CAPAS is a deterministic gate with no language model in the decision and no training step; submitted claims never train anything. This page separates what is available today from the roadmap.
pip install capas-claim-gate. The gate runs in-process on your machine; evidence never leaves it.
python3 capas_mcp.py as a local tool for your agent. Same in-process gate, no outbound calls.
A managed endpoint for signed certificates. Opt-in, auth-gated and isolated per pilot.
source_urls and the process runs with CAPAS_VERIFY_REMOTE_SOURCES=1, CAPAS fetches those URLs to confirm they resolve and hash-match. Without that variable nothing is fetched and the criterion source_urls_recoverable_hashable reads false. The hosted API does not set it; a payload with no URLs triggers no network call.Source text is not used to train a model. There is no model in the decision path and no training step. The same payload always yields the same verdict.
In library and MCP mode no data leaves the environment: the gate runs in-process. The one exception is opt-in provenance verification: with source_urls present and CAPAS_VERIFY_REMOTE_SOURCES=1 set, those URLs are fetched. The hosted API receives only the payloads sent to it.
Certificate issuance requires authentication: Authorization: Bearer or X-API-Key via CAPAS_API_KEY. Certificates are signed with HMAC-SHA256 and content-addressed, so tampering is detectable on verification. The certificate store path is set by CAPAS_DATA_DIR; retention is defined per pilot.
The library or MCP server runs inside the customer environment on full or redacted records. With no source URLs, nothing is transmitted. Pilots can start on synthetic samples or customer-redacted payloads.
The local gate writes no telemetry. Hosted API logs are operational, configurable per deployment, and retain no payload content by requirement.