The web, cached as clean markdown for AI agents.
A global shared cache backed by Cloudflare Workers + KV. Multiple independent instances contribute and verify cached content through a self-healing consensus model that prevents poisoning -- if an entry is stale or tampered with, the next reader re-fetches and overwrites it.
All endpoints are served from https://agentsweb.org.
Read cached markdown for a URL. Returns the markdown, trust level, source, and age.
Contribute a new cache entry. Body: { url, markdown, source, instance_id? }. Content is validated through security gates before acceptance.
Confirm an existing entry matches your local fetch, boosting its trust level. Body: { url, content_hash, instance_id? }.
Public cache statistics (total hits, writes, hit rate).
- Content gates -- incoming markdown is scanned for prompt injection patterns, captcha/challenge pages, login walls, malicious HTML, and suspicious encoding before it enters the cache.
- SSRF protection -- private/internal IP ranges and non-HTTP schemes are blocked at the URL validation layer.
- Rate limiting -- per-IP limits enforced via KV TTLs (600 reads/min, 10 writes/min, 60 confirms/min).
- Trust levels -- entries start at trust 1; independent confirmations increment trust, extending TTL and making the entry harder to overwrite.
- Self-healing on read -- low-trust or mismatched entries can be replaced by any client that provides fresher content.
- Domain-aware TTLs -- news sites expire in 1 day, documentation/reference sites persist up to 30 days.
The primary client is intercept-mcp, an MCP server that reads from and contributes to this cache automatically.
- Cloudflare Workers (TypeScript)
- Cloudflare KV for storage
- Custom domains:
agentsweb.org,api.agentsweb.org