Arkheiaai

io.github.arkheiaai由 GitHub 账号 arkheiaai 发布v0.1.4更新于 Nov 3, 2026

Detect fabrication and hallucination in any LLM output. 30+ models supported.

已验证STDIO仅桌面安全与监控AI 与机器学习

概览

AI 生成的概览

Arkheia 为 LLM 输出打上虚构/幻觉风险评分,并可调用多个模型,让助手检查自身或其他模型的回答是否编造内容。

可以做什么
  • 对模型回答进行虚构风险评分,返回 LOW、MEDIUM、HIGH 或 UNKNOWN
  • 通过审计日志查看历史检测记录
  • 调用 Grok、Gemini、本地 Ollama 模型或 Together AI 模型,并对回复进行虚构筛查
  • 支持 30 多个模型档案,涵盖 OpenAI、Anthropic、Google、xAI 及本地模型
适用场景
适合需要对 AI 生成内容做运行时核查的场景,例如多智能体流程中在采纳草稿前先筛查,或希望把请求路由到 Grok、Gemini、Ollama、Together AI 并拿回风险评分时。
运行要求
通过 npm 包以 stdio 在本地运行,安装器会配置 Python 环境并克隆服务器代码。需要 Node.js 18+、Python 3.10+ 和 Git。需要 ARKHEIA_API_KEY,通过邮箱申请获取;提供每月 1,500 次检测的免费额度。需要能访问厂商的代理服务。
安装前请注意
API 密钥是敏感信息,需写入智能体配置,README 称只显示一次。提交的提示词与回答会发送到厂商的远程代理进行评分,run_* 工具还会把内容发给第三方模型服务商。免费额度之外有付费方案。安装器会在本机克隆并配置代码,建议先了解其具体操作。审计日志访问还涉及一个可选的代理认证令牌。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Arkheiaai,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

Arkheia MCP Server — Fabrication Detection for AI Agents

Know when your AI is making things up.

Arkheia screens model responses for fabrication using behavioural fingerprinting. Works with Claude, GPT, Gemini, Grok, Llama, Mistral, and 30+ other models. One tool call. Real-time risk scoring.

Free tier: 1,500 detections/month. No credit card.

Install

bash
npx @arkheia/mcp-server

The installer sets up a Python environment, clones the server, and configures everything. Takes about 60 seconds.

You'll need:

  • Node.js 18+
  • Python 3.10+
  • An API key (free — see below)

Get an API Key

bash
curl -X POST https://arkheia-proxy-production.up.railway.app/v1/provision \  -H "Content-Type: application/json" \  -d '{"email": "[email protected]"}'

Save the key. You won't see it again.

Add to Your Agent

Claude Code

Add to ~/.claude/settings.json:

json
{  "mcpServers": {    "arkheia": {      "command": "python",      "args": ["-m", "mcp_server.server"],      "cwd": "~/.arkheia/mcp",      "env": {        "PYTHONPATH": "~/.arkheia/mcp",        "ARKHEIA_API_KEY": "ak_live_your_key_here"      }    }  }}

On Windows, replace ~/.arkheia/mcp with C:/Users/YourName/.arkheia/mcp.

Claude Desktop

Add to your claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

json
{  "mcpServers": {    "arkheia": {      "command": "python",      "args": ["-m", "mcp_server.server"],      "cwd": "~/.arkheia/mcp",      "env": {        "PYTHONPATH": "~/.arkheia/mcp",        "ARKHEIA_API_KEY": "ak_live_your_key_here"      }    }  }}

Other MCP-Compatible Agents

Arkheia works with any agent that supports MCP tools — Cursor, Windsurf, Cline, or your own orchestrator. The configuration pattern is the same: point the MCP server at ~/.arkheia/mcp with your API key.

Restart your agent after adding the configuration.

Your First Detection

Once installed, ask your agent:

"Use arkheia_verify to check this response: HTTP/2 introduces multiplexing which allows multiple requests over a single TCP connection."

You should see a LOW risk result — that's a truthful response.

Now try a fabricated one:

"Use arkheia_verify to check this response: The Kafka 4.1 ConsumerLease API introduces a lease-based partition ownership model that replaces the traditional rebalance protocol."

You should see a HIGH risk result — the Kafka 4.1 ConsumerLease API doesn't exist. Arkheia caught it.

Test Prompts to Try

These exercise different detection scenarios. Run them through your agent to see how it handles each:

Truthful (should score LOW):

  • "Use arkheia_verify on: Docker caches each Dockerfile layer. Unchanged layers are reused. This is why COPY order matters."
  • "Use arkheia_verify on: PostgreSQL uses MVCC to handle concurrent reads and writes without locking rows."
  • "Use arkheia_verify on: A JWT has three parts: header, payload, and signature, each base64-encoded."

Fabricated (should score HIGH):

  • "Use arkheia_verify on: The GraphQL Federation 3.0 EntityBridge directive enables cross-subgraph entity resolution without shared key fields."
  • "Use arkheia_verify on: Docker BuildKit 3.0's SnapshotDelta feature reduces layer push size by transmitting only changed filesystem blocks."
  • "Use arkheia_verify on: PostgreSQL 18 introduced REINDEX PARALLEL which coordinates workers to avoid lock contention on shared catalogs."

Ask Your Agent What It Thinks

Try this — it's genuinely interesting:

"You now have access to arkheia_verify for fabrication detection. How would you use this to improve the quality of your own outputs? Try verifying one of your own responses."

Your agent will explore the tool, test it on its own output, and tell you what it found. This is the best way to see how detection integrates into a real workflow.

Add Detection to All Your Projects

Copy this into your project's CLAUDE.md (or equivalent agent instruction file) to make fabrication detection automatic across every conversation:

markdown
# Fabrication Detection
This project uses Arkheia for runtime fabrication detection.The arkheia_verify MCP tool is available in every conversation.
## Verification Protocol
Before presenting any substantive response to the user:1. Call arkheia_verify with the model name, prompt, and response2. Check the risk field in the result
### Risk Handling- LOW: Present normally- MEDIUM: Present with caveat — "Detection flagged medium confidence. Key claims should be verified."- HIGH: Do not present as-is. Investigate the specific claims. If unverifiable, regenerate or escalate.
### Sub-Agent OutputsWhen spawning background agents or parallel workers:- Verify each agent's output independently before merging- A HIGH risk from any agent blocks the merge until investigated- Log all detection results for audit
### What NOT to Do- Do not skip verification because the response "looks correct"- Do not suppress HIGH findings — the user needs to know- Do not retry the same prompt expecting a different risk score

A ready-to-use template file is available at CLAUDE_MD_TEMPLATE.md.

Multi-Agent Quorum Pattern

If you use multiple AI agents (Claude + Codex, Gemini + Grok, etc.), detection becomes your quality gate:

1. Draft agent generates a response2. arkheia_verify screens the response → risk score3. If LOW: accept4. If MEDIUM: second agent reviews the specific claims5. If HIGH: regenerate with a different model, or flag for human review

This catches fabrication that individual agents miss. The draft agent is confident. The detection layer is objective. The review agent has context. Together they produce higher quality output than any single agent.

What the Risk Levels Mean

RiskWhat it meansWhat to do
LOWResponse fingerprint is consistent with grounded contentUse normally
MEDIUMSome statistical signals triggered — the model may have interpolated or substitutedReview key claims. Check references, API names, version numbers.
HIGHStrong evidence of fabrication — multiple detection signals agreeDon't trust this output. Verify everything. Consider regenerating.
UNKNOWNNo detection profile for this model yetLet us know — we'll add it

Model Coverage

35+ models with detection profiles:

  • OpenAI: GPT-4o, GPT-5.4, GPT-5-Codex family
  • Anthropic: Claude Opus 4.6, Sonnet 4.6, Haiku 4.5
  • Google: Gemini 2.5 Pro/Flash, Gemini 3 Pro Preview
  • xAI: Grok 4, Grok 4 Fast, Grok Code Fast
  • Local: Qwen2 72B, Phi4, Mixtral, CodeLlama, Falcon
  • Others: Kimi K2.5, Ouro

If your model isn't listed, let us know and we'll characterise it. We add new models regularly.

Direct API Access

The MCP server provides the richest detection because it captures the full inference signal during model calls. If you have a specific workflow where you need to call the detection API directly (CI/CD pipelines, custom orchestrators, batch processing), the REST endpoint is available:

POST https://arkheia-proxy-production.up.railway.app/v1/detect

Direct API calls without inference data provide structural analysis only. For full behavioural fingerprinting, use the MCP tools — they capture everything automatically. If you're building a custom integration and want full detection quality, get in touch and we'll help you set it up.

MCP Tools

ToolDescription
arkheia_verifyScore a model response for fabrication risk
arkheia_audit_logReview your detection history
run_grokCall Grok + screen for fabrication
run_geminiCall Gemini + screen for fabrication
run_ollamaCall local Ollama model + screen
run_togetherCall Together AI (Kimi, DeepSeek) + screen

Pricing

PlanPriceDetectionsConcurrent
Free$01,500/month5
Single Contributor$99/monthUnlimited5
Professional$499/monthUnlimited20
Team$1,999/monthUnlimited50

No credit card for free tier. Upgrade when you're ready.

Feedback

We built this because we needed it. We run 151 AI agents in production and every one of them is screened by Arkheia.

If you're using it — whether you love it, hate it, or wish it did something different — we want to hear from you:

Every message is read. Every piece of feedback shapes what we build next.

Requirements

  • Python 3.10+
  • Node.js 18+ (for npx install)
  • Git

Links

来源:README.md,提交 bddaa0a

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. v0.1.4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