AgentBadge — WebMCP Agent-Native Compliance Platform

First agent-native compliance platform powered by WebMCP. 6 imperative tools, 1 declarative form, zero API keys. 145+ readiness rules, on-chain agent passports on Hedera. Browser-mediated tool calls.

Comment

Inspiration

AI agents today face a fundamental integration problem. To use any web service, an agent must: read API docs, obtain API keys, install SDKs, configure authentication, and manage secrets. This is the API key bottleneck — every integration requires human setup.

Meanwhile, MCP (Model Context Protocol) emerged as a standard for exposing tools to LLMs, but it still requires running a separate MCP server, configuring transport (stdio/SSE), and managing permissions server-side. Agents without MCP clients are left out.

We asked: what if agents could discover and use tools directly from the browser — no keys, no servers, no setup? WebMCP answers this by making the browser the mediator. AgentBadge is our answer to "what does an agent-native web look like?"

What it does

AgentBadge is the first agent-native compliance platform powered by WebMCP. It exposes:

  • 6 imperative tools via navigator.modelContext.provideContext():
    1. agent-readiness-scan — scans any URL for 145+ AI readiness checks across 18 categories
    2. badge-generate — generates compliance badge SVG
    3. passport-issue — mints on-chain NFT passport on Hedera
    4. passport-verify — verifies agent identity by token ID or DID (Hedera + EVM)
    5. get-compliance-score — quick compliance score (0-100, grade A-F)
    6. search-rules — searches 145+ compliance rules by keyword
  • 1 declarative form — HTML form-based scan submission for declarative agents
  • Automatic discovery via /.well-known/webmcp.json — agents find tools without reading docs
  • Browser-mediated security — per-tool trust levels (trusted/untrusted, readOnly/write), no API keys
  • 145+ readiness rules across 18 categories — the most comprehensive AI agent readiness scanner

How we built it

WebMCP Layer

  • Discovery endpoint: GET /.well-known/webmcp.json returns a manifest with tool namespaces, endpoints, and descriptors
  • Tool registration: Inline <script> on every page calls navigator.modelContext.provideContext() with 6 tool definitions (input schemas + execute functions)
  • Trust levels: Each tool declares readOnly/write + trusted/untrusted — the browser enforces per-tool access

Scanner (145+ rules, 18 categories)

  • 41 parallel HTTP fetchers — robots.txt, sitemap, llms.txt, OpenAPI, WebMCP manifest, WebMCP runtime, content negotiation, payment protocols, identity/DID, bot auth, SEO/AEO, accessibility, and more
  • WebMCP-specific rules: AB-050 (manifest found), AB-051 (form annotations), AB-069 (browser-side tools), AB-102 (descriptor validation), AB-103 (runtime probe), AB-115 (tool schema), AB-116 (content negotiation)
  • Ed25519-signed reports with SHA-256 integrity hashes

Blockchain Layer

  • Hedera HTS — non-transferable NFT passports ($0.001/tx, 3-5s finality, no smart contracts)
  • HCS — immutable audit trail + agent directory (consensus-ordered, timestamped)
  • Mirror Node — free REST API for reads, no indexer needed
  • DID support: did:hcs:{tokenId}:{serial} (Hedera) and did:eip155:{chainId}:passport:{nftAddress}:{tokenId} (EVM)

Tech Stack

  • Runtime: Bun + Hono (TypeScript-native, server-side rendering)
  • Frontend: HTMX fragments — no React, no build step
  • Blockchain: Hedera Testnet (HTS + HCS + HBAR)
  • Metadata: IPFS via nft.storage (immutable JSON, CID = content hash)

Challenges we ran into

WebMCP Spec Readiness

WebMCP is an emerging spec — browsers haven't shipped navigator.modelContext yet. We built the full integration (tool registration, discovery endpoint, schemas) so it's spec-ready today: all endpoints work as standard HTTP APIs, and will work natively via WebMCP when browsers implement it.

Scanner Complexity

145+ rules across 18 categories with 41 parallel fetchers is complex to orchestrate. We built a parallel fetcher architecture with SSRF protection, content-type validation, and timeout handling. Each fetcher returns structured data that feeds into the rule engine.

Browser-Mediated Trust

Designing per-tool trust levels (trusted/untrusted, readOnly/write) that the browser can enforce required careful thinking about the threat model. Untrusted tools (agent-readiness-scan, get-compliance-score) are safe for any agent to call. Trusted tools (passport-issue, passport-verify) require browser mediation.

Cross-Chain Identity

Supporting both Hedera (did:hcs:) and EVM (did:eip155:) DIDs required abstracting the verification layer. The same passport-verify tool handles both formats, resolving to on-chain NFT ownership checks via the appropriate chain's API.

Accomplishments that we're proud of

  • First WebMCP-native compliance platform — not a wrapper around an existing API, but built from the ground up for browser-mediated agent tool discovery
  • 145+ scanner rules across 18 categories — the most comprehensive AI agent readiness framework, covering discovery, documentation, actionability, payments, identity, infrastructure, SEO/AEO, accessibility, and WebMCP itself
  • 7 WebMCP-specific scanner rules — AgentBadge scans itself and other sites for WebMCP compliance (manifest, runtime, form annotations, tool schemas, content negotiation)
  • Zero API keys architecture — agents discover and use tools without any authentication setup
  • Cross-chain DID support — Hedera and EVM identity in one tool
  • Spec-ready today — full WebMCP integration that works as HTTP APIs now and natively when browsers ship navigator.modelContext

What we learned

  • WebMCP is the right abstraction — agents shouldn't need to read API docs or manage keys. Browser-mediated tool discovery is fundamentally better than the API key model.
  • Scanner-first approach works — before exposing tools, check if a site is agent-ready. 145+ rules across 18 categories gives a comprehensive readiness picture.
  • Declarative + imperative is the right combo — some agents prefer tool calls, others prefer HTML forms. Supporting both covers all agent types.
  • Trust levels are essential — not all tools should be callable by any agent. The trusted/untrusted split lets the browser enforce access without authentication.

What's next for AgentBadge

  • Browser-native WebMCP — when Chrome ships navigator.modelContext, AgentBadge tools will work natively without any changes to our code
  • More imperative tools — expand beyond 6 tools to cover marketplace tasks, A2A messaging, and agent directory search
  • WebMCP compliance certification — a badge program for sites that pass all 7 WebMCP scanner rules
  • Agent passport tiers via WebMCP — let agents upgrade their on-chain tier through browser-mediated tool calls
  • Cross-browser testing — validate WebMCP integration across Chrome, Firefox, Safari as implementations ship
  • Community rule contributions — open the 145+ rule framework for community-submitted readiness checks

Built With

  • base
  • bun
  • docker
  • ed25519
  • evm
  • fly.io
  • hbar
  • hcs
  • hedera
  • hono
  • htmx
  • hts
  • ipfs
  • json-ld
  • mcp
  • nft.storage
  • openapi
  • rest
  • tailwindcss
  • typescript
  • vitest
  • webmcp
  • x402
  • zod
Share this project:

Updates

Submission history