Ship a tool your users can trust — and prove it.
A signed, offline-recomputable safety score for the MCP server, package, or agent tool you publish. Free, no account. Claim it, put its badge on your README, or check a private repo first.
Prove it's yours
A claimed tool shows a “✓ Maintainer-claimed” chip on its result, and we watch it for you: you get an alert when its answer or score changes.
- Sign in and paste
owner/repo. - Add the GitHub topic we give you,
agentavow-verify-…, or rungh repo edit owner/repo --add-topic agentavow-verify-… - Press Verify topic. We start watching it for you.
X-AgentAvow-Verify header or serverInfo, or serve /.well-known/owners.json listing your verified account emailPut the answer on your README
One line of Markdown. It shows the current signed answer, trust score and adoption, and links to the full report. No account needed for a public tool.
[](https://agentavow.com/check/you/your-repo)Check it before it's public
Two ways. Both need you signed in.
- GitHub App. Install it on the repos you choose. We check them now and re-check daily.
- One-time token. A fine-grained, read-only token for a single check. It's used for that check and never stored.
What you get: a private report only you can open. It isn't listed in Browse or search, and it has no public badge. To get a badge, publish the result: publishing lists it in Browse and turns the badge on.
Your badge, your way.
The badge regenerates on every view, so it never goes stale in your README, and clicking it opens the full signed report. Pick a style, copy one line.
[](https://agentavow.com/check/you/your-repo)
More styles and themes: Badge docs →
Or embed the live widget
The trust card (the answer, the trust bar and the adoption dial), linked to the report, plus a Verify offline button that checks the Ed25519 signature in your reader's own browser. One script tag, no tracking.
<script src="https://agentavow.com/widget.js"
data-tool="you/your-repo"></script>Enter a repo or package to preview the live widget.
Not listed yet? Add it.
Pasting a link into Check is the fastest way to see an answer, and a GitHub repo or an npm, PyPI, crates, Docker or Hugging Face package you check there joins the public catalog on its own. A live MCP server checked by URL doesn't. Add it here: we scan it, sign the result and list it in Browse. Free, no account.
Why put a score on your tool.
It works because it's the one score a skeptic can check for themselves.
Every score is signed (Ed25519), so anyone can prove who issued it, when, and that nobody changed the number in your README. They verify it offline against our public keys.
MCP servers and agent tools are multiplying. A signed high score is a trust signal to your users for a new project nobody has vetted yet.
MCP server, npm, PyPI, GitHub repo, or skill: one recomputable score across all of them, that travels with the artifact.
Score poorly and the report points at the exact file and line, with the fix. Score well and you get the badge.
Gate pull requests and agents on the answer.
For a public repo, the hosted GitHub Action comments the answer and the trust score on every PR, and fails the build when the answer is Do not connect. When the repo publishes a package, AgentAvow also runs it in the behavioral sandbox and the Action prints a Sandbox: line; fail_on_behavioral: true fails the build on a high or critical sandbox finding. Prefer to keep it all local? agentavow scan . runs offline, and nothing leaves your machine.
- uses: AgentAvow/AgentAvow/github-action@main
with:
fail_on: do_not_connect # the default; "review" is stricter
fail_on_behavioral: true- GitLab CI/CD Catalog → gate merge requests on the answer in your own runner
- Bitbucket Pipelines → a pipe that posts a Code Insights report on the pull request
- Azure DevOps → a pipeline template with a run summary and SARIF
- Vercel AI SDK → wrap your tools so a Do not connect answer never runs
- Flue → gate the tools a Flue agent connects to
- Claude Agent SDK → leave out a server that fails, and deny each mcp__ call that does
- OpenAI Agents SDK → wrap an MCP server so connect, listTools and callTool are checked
- Open Policy Agent → a Rego policy your gateway or Envoy asks before a server or tool call
- LiteLLM proxy → a guardrail on every MCP tool call through LiteLLM's MCP gateway
- Behavioral sandbox → see what a package actually does when it runs
Carry the score into the MCP Registry.
The signed block rides the publisher-provided _meta slot into your registry record. It is offline-recomputable, so a reviewer checks it without trusting us. Live block: /api/v1/public/registry-snippet/you/your-repo
"_meta": {
"io.modelcontextprotocol.registry/publisher-provided": {
"com.agentavow/trust": {
"trustScore": 92,
"certified": false,
"subject": "github:you/your-repo",
"issuer": "did:web:agentgraph.co",
"attestation": "<signed JWS — recompute & verify offline>",
"jwks": "https://agentgraph.co/.well-known/jwks.json"
}
}
}Give your agent a safety check before it connects anything.
Add AgentAvow's own MCP server to Claude, ChatGPT, Cursor, or any MCP client. Your agent can ask "is this tool safe to install?" and get a signed 0-100 score and findings back, before it connects an unknown server or runs a package. Free, no account, read-only.
Claude Code: claude mcp add agentavow-trust -- uvx agentavow-trust
Or install the Claude Code plugin, which also checks each new MCP server at session start and warns you before your agent uses one: /plugin marketplace add AgentAvow/AgentAvow then /plugin install agentavow-trust@agentavow. How it works.
The plugin's checks run on your machine, so they cost Claude nothing: zero tokens per check, under 1,000 a session. The numbers.
{
"mcpServers": {
"agentavow-trust": {
"command": "uvx",
"args": [
"agentavow-trust"
]
}
}
}Prefer the terminal? npx agentavow-trust scan owner/repo gives a signed score from the command line, no install.
