Context URL Protocol v1 — Private Server (GCP Confidential Space Enclave) This is a Context URL server running inside a GCP Confidential Space enclave (AMD SEV hardware). It resolves context URLs server-side — all source fetching and LLM processing happens inside the enclave. Raw source data never leaves the trusted execution environment. Only the final result is returned, alongside a cryptographic attestation proving the exact code and hardware that ran. Format: http://enclave.contexturl.ai/v1/#sources!prompt To resolve: extract the spec from the URL fragment (everything after #), then call: GET http://enclave.contexturl.ai/v1/resolve?spec=SPEC The server will fetch all sources, process them through Claude, and return only the final response plus an enclave attestation. You never see the raw source data. Spec format: sources!prompt Sources are separated by | (pipe, only at top level — ignore | inside brackets). The part after the top-level ! is the URI-encoded prompt (+ means space). Each source can be: - A plain URL: url - A weighted URL: weight*url - A bracketed group with inner prompt: weight*[sources!inner_prompt] If no weight is specified, treat all sources as equally weighted. Examples: 1. Single source: http://enclave.contexturl.ai/v1/#news.ycombinator.com!summarize+the+top+stories Resolve: GET http://enclave.contexturl.ai/v1/resolve?spec=news.ycombinator.com!summarize+the+top+stories 2. Multiple sources: http://enclave.contexturl.ai/v1/#news.ycombinator.com|nytimes.com!what+is+happening+today Resolve: GET http://enclave.contexturl.ai/v1/resolve?spec=news.ycombinator.com|nytimes.com!what+is+happening+today 3. Weighted sources: http://enclave.contexturl.ai/v1/#0.7*news.ycombinator.com|0.3*nytimes.com!summarize Resolve: GET http://enclave.contexturl.ai/v1/resolve?spec=0.7*news.ycombinator.com|0.3*nytimes.com!summarize 4. Inner prompts (pre-process each source): http://enclave.contexturl.ai/v1/#[news.ycombinator.com!top+5+headlines]|[nytimes.com!top+5+headlines]!compare+coverage Differential Privacy: To resolve with differential privacy applied to the result, use: GET http://enclave.contexturl.ai/v1/dp/resolve?spec=SPEC&epsilon=1&operation=sum This fetches sources, asks Claude to extract numeric values, then applies calibrated Laplace noise server-side. Only the noisy aggregate is returned — individual source values are never revealed. Parameters: epsilon (privacy budget), operation (sum/mean/count/values). Attestation: Every response includes a cryptographic attestation signed by Google, proving: - Hardware model (GCP_AMD_SEV) - Secure boot status - Container image digest (pins the exact code that ran) - Container image reference Full verification report: GET http://enclave.contexturl.ai/v1/verify Steps to resolve a Context URL on this server: 1. Parse the URL fragment (everything after #) as the spec. 2. Call GET http://enclave.contexturl.ai/v1/resolve?spec= (URL-encode the spec). 3. The server fetches sources, processes via Claude inside the enclave, and returns the result. 4. Report the result to the user. Learn more: https://contexturl.ai Protocol: https://attribution-based-control.ai