Compliance mappings are bundled rather than fetched, so a control lookup costs nothing and works with the network unplugged — which is the point for on-prem and SCIF deployments. Responses carry an ML-DSA-65 signature and land in a DAG chain that later exports as an attestation package or a DFARS POA&M, so the audit artefact falls out of the work instead of being assembled afterwards. Two decisions to make early: the hosted endpoint keeps its DAG in memory, so run Docker or the binary if the trail has to outlive the session, and most of the catalogue sits above the free Community tier.
KHEPRA is a compliance and security engine that runs on your own hardware and answers as an MCP server. It bundles 36,195 STIG/CCI/NIST/CMMC mappings so lookups happen offline, signs responses with ML-DSA-65 (FIPS 204), and records what the agent did into a DAG audit chain. In sovereign and ironbank modes it makes no outbound network calls at all.
- pqc_stig assesses a codebase against PQC-01-STIG-V1R1 — the project's post-quantum STIG covering CNSA 2.0 and FIPS 203/204/205 — and returns per-control findings with a signed compliance score
- khepra_query_stig and nist_map search the bundled 36,195-mapping database of STIG rules, CCI items, NIST 800-53 Rev 5 and 800-171 Rev 2 controls and CMMC 2.0, BM25-ranked and air-gap safe
- cmmc_assess scores CMMC Level 1, 2 or 3 with gaps and a C3PAO readiness flag; khepra_export_poam produces a DFARS 252.204-7012 POA&M and khepra_export_attestation a signed evidence package
- ert_scan runs SBOM, CVE, secrets, STIG and PQC inventory inside a Docker sandbox, with ert_readiness, ert_architect, ert_crypto and ert_godfather as the packaged passes — 800-171 alignment, supply-chain risk enriched with CISA KEV and EPSS, PQC readiness including weak primitives like MD5 and RC4, and a board-level causal chain with dollar impact
- godfather_report can be staged: with approval_required it returns a single-use token, and godfather_approve consumes that token to release the report
- Scanning and forensics: vuln_scan across Go, NPM and Python dependencies, secret_scan by entropy and pattern, container_scan on Dockerfiles, port_scan, enumerate_host, fingerprint_device, packet_analyze over a PCAP, forensic_snapshot and fim_baseline
- Non-human identities: nhi_inventory, nhi_orphans, nhi_excessive, nhi_expired and nhi_revoke, with attack_graph turning that inventory into lateral-movement paths and blast radius
- Incidents and credentials: ir_incident and ir_add_ioc, acp_issue, acp_status and acp_revoke — each signed and written to the chain
- The audit trail is addressable in its own right: dag_write, dag_query, dag_audit, audit_dag_integrity, khepra_get_dag_chain and dag_attestation, with flight_record, agent_record and flight_export producing a CMMC-aligned evidence packet of what the agent itself did
- threat_lookup and khepra_query_threat_intel read an embedded CISA KEV and MITRE ATT&CK database offline; threat_model builds a STRIDE model with NIST 800-53 and ATT&CK mappings; owasp_agent_assess scores this deployment against the OWASP Agentic Top 10
- KASA is the autonomous half — kasa_start, kasa_status, kasa_scan, kasa_task, kasa_forensics and kasa_crypto_agent — with ouroboros_waf_eye, ouroboros_vuln_eye, ouroboros_fim_eye and ouroboros_stig_eye reporting continuous posture
Three ways in. The hosted endpoint at https://mcp.souhimbou.ai/sse needs no install, reached over SSE or through mcp-remote, but it runs in edge mode where the DAG is in memory and goes with the session. Self-hosting is Docker (ghcr.io/nouchix/pqc-khepra-mcp:latest) or a Go 1.21+ binary that starts in about 300ms. KHEPRA_MODE picks sovereign (the default), ironbank, hybrid or edge, and an unknown value is rejected at startup and falls back to sovereign. KHEPRA_LICENSE_KEY is needed only above the free Community tier, which is pqc_stig, nist_map, khepra_query_stig and a core set; Pro is $19/mo and Enterprise $499/mo, neither charging per token or per query. Also configurable: KHEPRA_HOME, KHEPRA_DAG_PATH, KHEPRA_AUDIT_LOG_PATH, KHEPRA_NETWORK_POLICY, KHEPRA_MAX_CONCURRENT (default 5) and MCP_PQC_ENABLED.
One command plus a key — docker pull ghcr.io/nouchix/pqc-khepra-mcp:latest, then supply credentials