ai-trust

ai-trust

OpenA2A: CLI · HackMyAgent · Secretless · AIM · Browser Guard · DVAA

Trust verification CLI for AI packages. MCP servers, A2A agents, skills, AI tools, LLMs. Queries the OpenA2A Registry trust graph for security scans, community consensus, dependency risk, and known advisories. Apache 2.0.

Website · Registry · Discord

For general-purpose libraries (express, typescript, chalk, etc.) use HackMyAgent instead. ai-trust is scoped to AI-native packages only.

Quick start

npx ai-trust check @modelcontextprotocol/server-filesystem

  @modelcontextprotocol/server-filesystem  mcp_server · scanned 2 days ago
  No known issues

  Trust     ━━━━━━━━━━━━━━━━━━━━ 87/100
  Level     Scanned (3/4)
  Blocked > Warning > Listed > Scanned > Verified

  ── Next Steps ──────────────────────────────────────────────
  Fresh scan:         ai-trust check @modelcontextprotocol/server-filesystem
  Full project audit: ai-trust audit package.json

Install

npm

npx ai-trust check <pkg>     # run once, no install
npm install -g ai-trust      # install globally

Requires Node.js 18 or later.

Homebrew

brew install opena2a-org/tap/ai-trust

From source

git clone https://github.com/opena2a-org/ai-trust.git
cd ai-trust
npm install
npm run build
node dist/index.js check express

Verifying what was installed

Every release publishes via npm Trusted Publishing with SLSA v1 provenance. No long-lived NPM_TOKEN. GitHub Actions exchanges its OIDC token with npm at publish time.

npm view ai-trust dist.attestations --json
# Expects non-empty result with predicateType "https://slsa.dev/provenance/v1"

Scope: AI packages only

ai-trust verifies trust for AI-native packages. For everything else, use HMA.

Your package is...	Use
MCP server, A2A agent, skill, AI tool, LLM	ai-trust
General-purpose library (express, chalk, typescript, etc.)	hackmyagent check <pkg>
Full codebase security audit	hackmyagent secure .

ai-trust audit package.json audits AI packages in the trust table and separately lists libraries in an "Out of scope" section with an HMA pointer.

Running ai-trust check express on a general-purpose library returns an "out of scope" verdict with a redirect to hackmyagent check express. Intentional. ai-trust is for AI packages only.

Commands

check

Look up the trust verdict for a single AI package.

ai-trust check @modelcontextprotocol/server-filesystem
ai-trust check my-custom-agent --type a2a_agent
ai-trust check @modelcontextprotocol/server-postgres --json
ai-trust check mcp-server-xyz --scan-if-missing --contribute  # download + scan + share
ai-trust check server-filesystem --no-scan                    # registry lookup only
ai-trust check /path/to/local --scan-path /path/to/local      # scan a local directory

Flags: --type, --scan-if-missing, --contribute, --no-scan, --no-deep, --scan-path, --json.

MCP server shorthand

# These are equivalent:
ai-trust check server-filesystem
ai-trust check @modelcontextprotocol/server-filesystem

# Third-party MCP servers use their own package names:
ai-trust check mcp-server-kubernetes
ai-trust check @supabase/mcp-server-supabase
ai-trust check @cloudflare/mcp-server-cloudflare

server-* resolves to @modelcontextprotocol/server-*. Third-party mcp-server-* packages are looked up by their actual name.

Scan on demand

When a package is not in the registry, ai-trust can download and scan it locally using HackMyAgent. In interactive mode, you are prompted. In CI:

ai-trust check mcp-server-xyz --scan-if-missing --contribute   # auto-scan + share
ai-trust check server-filesystem --no-scan                     # skip scanning entirely

Local scans run HMA with NanoMind semantic analysis enabled by default. Pass --no-deep for static-only.

audit

Parse dependency files and audit AI packages. Supports .json (package.json format) and .txt (requirements.txt format). Libraries get partitioned into an "Out of scope" section.

ai-trust audit package.json
ai-trust audit requirements.txt
ai-trust audit package.json --min-trust 2                      # custom threshold (default 3)
ai-trust audit package.json --scan-missing --contribute        # scan unknown AI packages

Example output (mixed AI + libraries):

  5 AI packages audited · 9 libraries out of scope

  PACKAGE                    TYPE          VERDICT   TRUST       SCORE         SCAN
  ──────────────────────────────────────────────────────────────────────────────────────
  @modelcontextprotocol/sdk  mcp_server    SAFE      Scanned     ━━━━━━━━ 87  passed
  @opena2a/aim-core          a2a_agent     SAFE      Scanned     ━━━━━━━━ 81  passed
  ...

  ── Out of scope (libraries) ────────────────────────────────
  ai-trust is for AI packages. For general security, use HackMyAgent.
  @noble/ed25519, @noble/post-quantum, commander, js-yaml, onnxruntime-node + 4 more

  ── Next Steps ──────────────────────────────────────────────
  Library security:  npx hackmyagent secure .

batch

Look up trust verdicts for multiple AI packages at once. Non-AI packages get partitioned into the "Out of scope" footer.

ai-trust batch @modelcontextprotocol/server-filesystem @modelcontextprotocol/server-postgres
ai-trust batch my-server-a my-server-b --type mcp_server
ai-trust batch react vue express lodash chalk

Flags: --type, --min-trust.

Output options

ai-trust check express --json                                    # JSON output for scripting
ai-trust audit package.json --json                               # JSON audit output
ai-trust check express --no-color                                # disable colored output
ai-trust check express --registry-url http://localhost:8080      # custom registry endpoint

Trust levels

Level	Label	Description
0	Blocked	Package is blocked due to security concerns
1	Warning	Package has known issues
2	Listed	Package is listed but not yet scanned
3	Scanned	Package has been scanned by HackMyAgent
4	Verified	Package is verified by the publisher

Exit codes

Code	Meaning
0	All queried packages are safe and meet the trust threshold
1	Operational error (network failure, file not found, server error)
2	Policy signal: one or more packages have warning or blocked verdict, or fall below --min-trust

Community contribution

Every scan you run can improve trust data for the entire community. Scan results are shared as anonymised telemetry: check pass/fail and severity only. No file paths, source code, or descriptions.

On first scan, ai-trust asks whether you want to contribute. Your choice is saved in ~/.opena2a/config.json and shared across all OpenA2A tools (opena2a-cli, hackmyagent).

ai-trust check chalk --contribute              # contribute for this scan (non-interactive / CI)
opena2a config set contribute true             # opt in globally
opena2a config set contribute false            # opt out globally

More scans contributed means packages move from "Listed" to "Scanned" faster, reducing risk for everyone.

Using with opena2a-cli

opena2a-cli is the unified CLI for the OpenA2A security toolchain. ai-trust powers opena2a trust.

npm install -g opena2a-cli
opena2a trust @modelcontextprotocol/server-filesystem
opena2a review                              # full security dashboard

Use cases

Guide	Time
Check if a package is safe before installing	2 min
Verify an MCP server's trust score	3 min
Contribute trust data to the community	3 min

Full index: docs/USE-CASES.md.

Contributing

Apache 2.0. PRs from outside the org welcome.

git clone https://github.com/opena2a-org/ai-trust.git
cd ai-trust && npm install && npm run build && npm test

Security issues: [e-mail] (coordinated disclosure, response within 24 hours).

Links

• Website
• OpenA2A Registry. Trust scores and scan data.
• OpenA2A CLI. Unified security CLI.
• HackMyAgent. Local scanning for unverified packages.
• aicomply. Inline PII, credential, and regulated-data classification for agent I/O before it reaches an LLM.

Part of the OpenA2A security platform.

License

Apache-2.0.