenticated API call (provide service + path, Aegis injects credentials)
aegis_list_servicesList available services (names only, never secrets)
aegis_healthCheck Aegis status

The MCP server replicates the full Gate security pipeline: domain guard, agent auth, body inspection, rate limiting, audit logging.

Setup Guides

Features

FeatureDescription
Encrypted VaultAES-256-GCM encrypted credential storage with PBKDF2 key derivation
HTTP Proxy (Gate)Transparent credential injection — agent hits localhost:3100/{service}/path
Domain GuardEvery outbound request checked against credential allowlists. No bypass
Audit LedgerEvery request (allowed and blocked) logged with full context
Agent IdentityPer-agent tokens, credential scoping, and rate limits
Policy EngineDeclarative YAML policies — method, path, rate-limit, time-of-day restrictions
Body InspectorOutbound request bodies scanned for credential-like patterns
MCP ServerNative Model Context Protocol for Claude, Cursor, VS Code, Windsurf, Cline
Web DashboardReal-time monitoring UI with WebSocket live feed
Prometheus Metrics/_aegis/metrics endpoint for Grafana dashboards
Webhook AlertsHMAC-signed notifications for blocked requests, expiring credentials
RBACAdmin, operator, viewer roles with 16 granular permissions
Multi-VaultSeparate vaults for dev/staging/prod with isolated encryption keys
Shamir's Secret SharingM-of-N key splitting for team master key management
Cross-Platform Key StorageOS keychain by default (macOS, Windows, Linux) with file fallback
TLS SupportOptional HTTPS on Gate with cert/key configuration
Configuration Fileaegis.config.yaml with env var overrides and CLI flag overrides

Example Integrations

Step-by-step guides with config files and policies included:

  • Slack Bot — Protect your Slack bot token with domain-restricted proxy access
  • GitHub Integration — Secure GitHub PAT with per-agent grants and read-only policies
  • Stripe Backend — Isolate Stripe API keys with body inspection and rate limiting
  • OpenClaw Skill — Aegis skill for OpenClaw personal AI assistant

Security

  • Published STRIDE threat model — 28 threats analysed, 0 critical/high unmitigated findings
  • Full security architecture documentation (trust boundaries, crypto pipeline, data flow)
  • AES-256-GCM + ChaCha20-Poly1305 encryption at rest
  • Domain guard enforced on every request — no bypass
  • Agent tokens stored as SHA-256 hashes — cannot be recovered, only regenerated
  • Request body inspection for credential pattern detection
  • Open source (Apache 2.0) — read the code

How Aegis Compares

.env filesVault/DopplerInfisicalAegis
Agent sees raw keyYesYes (after fetch)Yes (after fetch)No — never
Domain restrictionsNoNoNoYes
MCP-nativeNoNoAddingYes
Local-firstYesNoNoYes
Setup10 sec30+ min15+ min~2 min

See full comparison for detailed breakdowns against each approach.

Documentation

DocumentDescription
Usage GuideFull reference: CLI commands, configuration, RBAC, policies, webhooks, troubleshooting
Security ArchitectureTrust boundaries, crypto pipeline, data flow diagrams
Threat ModelSTRIDE analysis — 28 threats, mitigations, residual risks
ComparisonDetailed comparison with .env, Vault, Doppler, Infisical
FAQCommon questions and objections
RoadmapFeature roadmap
ContributingCode style, PR process, architecture overview

Install

bash
# npmnpm install -g @getaegis/cli
# Homebrewbrew tap getaegis/aegis && brew install aegis
# Dockerdocker run ghcr.io/getaegis/aegis --help

Requires Node.js ≥ 20 — check with node -v

Development

bash
git clone https://github.com/getaegis/aegis.gitcd aegisyarn installyarn buildyarn test

See CONTRIBUTING.md for code style, PR process, and architecture overview.

License

Apache 2.0

来源:README.md,提交 f1321b3

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. v1.0.5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