# TrustVector > Independent, evidence-based trust evaluations for 100+ AI models, agents, and tools. - **Type:** MCP server - **Install:** `agentstack add mcp-guard0-ai-trustvector` - **Verified:** Yes β€” security-reviewed for prompt injection and unsafe behavior - **Seller:** [guard0-ai](https://agentstack.voostack.com/s/guard0-ai) - **Installs:** 0 - **Category:** [Integrations](https://agentstack.voostack.com/c/integrations) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [guard0-ai](https://github.com/guard0-ai) - **Source:** https://github.com/guard0-ai/TrustVector - **Website:** https://trustvector.dev ## Install ```sh agentstack add mcp-guard0-ai-trustvector ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # πŸ›‘οΈ TrustVector ### Trust scores for the entire AI stack β€” models, agents, and MCP servers. **Benchmarks tell you how smart an AI is. TrustVector tells you whether you can trust it in production.** [](https://opensource.org/licenses/MIT) [](#-current-coverage) [](/data/models) [](/data/agents) [](/data/mcps) [](http://makeapullrequest.com) [](https://github.com/guard0-ai/TrustVector) 🌐 **[trustvector.dev](https://trustvector.dev)** Β· πŸ“– **[Methodology](/docs/METHODOLOGY.md)** Β· 🀝 **[Contribute](/CONTRIBUTING.md)** Β· πŸ—ΊοΈ **[Roadmap](/ROADMAP.md)** --- ## 🚨 What our data found (June 2026) This isn't a list of logos. Every entry is an evidence-linked evaluation across **security, privacy, performance, transparency, and operations** β€” and the findings are uncomfortable: - ⚠️ **21 of 60 models** in the registry are **retired, deprecated, superseded, or were never released** β€” including models still hardcoded in thousands of production apps (Grok 3 now *silently redirects* to a different model; Gemini 2.0 Flash was shut down June 1). - 🩸 **Archived MCP reference servers ship unpatched SQL injection.** The Postgres reference server was still pulling ~21k weekly downloads *after* being archived with a known SQLi β€” we score it 51/100 on security so you don't find out the hard way. - πŸ•³οΈ **Popularity β‰  safety.** Context7 (57kβ˜…, the most-starred MCP server on GitHub) scores 86/100 on performance but **59/100 on security** after the "ContextCrush" registry-poisoning vulnerability. Playwright MCP: 88 performance, 60 security. - πŸ”“ **The agent you let browse the web matters.** General-purpose autonomous agents score as low as **50/100 on privacy** in our registry; sandboxed, permission-gated coding agents score 20+ points higher. - ⏳ **The OpenAI Assistants API sunsets August 26, 2026.** If you're on it, your migration window is measured in weeks. It's flagged. **Every one of these claims links to a primary source with a date.** That's the whole point. --- ## πŸ“Š Frontier models, scored on what benchmarks ignore Overall = mean of 5 dimension scores. Full criteria, evidence URLs, and confidence levels in each JSON file. | Model | Overall | Perf | Security | Privacy | Transparency | Ops | |---|:---:|:---:|:---:|:---:|:---:|:---:| | **Claude Fable 5** (Anthropic) | **92** | 98 | 92 | 93 | 88 | 91 | | **Claude Opus 4.8** (Anthropic) | **92** | 96 | 92 | 93 | 88 | 91 | | **GPT-5.5** (OpenAI) | **91** | 97 | 89 | 87 | 90 | 94 | | **Gemini 3.1 Pro** (Google) | **91** | 96 | 88 | 88 | 88 | 93 | | **Mistral Large 3** (Mistral, open) | **85** | 88 | 83 | 87 | 80 | 86 | | **Grok 4.3** (xAI) | **83** | 94 | 83 | 76 | 82 | 82 | | **DeepSeek-V4** (open) | **83** | 92 | 83 | 78 | 80 | 83 | | **GLM-5** (Z.ai, open) | **82** | 92 | 80 | 75 | 81 | 83 | | **Kimi K2.6** (Moonshot, open) | **81** | 91 | 79 | 75 | 80 | 82 | Notice the spread: models within 5 points of each other on *capability* differ by **15+ points on privacy and security**. If you're choosing a model for healthcare, legal, or finance, the right-hand columns are the ones that get you fired. And it's not just models β€” the same lens on **coding agents** (Claude Code 80, OpenAI Codex 82, Devin 71, Manus 64) and **MCP servers** (GitHub 82, Playwright 80, Context7 79, archived Postgres 72) exposes exactly where the trust gaps are. --- ## 🎯 Why this exists Leaderboards answer *"which model is smartest?"* Nobody was answering: - Can this model touch **PHI under HIPAA**? What's its actual data-retention policy β€” with a link? - Is this MCP server **maintained**, or was it quietly archived with an open CVE? - Does this agent framework **sandbox tool execution**, or does prompt injection mean shell access? - Is this API **deprecated**, and what's the shutdown date? TrustVector evaluates every entity across **5 dimensions** β€” like a CVSS score for AI systems: | Dimension | What it covers | |---|---| | ⚑ **Performance & Reliability** | Benchmarks, latency, uptime, context limits | | πŸ”’ **Security** | Prompt-injection resistance, jailbreaks, sandboxing, CVE history | | πŸ” **Privacy & Compliance** | Data residency, retention, training opt-out, HIPAA/GDPR/SOC 2 | | πŸ” **Trust & Transparency** | Hallucination rate, explainability, model cards, open source | | πŸ› οΈ **Operational Excellence** | API/SDK quality, versioning policy, ecosystem, support | Three rules make it trustworthy: 1. **Every score has evidence** β€” a primary source URL, a date, and a methodology. 2. **Every score has a confidence level** β€” high / medium / low. We tell you when we're not sure. 3. **Everything is a JSON file in git** β€” disagree with a score? Open a PR with better evidence. That's the protocol. --- ## ⚑ 30-second start ```bash git clone https://github.com/guard0-ai/TrustVector.git cd TrustVector && npm install && npm run dev # β†’ http://localhost:3000 ``` Or skip the website β€” the data is just JSON: ```typescript import fable5 from './data/models/claude-fable-5.json'; fable5.trust_vector.security.overall_score; // 92 fable5.trust_vector.privacy_compliance.criteria.data_retention; // evidence, URL, date, confidence fable5.use_case_ratings['healthcare']; // { overall, notes, alternatives } ``` ### Weight it like CVSS β€” your risk profile, your score ```typescript import { calculateCustomScore, WEIGHTING_PROFILES } from '@/framework/calculator/custom-score'; calculateCustomScore(entity, WEIGHTING_PROFILES.healthcare); // HIPAA-weighted calculateCustomScore(entity, WEIGHTING_PROFILES.security_first); // or roll your own calculateCustomScore(entity, { performance_reliability: 0.20, security: 0.30, privacy_compliance: 0.25, trust_transparency: 0.15, operational_excellence: 0.10, }); ``` Predefined profiles: `balanced` Β· `security_first` Β· `performance_focused` Β· `enterprise` Β· `healthcare` Β· `financial` Β· `startup` --- ## πŸ“¦ Current Coverage **156 evaluations** across 3 categories (last refreshed June 2026 β€” yes, including the models that launched *this month*): 🧠 AI Models (60) β€” Claude Fable 5 β†’ archived also-rans, all scored **Frontier:** Claude Fable 5, Opus 4.8/4.7/4.6/4.5, Sonnet 4.6/4.5, Haiku 4.5 Β· GPT-5.5, GPT-5.4, GPT-5.3-Codex, GPT-5.2, GPT-5.1, GPT-5, o-series Β· Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3 Β· Grok 4.3, Grok 4.1 Β· Nova 2 Lite, Nova Pro **Open-weight:** DeepSeek V4 / V3.2 / R1 Β· Qwen3.5 Β· Kimi K2.6 Β· GLM-5 Β· MiniMax-M2 Β· Mistral Large 3 Β· Command A+ Β· Gemma 4 / 3 Β· gpt-oss-120b/20b Β· Llama 4 / 3.x Β· Nemotron **[Browse all models β†’](/data/models)** πŸ€– AI Agents (50) β€” coding agents, frameworks, enterprise platforms **Coding & autonomous:** Claude Code, Claude Agent SDK, OpenAI Codex, Devin, Cursor, GitHub Copilot coding agent, Google Jules, Gemini CLI, Manus **Frameworks:** OpenAI Agents SDK, Google ADK, Microsoft Agent Framework, AWS Strands, LangGraph, CrewAI, LlamaIndex, Pydantic AI, smolagents, Mastra, Dify **Enterprise:** Amazon Bedrock Agents, Azure Bot Service, Gemini Enterprise Agent Platform, IBM watsonx Assistant, Dialogflow, Lex, and more β€” plus deprecated/archived projects (Swarm, AgentGPT, BabyAGI…) clearly flagged **[Browse all agents β†’](/data/agents)** πŸ”Œ MCP Servers (46) β€” incl. security advisories on archived servers **Top ecosystem:** Context7, Chrome DevTools MCP, Playwright MCP, Serena **Official vendor:** GitHub, Figma, Stripe, Notion, Vercel, Hugging Face, Zapier, Apify, Firecrawl, shadcn **Reference:** the 7 actively maintained servers (fetch, git, filesystem, memory, time, sequential-thinking, everything) β€” plus the **archived** ones (Puppeteer, Postgres, SQLite, Slack, …) flagged with security advisories so you don't `npx` your way into a CVE **[Browse all MCPs β†’](/data/mcps)** --- ## 🀝 Contribute an evaluation (it's just a PR) The registry stays honest because anyone can challenge it. Add or update an evaluation: ```bash # 1. Start from an existing evaluation as your template cp data/models/claude-fable-5.json data/models/your-model-name.json # 2. Fill in scores β€” every score needs evidence (source, URL, date) + confidence level # 3. Validate against the schema npm run validate # 4. Open a PR git checkout -b evaluation/your-model-name && git add data/ && git commit -m "Add evaluation for X" ``` We review within 48 hours. Found a score you disagree with? **Bring a better source and open a PR** β€” that's how the system is supposed to work. See [CONTRIBUTING.md](/CONTRIBUTING.md). ### Scoring scale | Range | Meaning | |---|---| | 90–100 | Exceptional β€” industry leading | | 75–89 | Strong β€” meets enterprise requirements | | 60–74 | Adequate β€” usable with caveats | | 40–59 | Concerning β€” significant gaps | | 0–39 | Poor β€” not recommended | Full scoring rules, confidence definitions, and evidence requirements: [METHODOLOGY.md](/docs/METHODOLOGY.md) --- ## πŸ†š How it compares | | **TrustVector** | Leaderboards | Vendor model cards | |---|:---:|:---:|:---:| | Security & privacy scored | βœ… | ❌ | ⚠️ self-reported | | Evidence URL on every score | βœ… | ⚠️ | ❌ | | Confidence levels | βœ… | ❌ | ❌ | | Covers agents & MCP servers | βœ… | ❌ | ❌ | | Flags deprecated/archived/CVE'd entries | βœ… | ❌ | ❌ | | Custom CVSS-style weighting | βœ… | ❌ | ❌ | | Disagreement protocol | PR with sources | ❌ | ❌ | | License | MIT, all data in git | varies | proprietary | --- ## πŸ—οΈ Project structure ``` trustvector/ β”œβ”€β”€ data/ # The registry β€” one JSON file per evaluation β”‚ β”œβ”€β”€ models/ # 60 model evaluations β”‚ β”œβ”€β”€ agents/ # 50 agent evaluations β”‚ β”œβ”€β”€ mcps/ # 46 MCP server evaluations β”‚ └── use-cases/ # Use-case taxonomy (healthcare, finance, …) β”œβ”€β”€ framework/ # Schema, Zod validation, custom-score calculator β”œβ”€β”€ app/ # Next.js site (static export, zero tracking) └── scripts/ # CI validation β€” every PR is schema-checked ``` **TrustVector itself collects nothing:** no cookies, no tracking, static site generation, every evaluation version-controlled and validated in CI. --- ## ⭐ Star history If TrustVector saved you from a deprecated API, an archived dependency, or a compliance surprise β€” star the repo. Stars are how more teams find out their MCP server has a CVE. [](https://star-history.com/#guard0-ai/TrustVector&Date) --- ## πŸ™ Acknowledgments Methodology inspired by [CVSS](https://www.first.org/cvss/), [OWASP LLM Top 10](https://owasp.org/www-project-top-10-for-large-language-model-applications/), [RiskRubric.ai](https://riskrubric.ai/), and [LMSYS Chatbot Arena](https://lmsys.org/). Built with Next.js, TypeScript, Tailwind, Recharts, and Zod. ## πŸ“¬ Community - πŸ› [Issues](https://github.com/guard0-ai/TrustVector/issues) β€” bugs and evaluation corrections - πŸ’¬ [Discussions](https://github.com/guard0-ai/TrustVector/discussions) β€” questions and proposals - πŸ”’ [SECURITY.md](/SECURITY.md) β€” report vulnerabilities - πŸ—ΊοΈ [ROADMAP.md](/ROADMAP.md) β€” what's next ## πŸ“œ License MIT β€” see [LICENSE](/LICENSE). The data is yours to build on. --- **[⭐ Star on GitHub](https://github.com/guard0-ai/TrustVector)** Β· **[🀝 Contribute an evaluation](/CONTRIBUTING.md)** Β· **[πŸ“– Read the methodology](/docs/METHODOLOGY.md)** Made with ❀️ by [Guard0.ai](https://guard0.ai) and the TrustVector community *Trust, but verify β€” then version-control the verification.* ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source β€” we do not rehost the code. - **Author:** [guard0-ai](https://github.com/guard0-ai) - **Source:** [guard0-ai/TrustVector](https://github.com/guard0-ai/TrustVector) - **License:** MIT - **Homepage:** https://trustvector.dev Install and usage instructions live in the source repository linked above. ## Pricing - **Free** β€” Free ## Security capabilities Automated source analysis of v0.1.0 β€” what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source β€” more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** β€” security scan: passed β€” Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-guard0-ai-trustvector - Seller: https://agentstack.voostack.com/s/guard0-ai - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack β€” the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.